C H A P T E R

1
A systematic view of remote sensing
O U T L I N E

1.1 Introduction 2 1.6.1.1 Scene generation 28

1.6.1.2 Surface radiation
1.2 Platform and sensor systems 3
modeling 28
1.2.1 Geostationary satellites 3
1.6.1.3 Atmospheric radiative
1.2.2 Polar-orbiting satellites 5
transfer 28
1.2.3 Overview of major satellite missions
1.6.1.4 Sensor modeling 30
and programs 6
1.6.2 Inversion methods 30
1.2.3.1 USA 8
1.6.2.1 Statistical analysis and
1.2.3.2 Europe 9
machine learning
1.2.3.3 China 10
techniques 31
1.2.4 Small satellites and satellite
1.6.2.2 Optimization algorithms 34
constellations 11
1.6.2.3 Look-up table algorithms 35
1.2.5 Sensor types 14
1.6.2.4 Direct estimation
1.2.6 Data characteristics 16
methods 35
1.2.6.1 Spatial resolution 16
1.6.2.5 Data assimilation methods 36
1.2.6.2 Spectral resolution 16
1.6.2.6 Spatial and temporal
1.2.6.3 Temporal resolution 17
scaling 37
1.2.6.4 Radiometric resolution 17
1.6.2.7 Regularization method 39
1.3 Data transmission and ground 1.6.3 Use of multisource data 39
receiving system 18 1.6.4 Use of a prior knowledge 39
1.6.5 Spaceetime constraints 40
1.4 Data processing 19
1.6.6 Algorithm ensemble 41
1.4.1 Radiometric calibration 19
1.4.2 Geometric processing 20 1.7 Production, archiving, and distribution
1.4.3 Image quality enhancement 22 of high-level products 41
1.4.4 Atmospheric correction 23
1.8 Product validation 44
1.4.5 Image fusion and product integration 25
1.9 Remote sensing applications 44
1.5 Mapping category variables 25
1.10 Conclusion 45
1.6 Estimating quantitative variables 27
1.6.1 Forward radiation modeling 28 References 47

Advanced Remote Sensing, Second Edition

https://doi.org/10.1016/B978-0-12-815826-5.00001-5 1 © 2020 Elsevier Inc. All rights reserved.
2 1. A systematic view of remote sensing

moves on to the data transmission and ground

Abstract
This chapter provides an overview of the remote receiving system, the processing system for
sensing system, including the platform and sensor sys- handling the geometric and radiometric proper-
tem, data transmission and ground receiving system, ties of data, the analysis system for extracting in-
processing system of radiometric and geometric prop- formation on both category and numerical
erties, analysis system for mapping category variables
variables of the Earth surface environment, the
and generating high-level products of quantitative
variables, product production and distribution sys- product generation and distribution system, the
tem, product validation system, and remote sensing product validation system, and end-user appli-
applications. It aims to present a complete picture of cations. Applications largely define the data
the state-of-the-art development of remote sensing acquisition system, and end-users often need to
techniques by linking different chapters in the rest of
validate the products to quantify their errors
the book and filling in any possible gaps.
and uncertainties.

1.1 Introduction
We are living in a world where population is
rapidly increasing, depleting natural resources,
and experiencing the possible consequences of
human-induced climate change. Our ability to
meet these challenges partially depends on how
well we understand the Earth system and use
that information to guide our actions. Remote
sensing is a tremendous source of information
needed by policy-makers, resource managers,
forecasters, and other users, and it has become
increasingly vital for the effective and sustainable
future management of the Earth. A remote
sensing system consists of instrumentation, pro-
cessing, and analysis designed to measure,
monitor, and predict the physical, chemical, and
biological aspects of the Earth system. Sophisti-
cated new technologies have been developed to
gather vast quantities of data, and the mathemat-
ical and physical sophistication of the techniques
used to process and analyze the observed data
has increased considerably.
The first chapter of the book aims to link
diverse components to paint a full picture of a
remote sensing system as illustrated in Fig. 1.1.
It starts with a brief introduction to the platform FIGURE 1.1 Key components of the remote sensing
and sensor system for acquiring data and then system.
 1.2 Platform and sensor systems 3

1.2 Platform and sensor systems The satellite appears motionless at a fixed posi-
tion in the sky to ground observers. There are
The data acquisition system mainly consists of several hundred communication satellites and
the sensor and the platform on which the sensor several meteorological satellites in such an orbit.
resides. The platform may be on the surface, in Fig. 1.2 illustrates a few typical meteorological
the air, or in space. A surface platform may be satellites in the geostationary orbit relative to
a ladder, tower, cherry picker, crane, building, the polar-orbiting satellites.
or scaffolding that provides data used primarily US operational weather satellites include the
for validation. Geostationary Operational Environmental Satel-
Aerial platforms include aircraft and balloons. lite (GOES) used for short-range warning and
Unmanned aerial systems (UAS), commonly “now-casting” primarily to support the National
known as a drone, have considerable potential Weather Service requirements. The procure-
to radically improve Earth observation by ment, design, and manufacturing of GOES are
providing high spatial detail over relatively large overseen by the National Aeronautics and Space
areas in a cost-effective way and an entirely new Administration (NASA), while all operations of
capacity for enhanced temporal retrieval the satellites once in orbit are effected by the
(Manfreda et al., 2018). In addition to the National Oceanic and Atmospheric Administra-
increasing availability of UAS and affordability, tion (NOAA). Before being launched, GOES
recent advances in sensor technologies and satellites are designated by letters (-A, -B, -C).
analytical capabilities have stimulated an explo- Once a GOES satellite is launched successfully,
sion of interest from the remote sensing commu- it is redesignated with a number (
1,
2,
3).
nity. Increasing miniaturization allows Normally two GOES satellites are operational.
multispectral, hyperspectral, and thermal imag- Information on the GOES series is shown in
ing, as well as synthetic-aperture radar (SAR) Table 1.1. The third generation of GOES, the
and light detection and ranging (LiDAR) sensing new GOES-R satellite series program, consisting
to be conducted from UAS. of four satellites (from GOES-16), represents a
Spaceborne platforms are mainly satellites significant improvement in spatial, temporal,
and space shuttles. As the landmark of space- and spectral observations over the capabilities
borne remote sensing, Landsat 1 was launched of the previously operational GOES series. For
in 1972. Since then, there have been over 50 example, the Advanced Baseline Imager (ABI)
countries operating land remote sensing satel- is the primary instrument on the GOES-R Series
lites. The Committee on Earth Observation Satel- for imaging Earth’s weather, oceans, and envi-
lites (CEOS) database (http://database. ronment. The ABI provides three times more
eohandbook.com/database/missiontable.aspx) spectral information, four times the spatial reso-
lists the current and future satellite missions and lution, and more than five times faster temporal
sensors. The following will mainly discuss the coverage than the previous system.
satellite remote sensing. European operational missions are currently
operated by the European Organization for the
Exploitation of Meteorological Satellites
(EUMETSAT). EUMETSAT’s geostationary
1.2.1 Geostationary satellites satellite programs include the Meteosat First
A geostationary satellite is in an orbit that can Generation system (up to Meteosat-7) from
only be achieved at an altitude very close to 1977 to 2017, four Meteosat Second Generation
35,786 km (22,236 miles) and which keeps the (MSG) satellites (MSG-1,2,3,4 or Meteosat-
satellite fixed over one longitude at the equator. 8,9,10,11) from 2004 to 2025, and six Meteosat
4 1. A systematic view of remote sensing

FIGURE 1.2 Illustration of the distribution of a few common geostationary satellites compared to the polar-orbiting
satellites.

Third Generation (MTG) satellites from 2021 to are located in orbit at around 140.7 degrees east
39. The MSG satellites carry an impressive pair and will observe the East Asia and Western
of instruments: the Spinning Enhanced Visible Pacific regions for a period of 15 years. The
and Infrared Imager (SEVIRI), which has the ca- Advanced Himawari Imager (AHI), similar to
pacity to observe the Earth in 12 spectral chan- ABI, has six channel multispectral bands in the
nels and provide image data every half hour, visible to near-infrared spectrum with 500m
and the Geostationary Earth Radiation Budget spatial resolution and provides full disk observa-
(GERB) instrument supporting climate studies. tions every 10 min and images of Japan every
The Japanese Geostationary Meteorological 2.5 min.
Satellite (GMS) series had five satellites from China has launched eight of the first-
1977. The Multifunctional Transport Satellites generation geostationary satellites named Fen-
(MTSAT) are the successors to the GMS 1e5 sat- gyun (FY-2) from FY-2A to FY-2H since 1997.
ellite series. The MTSAT-2 from 2010 was also The second generation of geostationary meteoro-
known as Himawari-7. Himawari-8 was opera- logical satellites FY-4 was launched in December
tional from July 2015, and Himawari-9 started 2016, and multiple FY-4 satellites have been
backup operation on March 2017. Both satellites planned to provide service through 2037 when
 1.2 Platform and sensor systems 5
TABLE 1.1 Information on GOES satellite series. 1.2.2 Polar-orbiting satellites
Satellites Launch day Status Polar-orbiting satellites can provide an obser-
1 October 16, Decommissioned
vational platform for the entire Earth surface,
1975 while their geostationary counterparts are limited
to approximately 60 degrees of latitude of geosta-
2 June 16, 1977 Decommissioned
tionary meteorological satellites at a fixed point
3 June 16, 1978 Decommissioned over the Earth. Polar-orbiting satellites are able
4 September 9, Decommissioned to circle the globe approximately once every
1978 100 min. Most polar-orbiting Earth observation
5 May 22, 1981 Deactivated on July 18, 1990
satellites, such as Terra, ENVISAT, and Landsat,
have an altitude of about 800 km. They are in
6 April 28, 1983 Decommissioned sun-synchronous orbits passing directly over a
G May 3, 1986 Failed to orbit given spot on the ground at the same local time.
7 February 26, Used as a communications
A relatively low orbit allows detection and collec-
1987 satellite; decommissioned 2012 tion of data by instruments aboard a polar-
orbiting satellite at a higher spatial resolution
8 April 13, 1994 Decommissioned 2004
than from a geostationary satellite.
9 May 23, 1995 Decommissioned 2007 NASA has launched a series of polar-orbiting
10 April 25, 1997 Decommissioned 2009 satellite missions with the ability to characterize
the current state of the Earth system. The
11 May 3, 2000 Decommissioned 2011
currently active satellites are illustrated in
12 July 23, 2001 Decommissioned 2013 Fig. 1.3. All the missions fall into three types:
13 May 24, 2006 On-orbit storage exploratory, operational precursor and technol-
ogy demonstration, and systematic.
14 June 27, 2009 On-orbit spare
Exploratory missions are designed to yield
15 March 4, 2010 Operational West backup new scientific breakthroughs. Each exploratory
16 November 19, Currently operating as GOES satellite project is expected to be a one-time
(GOES-R) 2016 East mission that can deliver conclusive scientific
17 March 1, 2017 Currently operating as GOES
results addressing a focused set of scientific
(GOES-S) West questions. In some cases, an exploratory mission
may focus on a single pioneering measurement
GOES-T Planned to
launch in 2020
that opens a new window on the behavior of
the Earth system. These missions are managed
GOES-U Planned to in the NASA Earth System Science program
launch in 2024
(ESSP). Examples include the Gravity Recovery
and Climate Experiment (GRACE) and Cloud-
SAT. GRACE data can be used for estimating
a successor program will be inaugurated. The soil moisture and surface/underground water
Advanced Geosynchronous Radiation Imager (Section 20.4).
(AGRI) aboard FY-4 is the corresponding version Operational precursor and technology
of ABI in the GOES-R series. It has 14 spectral demonstration missions enable major upgrades
bands, delivering full disk images every 15 min of existing operational observing systems.
at a significantly improved resolution of NASA is investing in innovative sensor technol-
0.5e4 km. ogies and developing more cost-effective
6 1. A systematic view of remote sensing

FIGURE 1.3 Illustration of the current NASA Earth observing satellites, downloaded from https://eospso.gsfc.nasa.gov/
in February 2019.

versions of its pioneer scientific instruments that centerpiece of NASA’s recent Earth observation
can be used effectively by operational agencies. program. It was conceived in the 1980s and
An example is the NMP EO-1 (New Millennium began to take shape in the early 1990s. It is
Program Earth Observing-1) mission launched composed of a series of satellites and sensors, a
on November 21, 2000, which includes three science component, and a data system support-
advanced land imaging sensors and five revolu- ing a coordinated series of polar-orbiting and
tionary crosscutting spacecraft technologies. The low inclination satellites for long-term global ob-
three sensors led to a new generation of lighter servations of the land surface, biosphere, solid
weight, higher performance, and lower cost Earth, atmosphere, and oceans. Complete and
Landsat-type Earth surface imaging instru- still active EOS satellites are shown in Tables 1.2
ments. The hyperspectral sensor Hyperion is and 1.3.
the first of its kind to provide images of land sur-
face in more than 220 spectral bands.
Systematic missions provide systematic
1.2.3 Overview of major satellite
measurements of key environmental variables
missions and programs
that are essential to specify changes in forcings
caused by factors outside the Earth system There exist 72 different government space
(e.g., changes in incident solar radiation) and to agencies as of 2018, and 14 of those have launch
document the behavior of the major components capability. Six government space agencies have
of the Earth system. An example is the Earth full launch capabilities, i.e., launch and recover
Observing System (EOS) program. EOS is the multiple satellites, deploy cryogenic rocket
 1.2 Platform and sensor systems 7
TABLE 1.2 Active EOS satellites as of April 2019. TABLE 1.2 Active EOS satellites as of April
2019.dcont’d
Satellites Launch day
Satellites Launch day
Aqua May 4, 2002
Terra December 18,
Aura July 15, 2004 1999
Cloud-aerosol LiDAR and infrared April 28, 2006 The global change observation May 18, 2012
pathfinder Satellite observation (CALIPSO) mission-water (GCOM-W1)
CloudSat April 28, 2006 Total solar irradiance spectral solar December 15,
Cyclone global Navigation Satellite system December 15, irradiance 1 (TSIS-1) 2017
(EVM-1) (CYGNSS) 2016
Deep space climate observatory (DSCOVR) February 11,
2015
TABLE 1.3 Completed EOS satellites as of April
ECOsystem spaceborne thermal radiometer June 29, 2018 2019.
experiment on space station (EVI-2)
(ECOSTRESS) Satellites Lunch day
Global ecosystem dynamics investigation December 5, Combined Release and Radiation Effects July 25, 1990
LiDAR (EVI-2) (GEDI on ISS) 2018 Satellite (CRRES)
Global precipitation measurement core February 27, Upper Atmosphere Research Satellite September 12,
observatory (GPM Core) 2014 (UARS) 1991
Gravity recovery and climate experiment May 22, 2018 Atmospheric Laboratory of Applications March 24,
follow on (GRACE-FO) and Science (ATLAS) 1992
Ice, cloud, and land Elevation Satellite-2 September 15, TOPEX/Poseidon August 10,
(ICESat-2) 2018 1992
Jason-3 January 17, Spaceborne Imaging Radar-C (SIR-C) April 19, 1994
2016
Radar Satellite (RADARSAT) November 4,
Landsat 7 April 15, 1999 1995
Landsat 8 February 11, Total Ozone Mapping Spectrometer-Earth July 2, 1996
2013 Probe (TOMS-EP)
Lightning imaging sensor on ISS February 19, Advanced Earth Observing Satellite August 17,
(LIS on ISS) 2017 (ADEOS) 1996
Ocean surface topography Mission/Jason-2 January 20, Orbview-2/SeaWiFS August 1, 1997
(OSTM/Jason-2) 2008
Tropical Rainfall Measuring Mission November 27,
Orbiting carbon observatory 2 (OCO-2) July 2, 2014 (TRMM) 1997
Quik Scatterometer (QuikSCAT) June 19, 1999 Tomographic Experiment using Radiative May 18, 1999
Soil moisture active-passive (SMAP) January 31, Recombinative ionospheric EUV and Radio
2015 Sources (TERRIERS)

Solar radiation and climate experiment January 25, Active Cavity Radiometer Irradiance December 20,
(SORCE) 2003 Monitor Satellite (ACRIMSAT) 1999

Stratospheric aerosol and gas experiment February 18, Challenging Mini-Satellite Payload July 15, 2000
III on ISS (SAGE III-ISS) 2017 (CHAMP)

(Continued)
8 1. A systematic view of remote sensing

TABLE 1.3 Completed EOS satellites as of April The following will introduce the major satellite
2019.dcont’d programs of the United States, Europe, and
Satellites Lunch day China.

Earth Observing-1 (EO-1) November 21, 1.2.3.1 USA

2000
The United States has three major federal
Jason-1 December 7, agencies involved in the EO satellites: NASA,
2001 NOAA, and US Geological Survey (USGS), but
Stratospheric Aerosol and Gas Experiment December 10, only NASA is responsible for launching all satel-
(SAGE III) 2001 lites for these agencies. The satellite missions
Gravity Recovery and Climate Experiment March 17, managed by NASA have been briefly presented
(GRACE) 2002 in Section 1.2.2, and USGS is currently managing
SeaWinds (ADEOS II) December 14,
the Landsat program. In the following, we will
2002 discuss the satellites operated by NOAA.
NOAA’s operational environmental satellite
Ice, Cloud, and land Elevation Satellite January 12,
(ICESat) 2003
system is composed of both geostationary and
polar-orbiting satellites. GOES satellites, as dis-
Polarization & Anisotropy of Reflectances December 4, cussed in Section 1.2.1, are mainly for national,
for Atmospheric Sciences coupled with 2006
Observations from a LiDAR (PARASOL)
regional, short-range warning, and “now-
casting.” Complementing the GOES geostation-
Aquarius June 10, 2011 ary satellites are the polar-orbiting satellites
ISS-Rapid Scatterometer (ISS-RapidScat) September 21, known as Polar Operational Environmental Sat-
2014 ellites (POES), Suomi National Polar-orbiting
Cloud-Aerosol Transport System on ISS January 10, Partnership (S-NPP), and Joint Polar Satellite
(CATS) 2015 System (JPSS) for global, long-term forecasting
and environmental monitoring. Both types of
satellite are necessary for providing a complete
engines and operate space probes. They are the global weather monitoring system.
China National Space Administration (CNSA), The POES system includes the Advanced
the European Space Agency (ESA), the Indian Very High Resolution Radiometer (AVHRR)
Space Research Organization (ISRO), the Japan and the Television Infrared Observation Satellite
Aerospace Exploration Agency (JAXA), the (TIROS) Operational Vertical Sounder (TOVS).
NASA, and the Russian Federal Space Agency The world’s first meteorological satellite, TIROS,
(RFSA or Roscosmos). was launched on April 1, 1960 and demonstrated
According to the Union of Concerned Scien- the advantage of mapping Earth’s cloud cover
tists Database: As of November 30, 2018, there from satellite altitudes.
are 1957 Earth-orbiting satellites (US 849, China On January 23, 1970, the first of the improved
284, Russia 152), 36% of these have a main TIROS Operational Satellite (ITOS) was
purpose of either Earth Observation (EO) or launched. Between December 11, 1970 and July
Earth Science. Among active EO satellites (620) 29, 1976, five ITOS satellites designated
in 2017, a massive increase by 66% from the NOAA-1 through 5 were launched. From
year 2016, their purposes can be grouped as October 13, 1978 to July 23, 1981, satellites in
following: 327 for optical imaging, 45 for radar the TIROS-N series were launched, where N rep-
imaging, 7 for infrared imaging, 7 for satellites, resents the next generation of operational satel-
64 for meteorology, and 60 for Earth Science. lites. NOAA-6 and NOAA-7 were also
 1.2 Platform and sensor systems 9
launched during this time frame. On March 28, satellites have been planned: JPSS-2 (2021),
1983, the first of the Advanced TIROS-N (or JPSS-3 (2026), and JPSS-4 (2031).
ATN) satellites, designated NOAA-8, was The Landsat missions have provided the long-
launched. NOAA continues to operate the ATN term land surface observations at fine spatial
series of satellites today with improved instru- resolutions (Fig. 1.4). On July 23, 1972, the Earth
ments. Complementing the geostationary satel- Resources Technology Satellite (ERTS-1) was
lites are two NOAA polar-orbiting satellites, launched and later renamed Landsat 1. The
one crossing the equator at 7:30 a.m. local time launches of Landsat 2, Landsat 3, Landsat 4,
and the other at 1:40 p.m. local time. The latest and Landsat 5 followed in 1975, 1978, 1982,
is NOAA-19, launched on February 6, 2009. and 1984, respectively. Landsat 5 provided the
NOAA-18 (PM secondary), NOAA-17 (AM Thematic Mapper (TM) imagery for 28 years
backup), NOAA-16 (PM secondary), and and 10 months; Landsat 6 failed to achieve orbit
NOAA-15 (AM secondary) all continue transmit- in 1993. Landsat 7 successfully launched in 1999
ting data as standby satellites. NOAA-19 is the with the ETM þ sensor, Landsat 8 in 2013, and
“operational” PM primary satellite, and both satellites continue to acquire data. The
METOP-A, owned and operated by EUMET- Landsat 9 satellite is expected to launch in
SAT, is the AM Primary satellite. December 2020.
The first AVHRR sensor was a 4-channel radi-
ometer, first carried on TIROS-N (launched 1.2.3.2 Europe
October 1978). This was subsequently improved The EUMETSAT Polar System mainly
to a 5-channel instrument (AVHRR/2) that was includes Metop that is a series of three polar-
initially carried on NOAA-7 (launched June orbiting meteorological satellites: Metop-A
1981). The latest instrument version is (launched on October 19, 2006), Metop-B
AVHRR/3, with six channels, first carried on (launched on September 17, 2012), and Metop-
NOAA-15, launched in May 1998. Multiple C (launched on November 7, 2018). They are
global vegetation index datasets have been currently operated in unison.
developed from NOAA-7 to now. The European Space Agency (ESA) has oper-
From 2011, NOAA has started the new JPSS ated a series polar-orbiting satellite, such as
program. JPSS is a collaborative program be- CryoSat, Soil Moisture and Ocean Salinity
tween the NOAA and NASA. This interagency (SMOS), and Envisat. Envisat was launched in
effort is the latest generation of US polar- 2002 with 10 instruments aboard but ended on
orbiting environmental satellites. The S-NPP April 08, 2012, following the unexpected loss of
satellite, launched in October 2011, is the prede- contact with the satellite. SMOS mission is a
cessor to the JPSS series spacecraft and is consid- radio telescope in orbit, but pointing back to
ered the bridge between NOAA’s legacy polar Earth not space. It was launched on November
satellite fleet, NASA’s EOS missions, and the 2, 2009. CryoSat is Europe’s first ice mission
JPSS constellation. S-NPP was constructed with with an advanced radar altimeter specifically
a design life of 5 years but is still functioning designed to monitor the most dynamic sections
normally. NOAA-20 (formerly JPSS-1), which of Earth’s cryosphere. CryoSat was launched
launched into space on November 18, 2017, is on April 8, 2010.
the first spacecraft of NOAA’s next generation ESA is developing the Sentinel satellite series.
of polar-orbiting satellites. Visible Infrared Imag- Each Sentinel mission is based on a constellation
ing Radiometer Suite (VIIRS) is very similar to of two satellites to fulfill revisit and coverage
MODIS. NOAA-20 carries five similar instru- requirements. These missions carry a range of
ments to the Suomi NPP. The following-on technologies, such as radar and multispectral
10 1. A systematic view of remote sensing

FIGURE 1.4 Landsat satellites timelines, downloaded from https://www.usgs.gov/land-resources/nli/landsat/ in

February 2019.

imaging instruments. Sentinel-1 is a polar- monitor the atmosphere from polar orbit aboard
orbiting, all-weather, day-and-night radar imag- a MetOp Second Generation satellite. Sentinel-6
ing mission for land and ocean service with carries a radar altimeter to measure global sea-
Sentinel-1A launched on April 3, 2014 and surface height, primarily for operational oceanog-
Sentinel-1B on April 25, 2016. Sentinel-2 is a raphy and for climate studies.
polar-orbiting, multispectral high-resolution Similar to the Landsat program, the SPOT
imaging mission for land monitoring, with program has also provided the long-term high-
Sentinel-2A launched on June 23, 2015 and resolution satellite observations. The Landsat
Sentinel-2B on March 7, 2017. Sentinel-3 is a mul- program has been mostly funded by the US
tiinstrument mission to measure sea surface government, but the SPOT program has been
topography, sea- and land surface temperature, operating commercially. The SPOT satellites are
and ocean color and land color. Sentinel-3A was summarized in Table 1.4. It is able to take stereo-
launched on February 16, 2016 and Sentinel-3B pair images almost simultaneously to map
on April 25, 2018. Sentinel-4 is for atmospheric surface topography.
monitoring that will be embarked upon a Meteo-
sat Third Generation-Sounder (MTG-S) satellite 1.2.3.3 China
in geostationary orbit. Sentinel-5 Precursor China has developed several satellite series,
(Sentinel-5P) is to provide timely data on a multi- such as meteorological satellite series Fengyun
tude of trace gases and aerosols. Sentinel-5P was (FY), ocean satellite series Haiyang (HY), Earth
launched on October 13, 2017. Sentinel-5 will resources satellite series Ziyuan (ZY),
 1.2 Platform and sensor systems 11
TABLE 1.4 Overview of the SPOT satellites and on October 21, 2003. ZY-1 02C was launched
data characteristics. on December 22, 2011. ZY-3 is China’s first
SPOT Launch Ending Spatial and spectral
high-resolution civilian optical transmission-
satellites date date resolutions type stereo mapping satellite that integrates the
functions of surveying, mapping, and resources
1 February December One 10 m panchromatic
investigation. ZY-3 is equipped with two front
22, 1986 31, 1990 band (0.51e0.73 mm);
three 20 m multispectral and back view CCD cameras having the resolu-
bands: green (0.50e0.59), tion better than 3.5 m, one CCD camera with
red (0.61e0.68 mm), near- the resolution better than 2.1 m, and one multi-
infrared (0.79e0.89 mm) spectral camera with the resolution better than
2 January July 2009 Same as SPOT 1 5.8 m. The swath is about 50 km.
22, 1990 The meteorological FY satellite series include
3 September November Same as SPOT-1 both geostationary (FY2 and FY4) and polar-
26, 1993 14, 1997 orbiting (FY-1 and FY-3) satellites. FY3 is the
second generation of the Chinese meterological
4 March 24, July 2013 One 10 m monospectral
1998 band (0.61e0.68 mm); polar-orbiting satellites. FY-3A was launched on
three 20 m multispectral May 27, 2008 and carried 11 sensors. FY-3D is
bands: green (0.50e0.59), the latest one in the series launched on November
red (0.61e0.68 mm), near- 15, 2017. Additional satellites in this series have
infrared (0.79e0.89 mm)
also been planned with FY-3E (2019), FY-3F
5 May 4, March 31, 2.5/5 m panchromatic (2019), and FY-3G (2022). FY-4A was launched
2002 2015 band; three 10 m on December 10, 2016, and additional five new
multispectral bands:
FY-4’s launches were also planned.
green (500e590 nm), red
(610e680 nm), near-IR Gaofen (GF), meaning high resolution in
(780e890 nm) bands, and Chinese, satellite series is part of the China
one 20m resolution on High-Resolution Earth Observation System
shortwave-infrared (1.58 (CHEOS), an analog to Europe’s Copernicus
e1.75 mm)
program of Sentinel Earth observation satellites.
6 September One 1.5 m panchromatic The first few satellites and some characteristics
9, 2012 band; four 6m are shown in Table 1.5.
multispectral bands: blue
(450e525 nm), green (530
e590 nm), red (625
e695 nm), near-infrared 1.2.4 Small satellites and satellite
(760e890 nm)
constellations
7 June 30, Same sensors as SPOT 6
2014 All satellites can be classified into seven clas-
ses based on their masses (see Table 1.6). There
environment and disaster monitoring small sat- is a trend in using small satellites for Earth obser-
ellite constellation (HJ). Their launch times are vation for reducing the cost: heavier satellites
shown in Fig. 1.5. require larger rockets with greater thrust, which
The resource ZY satellite series started with also has greater cost to finance. Small satellites,
the ChinaeBrazil Earth resource satellites also known as miniaturized satellites, are artifi-
(CBERS) jointly developed by China and Brazil cial satellites of low mass and size, usually under
with CBERS-1 launched in 1999 and CBERS-2 500 kg. Most of these small satellites have been
used with “mother” satellites that provide
12 1. A systematic view of remote sensing

FIGURE 1.5 Major Chinese satellites relevant to land remote sensing and their launch times (Liang et al., 2018).

operating signals; however, more recent versions cover might be a good example of a mission
are operating independently. Femto satellites where higher refresh rate at lower accuracy is a
and other types of small satellites are beginning preferred approach.
to revolutionize not only who can send satellite The Planet Labs, a private company based in
systems in space but also they have now given San Francisco, CA, USA, had launched 298 satel-
unprecedented access for data collection. lites, 150 of which were active, as of September
According to the Union of Concerned Scien- 2018. The company is operating several Earth
tists Database, all 620 active EO satellites in observation satellite constellations: Flock, Rapi-
2017 include 186 large satellites, 74 small satel- dEye, and Skysat. The Flock constellation consists
lites, 100 microsats, 215 Nanosats/CubeSats, of the Dove Cubesats that weigh 4 kg (8.8 lb),
and the remaining 45 satellites that do not have 10  10  30 cm (3.9 in  3.9 in  11.8 in) in
a launch mass specified. In particular, the num- length, width, and height. Each Dove satellite
ber of Nanosats/CubeSats increased by 34.68% is tiny and has a lifespan of 1e3 years but can
from 2016. observe the Earth at 3e5 m spatial resolution.
Satellite systems are generally transitioning The RapidEye constellation consists of five sat-
from the single satellite model to the cooperative ellites producing 5-m (16 ft) resolution imagery
sensing approach. For missions requiring global that Planet acquired from the German company
or continuous coverage in real time or within a BlackBridge in 2015. The five satellites travel on
very short temporal period, there is a potential the same orbital plane (at an altitude of
advantage in deploying a constellation of satel- 630 km) and together are capable of collecting
lites. A satellite constellation is a group of satel- over 4 million km of 5-m resolution, 5-band im-
lites operating in a coordinated format. The agery every day in the blue (440e510 nm), green
well-known example is the global positioning (520e590 nm), red (630e690 nm), red-edge
system (GPS) constellation (Fig. 1.6). (690e730 nm), and near infrared (760e880 nm).
Constellations of small satellites may offer a The Skysat constellation, purchased from
new approach to those science missions that Google in 2017, is composed of CubeSat that
would benefit from more frequent sampling by can observe the Earth surface at a spatial resolu-
a larger number of lower cost sensors. Moni- tion of 0.9 m in its 400e900 nm panchromatic
toring of time-varying phenomena such as cloud band, making it the smallest satellite to be put
 1.2 Platform and sensor systems 13
TABLE 1.5 Overview of the Gaofen first seven sat- in orbit capable of such high-resolution imagery.
ellites and data characteristics. The four multispectral bands have a spatial reso-
GF- Launch date Notes
lution of 2 m in blue (450e515 nm), green
(515e595 nm), red (605e695 nm), and near-
1 April 26, 2013 Two sensors: high-resolution cameras infrared (740e900 nm). As of September 2016,
(HRC) and wide field imagers (WFI).
six SkySat satellites were launched. In October
HRC includes pan at 2 m and four
multispectral bands (blue, green, red, 2017, four additional Dove satellites were also
and near-IR) at 8m with the swath of launched being part of this constellation.
68 km. WFI has similar four Besides the constellations with actual satel-
multispectral bands to HRC at 16m lites simultaneously orbiting in space, the
resolution with the swath of 830 km.
concept of virtual constellation has also been
Repeating cycle:  4 days at the equator
proposed. The CEOS defines virtual constella-
2 August A single camera: one 1 m Pan and 4 m tions as a “set of space and ground segment capabil-
19, 2014 multispectral bands (blue, green, red,
ities that operate in a coordinated manner to meet a
and near-IR). Swath: 45 km. Repeating
cycle: 4 days at equator combined and common set of Earth Observation
requirements.”
3 September A quad-polarization (vertical-vertical
We are increasingly faced the challenging dif-
8, 2016 (VV); horizontalhorizontal (HH);
vertical-horizontal (VH); ficulty to address rapid changes in the global
horizontalvertical (HV)) C-Band SAR at environment using data from single-satellite
25 m spatial resolution, a 26-day repeat sensors or platforms due to the underlying limi-
cycle tations of data availability and tradeoffs that
4 December A geostationary satellite with a camera govern the design and implementation of exist-
28, 2015 of 5 bands. The first four bands (blue, ing satellite systems. Virtual constellations can
green, red, and near-IR) at 50 m principally be used to add value to Earth obser-
resolution, the middle-IR (3.5e4.1 mm)
vation by combining sensors with similar spatial,
at 400m resolution. Swath: 400 km.
spectral, temporal, and radiometric characteris-
5 September Two hyperspectral/multispectral tics. Virtual constellations of planned and exist-
5, 2018 sensors for terrestrial earth observation
ing satellite sensors may help to overcome the
and four atmospheric observation
sensors: visible shortwave infrared limitation by combining existing observations
hyperspectral camera, full-spectrum to mitigate limitations of any one particular
spectral imager, atmospheric aerosol sensor. While multisensory applications are not
multiangle polarization detector, new, the integration and harmonization of multi-
atmospheric trace gases differential
sensor data is still challenging, requiring tremen-
absorption spectrometer, main
atmospheric greenhouse gases monitor, dous efforts of science and operational user
ultrahigh-resolution infrared communities.
atmospheric sounder The CEOS has formed seven virtual constella-
6 February Similar to the GF-1 satellite, but using a tions so far to coordinate space-based, ground-
6, 2018 different instrument suit, consisting of a based, and/or data delivery systems to meet a
2/8 m resolution panchromatic/ common set of requirements within a specific
hyperspectral camera and a 16 m domain, including Atmospheric Composition
resolution wide angle camera
(AC-VC), Land Surface Imaging (LSI-VC),
7 2019 Similar to ZY-3 with 3D topographical Ocean Color Radiometry (OCR-VC), Ocean
mapping Surface Topography (OST-VC), Ocean Surface

TABLE 1.6 Classification of EO satellite system.

Large Medium-sized Microsatellite or Nanosatellite or Picosatellite or Femtosatellite or

Satellites satellite satellite Minisatellitemicrosat nanosat picosat femtosat

Mass >1000 kg 500e1000 kg 100e500 kg 10e100 kg 1e10 kg 0.1ei kg <0.1 kg

14 1. A systematic view of remote sensing

FIGURE 1.6 Global positioning system satellite constellation, downloaded from https://upload.wikimedia.org/
wikipedia/commons/e/e2/GPS-constellation-3D-NOAA.jpg.

Vector Wind (OSVW-VC), Precipitation (P-VC), that is emitted by the object being viewed or re-
and Sea Surface Temperature (SST-VC). They flected by the object from a source other than the
leverage inter-Agency collaboration and part- instrument. Reflected sunlight is the most com-
nerships to address observational gaps, sustain mon external source of radiation sensed by pas-
the routine collection of critical observations, sive sensors. Typical passive sensors include the
and minimize duplication/overlaps, while following:
maintaining the independence of individual
• Radiometer: An instrument that
CEOS Agency contributions.
quantitatively measures the radiance of
electromagnetic radiation in the visible,
infrared, or microwave spectral region.
1.2.5 Sensor types • Imaging radiometer: A radiometer that
The sensor technology has been well includes a scanning capability to provide a
reviewed recently (Toth and Jozkow, 2016). two-dimensional array of pixels from which
There are two types of sensors: passive and an image may be produced. It is often called a
active. Passive sensors detect natural radiation scanner. Scanning can be performed
 1.2 Platform and sensor systems 15
mechanically or electronically by using an structure information, Section14.4.2 for
array of detectors. Across-track scanners, estimating above-ground forest biomass, and
scanning from one side of the sensor to the Section 18.3.2.2 for mapping soil moisture.
other across the platform flight direction • Interferometric synthetic-aperture radar
using a rotating mirror, are called (InSAR): A technique that compares two or
Whiskbroom Scanners, such as AVHRR. more amplitude and phase images over the
Alone-track scanners, scanning a swath with same geographic region received during
a linear array of charge-coupled devices different passes of the SAR platform at
(CCD) arranged perpendicular to the flight different times. InSAR can survey height
direction of the platform without using a information of the illuminated scene with
mechanical rotation device, are called cm-scale vertical resolution and 30-m pixel
Pushbroom scanners, such as High resolution, and covering areas 100  100 km
Resolution Visible of SPOT and Advanced (in standard beam modes). Examples include
Land Imager of EO-1. ERS-1 (1991), JERS-1 (1992), RADARSAT-1
• Spectroradiometer: A radiometer that can and ERS-2 (1995), and ASAR (2002). While the
measure the radiance in multiple spectral majority of InSAR missions to date have
bands, such as the Moderate Resolution utilized C-band sensors, recent missions such
Imaging Spectroradiometer (MODIS) and the as ALOS PALSAR, TerraSAR-X, and COSMO
Multi-angle Imaging SpectroRadiometer SKYMED are expanding the available data in
(MISR). the L- and X-bands.
• Scatterometer: A high frequency microwave
Active sensors provide their own electromag-
radar designed specifically to determine the
netic radiation to illuminate the scene they
normalized radar cross section of the surface.
observe. They send a pulse of energy from the
Over ocean surfaces, measurements of
sensor to the scene and then receive the radiation
backscattered radiation in the microwave
that is reflected or backscattered from that scene.
spectral region can be used to derive maps of
Typical active sensors include:
surface wind speed and direction. It has also
• Radar (Radio Detection and Ranging): A been used for mapping surface soil moisture
microwave radar that uses a transmitter and freeze/thaw states. Examples include the
operating at microwave frequencies to emit Advanced Microwave Instrument (AMI) of
electromagnetic radiation and a directional ERS-1 and ERS-2.
antenna or receiver to measure the time of • LiDAR: An active optical sensor that uses a
arrival of reflected or backscattered pulses of laser in the ultraviolet, visible, or near-
radiation from distant objects for determining infrared spectrum to transmit a light pulse
the distance to the object. and a receiver with sensitive detectors to
• SAR: A side-looking radar imaging system measure the backscattered or reflected light.
that uses relative motion between an antenna Distance to the object is determined by
and the Earth surface to synthesize a very recording the time between the transmitted
long antenna by combining signals (echoes) and backscattered pulses and using the speed
received by the radar as it moves along its of light to calculate the distance traveled. The
flight track for obtaining high spatial details are given in Chapters 13 and 14.
resolution imagery. There are multiple SAR • Laser Altimeter: A laser altimeter that uses a
systems in operation, and some examples can LiDAR to measure the height of the
be seen in Section 13.4 for estimating instrument platform above the surface. By
vegetation canopy height and vertical independently knowing the height of the
16 1. A systematic view of remote sensing

platform with respect to the mean Earth’s of view (IFOV) of the sensor, or the linear dimen-
surface, the topography of the underlying sion on the ground represented by each pixel.
surface can be determined. The Geoscience Fig. 1.7 shows the campus of the University of
Laser Altimeter System (GLAS) of ICESat is a Maryland at College Park at four different
typical example of a space-based Laser spatial resolutions. Table 1.7 shows the spatial
Altimeter. resolution of some common sensors.

1.2.6.2 Spectral resolution

1.2.6 Data characteristics The spectral resolution describes the number
The specifications of the platform and the and width of spectral bands in a sensor system.
sensor determine the resolutions of the remotely Many sensor systems have a panchromatic
sensed data: spatial, spectral, temporal, and band, which is one single wide band in the
radiometric. visible spectrum, and multispectral bands in
the visible-near-IR or thermal-IR spectrum see
1.2.6.1 Spatial resolution (Table 1.7). Hyperspectral systems usually have
Spatial resolution is a measure of the smallest hundreds of spectral narrow bands; for example,
object that can be resolved by the sensor, or the Hyperion on EO-1 satellite has 220 bands at 30-m
ground area imaged for the instantaneous field spatial resolution.

(A) 1 m (B) 10 m

(C) 30 m (D) 250 m

FIGURE 1.7 Campus of the University of Maryland at College Park at four Spatial resolutions.
 1.2 Platform and sensor systems 17
TABLE 1.7 Characteristics of some commonly used satellite sensors.

Spatial Radiometric Temporal Temporal

Satellite sensors Spectral bands resolution (m) resolution (bit) resolution (day) coverage

Coarse resolution POLDER B1eB9 6000*7000 12 4 POLDER 1:

(>1000 m) October 1996 to
June 1997
POLDER2: April
to October 2003
Medium MODIS B1eB2 250 12 Daily 1999
resolution (100
e1000 m) B3eB7 500
B8eB36 1000
AVHRR B1eB5 1100 at nadir 10 Daily

Fine resolution ALI/EO1

(5e100 m)
ASTER/Terra B1 15 8
B2eB9 30
B11eB14 90 12
ETMþ/Landsat 7 Pan 15 8 16 1999-
B1eB5,B7 30

B6 60
HRV/SPOT5 Pan 2.5 or 5 8 26/2.4 2002-
B1eB3 10
SW-IR 20
Very high Ikonos Panchromatic 0.82 at nadir 11 3 days at 40 1999-
resolution band degrees latitude
(<5 m)
B1-b4 3.2 at nadir
Quickbird Pan 0.61 11 1e3.5 2001-

B1eB4 2.44
World view Pan 0.5 at nadir 11 1.7e5.9 2007-
Geoeye-1 Pan 1.41 at nadir 11 2.1e8.3 days at 40 2008-
degrees latitude
B1eB4 1.65 at nadir

1.2.6.3 Temporal resolution The temporal resolutions of common sensors are

Temporal resolution is a measure of the repeat also shown in Table 1.7.
cycle or frequency with which a sensor revisits
the same part of the Earth’s surface. The fre- 1.2.6.4 Radiometric resolution
quency characteristics are determined by the Radiometric resolution refers to the dynamic
design of the satellite sensor and its orbit pattern. range, or the number of different output numbers
18 1. A systematic view of remote sensing

(A) 8 bits (256 levels) (B) 4 bits (16 levels)

(C) 2 bits (4 levels) (D) 1 bit (2 levels)

FIGURE 1.8 Campus of the University of Maryland at College Park at four radiometric resolutions.

in each band of data, and is determined by the (2) the data can be recorded on board the
number of bits into which the recorded radiation satellite for transmission to a GRS at a later
is divided. In 8-bit data, the digital numbers (DN) time; and
can range from 0 to 255 for each pixel (28 1/4 256 (3) the data can also be relayed to the GRS
total possible numbers). Obviously more bits re- through the Tracking and Data Relay
sults in higher radiometric accuracy of the sensor, Satellite (TDRS) System (TDRSS), which
as shown in Fig. 1.8. The radiometric resolutions consists of a series of communications
of common sensors are shown in Table 1.7. satellites in geosynchronous orbit. The data
are transmitted from one satellite to another
until they reach the appropriate GRS.
NASA’s TDRS started in the early 1970’s and
1.3 Data transmission and ground has evolved for three generations. The
receiving system current TDRSS consists of 10 in-orbit
satellites (four first generation, 3 s generation
There are three main options for transmitting and two third generation satellites)
data acquired by satellite sensors to the surface: distributed to provide near-constant
(1) the data can be directly transmitted to Earth communication links between the ground
if a Ground Receiving Station (GRS) is in the and orbiting satellites (e.g., Landsat).
line of sight of the satellite;
 1.4 Data processing 19
There are two types of GRSs: fixed and mo- 1.4 Data processing
bile. Most GRSs are fixed, and Fig. 1.9 shows
the locations of all currently active ground sta- A series of preprocessing tasks are needed to
tions operated by the United States (only two undertake before environmental information
in South Dakota and Aelaska) and International can be accurately extracted from remotely
Cooperator ground station network for the sensed data. Two types of preprocesses are
direct downlink and distribution of Landsat 7 conducted: radiometric processing and geomet-
(L7) and/or Landsat 8 (L8) image data. As ric processing. The radiometric processing may
coverage of the globe by ground receiving sta- include sensor radiometric calibration, image
tions is not complete, as seen from Fig. 1.9, the enhancement (mostly filtering noises), atmo-
mobile station is an attractive solution to fill spheric correction, and image fusion.
the holes and also an efficient means to perform
acquisition in a remote location for a long period
of time when a lot of images are needed for 1.4.1 Radiometric calibration
particular work (cartography of a region for
Radiometric calibration is a process that con-
example).
verts recorded sensor voltages or digital
The ground receiving stations acquire,
numbers (DN) to an absolute scale of radiance
preprocess, archive, and process data. Their
or reflectance. Because outer space is such a
typical components and functions may include
harsh environment, the performance of all satel-
the data acquisition facility, the data processing
lite sensors degrades over time. To achieve
facility, the value added facility, and user sup-
consistent and accurate measurements that can
port services.
be used to detect climatic and environmental

FIGURE 1.9 The locations of all active ground stations operated by our US and International Cooperator (IC) ground sta-
tion network for the direct downlink and distribution of Landsat 8 and/or Landsat 9 data. The circles show the approximate
area over which each station has the capability for direct reception of Landsat data. Downloaded from https://landsat.usgs.gov/igs-
network in February 2019.
20 1. A systematic view of remote sensing

change, the digital numbers (DNs) need to be reflective solar sensors. There are several
transformed into physical quantities. common desired characteristics of an
Calibration measurements can be conducted invariant sited for example, temporal
in three stages: preflight, in-flight, and post- stability, spatial uniformity, little or
launch: nonvegetation, and relatively high surface
reflectivity with approximately Lambertian
• Preflight calibration measures a sensor’s
reflectance. Commonly used sites include
radiometric properties before that sensor is
stable desert areas of the Sahara, Saudi
sent into space. Preflight instrument
Arabia, Sonoran, White Sand, and regions in
calibration is performed at the instrument
Bolivia. By observing these sites with satellite
builder’s facilities. The controllable and stable
sensor systems over extended periods of time,
environment in the laboratory guarantees
degradations (trends) in sensor responsivity
high calibration accuracy and precision
can be monitored and quantified.
• In-flight calibration is usually performed on a
routine basis with on-board calibration Some sensors have neither on-board calibra-
systems. More and more optical sensors have tion devices nor regular post-launch calibration,
on-board calibration devices. For example, for example, AVHRR. One solution to calibrate
the AVHRR optical sensor does not have an the sensor is through cross calibration. MODIS
on-board calibration capability, but the as a well calibrated instrument (Xiong et al.,
ETMþ has three on-board calibration devices: 2018) has been used as a reference to calibrate
the Internal Calibrator, the Partial Aperture other sensors using coincident observations of
Solar Calibrator, and the Full Aperture Solar MODIS and the target sensors over the pseudo-
Calibrator. MODIS also has three dedicated invariant calibration sites. For example, Vermote
calibration devices for the reflective bands: and Kaufman (1995) proposed a cross-
Solar Diffuser, Solar Diffuser Stability calibration method using a time series of MODIS
Monitor, and the Spectroradiometric and AVHRR data over a Saharan Desert site.
Calibration Assembly. In addition, MODIS These two methods using data over the
has two additional calibration techniques: pseudo-invariant calibration sites provide abso-
looking at the Moon and at deep space. The lute radiometric calibration of the sensors.
MODIS sensor has such an onboard Many sensors contain multiple detectors that
calibration system that promises an absolute have slightly different responsivities. As a result,
error better than 2%. the imagery produced by these sensors may
• Post-launch calibration data have to be contain a significant level of striping. One solu-
obtained from vicarious calibration tion is to match the mean values of each detector
techniques that typically make use of selected over a period of times so that all detectors pro-
natural or artificial sites on the surface of the duce the relatively uniform values.
Earth. Prelaunch and onboard methods are
better established, and postlaunch methods
using invariant sites in vicarious calibration is 1.4.2 Geometric processing
becoming more popular with the changing
No image acquired by sensors can perfectly
design and demands of new instruments.
represent the true spatial properties of the land-
Vicarious calibration using pseudo-invariant
scape. Many factors can also distort the geomet-
sites has become increasingly accepted as a
ric properties of remote sensing data, such as
fundamental postlaunch calibration method
variations in the platform altitude, attitude and
to monitor long-term performance of satellite
velocity, Earth rotation and curvature, surface
 1.4 Data processing 21

(A) (B)
TGSD
1

0.8 Ideal PSF 1

0.8

0.6 0.6

PSF
footprint decided
PSF

0.4 by TGSD
0.4 0.2

Realistic PSF 0
0.2 10
5 10
0 5
Y( 0
km –5 –5
0 ) )
–20 –15 –10 –5 0 5 10 15 20 –10 –10 X (km
X distance from scene center (km)

FIGURE 1.10 Point spread functions: (A) Realistic and ideal PSF model for simulating GEO radiances, (B) Sketch of EE in
terms of the PSF, Dashed square shows the integral area for EE, which is determined by the TGSD (Zhang et al., 2006).

ACK
GTR
ALON
N
GSCA
ALON

Nadir

–
a–
– Major axis of scan spot ellipse
a
b b–
– Minor axis of scan spot ellipse
–

FIGURE 1.11 Sketch of pixel geometry for the AVHRR for adjacent scan lines to illustrate autocorrelation (Breaker, 1990).

relief displacement, and perspective projection. called the modulation transfer function (MTF), a
Some of these resulting distortions are system- precise measurement of details and contrast
atic and can be corrected through analysis of made in the frequency domain. The sensor PSF
sensor characteristics and platform ephemeris is often modeled as a Gaussian. Fig. 1.10 illus-
data, but others are random and have to be cor- trates the PSF in two- and three-dimensions
rected by using ground control points (DCP). (Fig. 1.11), where TGSD is the threshold ground
In the sensor ground instantaneous field of sample distance, which is the centroid-to-
view (IFOV), surface elements do not contribute centroid distance between adjacent pixels.
to the pixel value equally, but rather, the central The actual response function of the ground
part contributes most to the pixel value. This IFOV is often not square; for example, for
kind of spatial effect is usually specified by the MODIS, it is twice as wide cross-track as
sensor point spread function (PSF) in the spatial in-track because of time integration during scan-
domain, and the Fourier transform of the PSF is ning. For most whiskbroom scanners, such as
22 1. A systematic view of remote sensing

FIGURE 1.12 The actual size of the ground IFOV of AVHRR as a function of the view zenith angle.

AVHRR and MODIS, the actual size of the example, dropped lines are usually filled with
ground IFOV is a function of the scanning angle the values of the previous lines or the averages
(see Fig. 1.12). of the neighboring lines. Strips can be removed
This topic is significant because level-1 radi- by using simple along-line convolution, high-
ance data or level-2 reflectance data should be pass filtering, and forward and reverse principal
corrected for geometric distortions before calcu- component transformations.
lating geophysical parameters in order to obtain To assist human visual interpretation, various
a truly absolute geophysical parameter. The image enhancement techniques have been incor-
details are discussed in Chapter 2. porated in many remote sensing digital image
processing systems. These enhancement
methods can be divided into spatial domain
1.4.3 Image quality enhancement and frequency domain categories. In spatial
domain techniques, we directly deal with the im-
Imperfections or image artifacts are continu-
age pixels. Fig. 1.13 illustrates the effects of a
ously caused by the instrument’s electronics,
linear enhancement technique. The pixel values
dead or dying detectors, and downlink errors.
are manipulated to achieve the desired enhance-
Known artifacts include the scan-correlated shift, ment. In frequency domain methods, the image
memory effect, modulation transfer function, is first transferred to the frequency domain.
and coherent noise. Dropped lines and That is, the Fourier transform of the image is
in-operable detectors also exist as a result of computed first, all the enhancement operations
decommutating errors and detector failure.
are performed on the Fourier transform of the
Potential remnant artifacts include banding and
image, and then the inverse Fourier transform
striping. In the past, these effects were ignored
is performed to obtain the resultant image.
or artificially removed using cosmetic algo-
rithms during radiometric preprocessing. For
 1.4 Data processing 23

(A) (B)

2.0×104

1.5×104

1.0×104

5.0×103

0
500 1000 1500 2000
Data value

(C) (D)

8000

6000

4000

2000

0
500 1000 1500 2000
Data value

FIGURE 1.13 An example of linear enhancement: original image and its histogram (A) and (B); linearly enhanced image
and its histogram (C) and (D). Note that the radiometric properties characterizing environmental conditions are artificially
altered by image enhancement methods. Most image enhancement techniques for assisting visual interpretation should not
be performed before quantitatively estimating biophysical variables. (A) Original image. (B) Histogram. (C) Enhanced image.
(D) New histogram.

1.4.4 Atmospheric correction applications. Various cloud and shadow detec-

tion algorithms have been developed, and cloud
Since the observed radiance recorded by a mask is one of the high-level atmospheric prod-
spaceborne or airborne sensor contains both ucts. However, this is still an active research
atmospheric and surface information, atmo- area, and more effective and reliable algorithms
spheric effects must be removed to estimate are needed. For km coarse-resolution imagery
land surface biogeophysical variables, particu- (e.g., AVHRR and MODIS), there usually remain
larly from the reflective and thermal IR data, many cloudy or mixed cloudy pixels after
since microwave signals are not very sensitive applying the cloud mask. Various solutions
to changes in atmospheric conditions. have been used to address this issue. One solu-
Clouds in the atmosphere largely block Earth tion relies on temporal compositing techniques,
surface information and make most optical and converting daily observations to weekly or
thermal-IR imagery useless for terrestrial
24 1. A systematic view of remote sensing

satellite imagery, different methods (Sun et al.,

2017b), such as the threshold, radiative transfer
and statistical methods have been developed to
identify the clouds and the associated shadows.
One example is shown in Fig. 1.14.
For optical imagery, both aerosol and water
vapor scatter and absorb the radiation reflected
from the surface. There are two approaches for
atmospheric correction:
• The first assumes known atmospheric
properties, usually the total amounts of
aerosol and water vapor in the atmospheric
column, which may be estimated from other
sensors and/or other sources. Many
atmospheric radiative transfer codes (e.g.,
MODTRAN, 6S) can be used to calculate the
quantities required for atmospheric
correction.
• The second relies only on the imagery itself
FIGURE 1.14 Examples of cloud detection. The left col- without any external information.
umn shows two false-color composite imagery of Landsat
data and the detected clouds are shown in the right column If the atmospheric information can be accu-
(Sun et al., 2017b). rately estimated from other sources, the first
approach is preferable, but quite often we are
monthly data based on maximum vegetation in- not that lucky. Fig. 1.15 shows the significant dif-
dex or other criteria; other solutions include ferences of MODIS surface reflectance before and
replacement of these contaminated pixels using after atmospheric correction (Liang et al., 2002).
smoothing algorithms. This topic will be This topic will be discussed in Chapter 4.
discussed in Chapter 3. For high-resolution

FIGURE 1.15 An example of atmospheric correction of MODIS imagery before (A) and after (B) (Gui et al., 2010). From
Liang, S., Fang, H., Chen, M., Shuey, C., Walthall, C., Daughtry, C., 2002. Atmospheric correction of landsat ETMþ land surface
imagery: II. validation and applications. IEEE Trans. Geosci. Remote Sens. 40, 2736e2746. © 2002, IEEE.
 1.5 Mapping category variables 25
For thermal-IR imagery, if we can acquire at- • Multispatial images from the same or multiple
mospheric profile information (mainly tempera- sensors (e.g., merge ETM panchromatic and
ture and water vapor) from sounding data, multispectral images);
atmospheric correction is straightforward. The • Multiple images of different spectral regions
split-window approach based on two thermal- from the same or multiple sensors (e.g., merge
IR bands, when no such atmospheric profile SAR with optical imagery or visible bands
information is available, is often used to estimate with thermal bands);
land surface temperature without atmospheric • Remote sensing images with ancillary data
correction. The details are available in Chapter 7. (e.g., topographic map).
When the high-level satellite products are
evaluated, it is surprising to see that most prod-
1.4.5 Image fusion and product ucts are mainly generated from a single sensor.
integration For example, the MODIS albedo product is
mainly from MODIS data, which is also true
There are many cases where we need to inte- for MISR, MERIS, etc. The same product from
grate image data through image fusion tech- different satellite sensors may have different
niques. Definitions of image fusion in the characteristics (e.g., spatial and temporal resolu-
literature are very diverse. Image fusion can be tions, accuracy). Instead of asking the user to
viewed as a process that produces a single image pick the “best” product, we can generate a
from a set of input images. The fused image blended/integrated product from multiple-
should have more complete information and is sensor products. Chapter 21 is devoted to
more useful for estimating land surface vari- addressing this topic in the example of leaf
ables. It can improve both reliability by using area index (LAI).
redundant information and capability by using
complementary information, as illustrated in
Fig. 1.16. 1.5 Mapping category variables
Image fusion is not distinguished from image
merging or image integration, which at the pixel We are interested in two types of land surface
level may be in many different forms, for variables: category and quantitative. The cate-
example: gory variables represent the types of objects on
• Multitemporal images from the same or the land surfaces and are usually mapped out
multiple sensors for change detection (e.g., through image classification. The purpose of im-
merge TM images acquired at different age classification is to group together pixels that
times); have similar properties into a finite set of classes.
An example of a classified image is a land cover
map. Fig. 1.17 is a global land cover map map-
ped from MODIS data. The key steps in the clas-
sification process are as follows:
(1) Definition of classification system (scheme):
This depends on the objective and the
characteristics of the remote sensing data.
The purpose of such a scheme is to provide
(2) A framework for organizing and
FIGURE 1.16 Illustration of image fusion. categorizing the information that can be
 26
1. A systematic view of remote sensing
FIGURE 1.17 Global land cover classification map from MODIS.
 1.6 Estimating quantitative variables 27
extracted from the data. A number of (6) Accuracy assessment: The classified results
classification schemes have been developed should be checked and verified for their
for mapping regional and global land cover accuracy and reliability. The training data
and land use maps. The IGBP land cover are usually divided into two parts, one for
classification system for global mapping training and the other for validation. In the
using MODIS data is shown in Fig. 1.17. evaluation of classification errors, a
(3) Selection of features: Classification is classification error matrix is typically
executed based on a series of features in the formed, which is sometimes called a
feature space. It divides the feature space confusion matrix or contingency table.
into several classes based on a decision rule.
The details of image classification techniques
Instead of using the original bands, they are
are not covered by this book, but the basic prin-
often transformed into feature space to
ciples and progress can be found elsewhere
discriminate between the classes. Examples
(Dash and Ogutu, 2016; Lu and Weng, 2007).
of features include various vegetation
Some typical techniques for mapping land use
indexes, principal components and those
types will be discussed in Chapters 23e25. At
from the Tasseled-Cap transformation, and
most spatial resolutions, the majority of pixels
other spatial, temporal, and angular features.
are mixed. If a pixel is required not just to be
The subset of features is selected to
labeled as one of the cover types but to estimate
maximally distinguish different classes.
the percentages of the cover types, it would be
(4) Sampling of training data: Training is the
more challenging. How to estimate the fractional
process of defining the criteria by which these
vegetation coverage within one pixel will be dis-
classes are recognized and is performed with
cussed in Chapter 12.
either a supervised or an unsupervised
method. Supervised training is closely
controlled by the analyst, who selects pixels
from each class based on high-resolution 1.6 Estimating quantitative variables
imagery, ground truth data, or maps, while
unsupervised training is more computer- To drive, calibrate, and validate the Earth pro-
automated and enables the user to specify cess models and support various applications,
some parameters that the computer uses to high-level products of quantitative variables
uncover statistical patterns that are inherent are much more desirable. How to generate these
in the data but do not necessarily correspond products is the main focus of this book. In the
to classes in the classification scheme. early stages of remote sensing technique devel-
(5) Classification: A parametric or opment, visual interpretation was the approach
nonparametric decision rule, which is often commonly used for extracting land surface infor-
called a classifier, is used to perform the mation. Statistical analysis later became a more
actual sorting of pixels into distinct class common method for quantitatively estimating
values. There are various classifiers, such as land surface information. As can be seen in the
the parallelepiped classifier, minimum following chapters, various inversion techniques
distance classifier, maximum likelihood based on physically based surface radiation
classifier, regression tree classifier, and models have become the subject of mainstream
support vector machine (SVM) classifier. research (Liang, 2007). It is necessary to provide
They are compared with the training data so an overview of these techniques. As many inver-
that an appropriate decision rule is selected sion algorithms are based on forward radiation
for classification. modeling, let us first begin with that.
28 1. A systematic view of remote sensing

1.6.1 Forward radiation modeling ellipsoid shape and the calculated sunlit and
shadowed components.
This is the process that links the pixel values Turbid-medium radiative transfer models
of an image with surface characteristics through treat surface elements (leaf or soil particle) as
mathematical models (Liang, 2004). We will small absorbing and scattering particles with
mainly present landscape generation, surface given optical properties, distributed randomly
and atmosphere radiative transfer modeling, in the scene and oriented in given directions. In
and sensor models. one-dimensional canopy models(Kuusk, 1995;
Liang and Strahler, 1993b; Liang and Town-
1.6.1.1 Scene generation shend, 1996; Verhoef, 1984), canopy elements
Scene generation is a quantitative description are assumed to be randomly distributed, but
of our understanding of the landscape. Strahler three-dimensional RT models (Kuusk, 2018;
et al. (1986) identify two different scene models Myneni et al., 1989) can take into account the
in remote sensing: H- and L-resolution models. structural information of the landscape, as
H-resolution models are applicable where the el- shown in Fig. 1.19. The further development of
ements of the scene are larger than the pixel size, geometric optical models has incorporated radi-
and L-resolution models are applicable when the ative transfer theory in calculating the individual
converse is true. H-resolution scenes can be sunlit/shadow components; the resulting
generated using computer graphics techniques; models are often called hybrid models. In com-
for example, a vegetation canopy can be created puter simulation models, the arrangement and
with Onyx software (http://www.onyxtree. orientation of scene elements are simulated on
com/). L-resolution scenes can be generated a computer, and the radiation properties are
using mathematical models or GIS (geographic determined based on the radiosity equations
information system) techniques. (Borel et al., 1991; Huang et al., 2013; Qin and
Gerstl, 2000) and/or Monte Carlo ray tracing
(Disney et al., 2006; Gastellu-Etchegorry et al.,
1.6.1.2 Surface radiation modeling 2015; Lewis, 1999; North, 1996; Qi et al., 2019)
Given landscape composition and its optical methods. Fig. 1.20 compares a photo of grass
properties, we can predict the radiation field. field and the simulated field using the Botanical
Three types of models characterize the radiation Plant Modeling System (Lewis, 1999) based on
field of the scene, and they are commonly used the ray-tracing technique.
in optical remote sensing: geometric optical
models, turbid-medium radiative transfer 1.6.1.3 Atmospheric radiative transfer
models, and computer simulation models. The radiation at the Earth’s surface is
In geometric optical models (Li and Strahler, disturbed by the atmosphere before being
1985, 1986), canopy or soil is assumed to consist captured by the sensor in the atmosphere
of geometric protrusions with prescribed shapes (airborne sensors) or above the atmosphere
(e.g., cylinder, sphere, cones, ellipsoid, (spaceborne sensors).
spheroid), dimensions, and optical properties Atmospheric gases, aerosols, and clouds scat-
that are distributed on a background surface in ter and absorb the incoming solar radiation and
a defined manner (regularly or randomly distrib- the reflected and/or emitted radiation from the
uted). The total pixel value is the weighted surface. As a result, the atmosphere greatly mod-
average of sunlit crown, sunlit ground, shad- ulates the spectral dependence and spatial distri-
owed crown, and shadowed ground. Fig. 1.18 il- bution of the surface radiation. The atmospheric
lustrates a simulated canopy field with an radiative transfer theory is quite mature, and
 1.6 Estimating quantitative variables 29

FIGURE 1.18 DART (discrete anisotropic radiative transfer) (Gastellu-Etchegorry, 2008).

(A) (B)

Sunlit
r
Projecon crown shadowed
b

(shadowed) crown
h
A(

shadowed
ground

FIGURE 1.19 Principles of the geometric-optical model (A) for the canopy with an ellipsoid crown and the simulated
canopy field (B).
30 1. A systematic view of remote sensing

(A) (B)

FIGURE 1.20 A photo of an actual canopy (A) and the simulated canopy field (B). Courtesy from Dr. Mathias Disney at Uni-
versity College London.

many computer software packages (e.g., MOD- covariance matrices. Radiometric noise sources
TRAN, 6S) have been developed to enable us come from the detector and electronics. Detector
to calculate all necessary quantities, such as noise includes photon (shot) noise, thermal
path radiance and transmittance. noise, and multiplexer/readout noise. Because
detector parameters are often specified in terms
1.6.1.4 Sensor modeling of electrons, the noise terms are summed in a
root sum squared sense in that domain before
As common detector materials do not
being converted to noise equivalent spectral
respond across the entire optical spectrum,
radiance. The noise processes originating in the
sensors have separate focal planes and noise
electronics include quantization noise, bit errors
mechanisms for each spectral region. The sensor
(in recording or transmitting the data), and noise
model (Kerekes and Baum, 2005) can describe
arising within the electrical components.
the effects of an imaging spectrometer on the
Besides the sensor spectral response function
spectral radiance mean and covariance statistics
and radiometric noise, the sensor model can
of a land surface. The input radiance statistics
also include spatial effects using PSF and MTF.
of every spectral channel are modified by elec-
tronic gain, radiometric noise sources, and
relative calibration error to produce radiance
signal statistics that represent the scene as
1.6.2 Inversion methods
imaged by the sensor. Development of the inversion algorithms for
The sensor model includes approximations estimating quantitative variables has a long
for the spectral response functions and radio- history, as shown in Fig. 1.21. The early days
metric noise sources. The spectral response func- focused on statistical methods and were then
tions of each instrument can be measured and followed by physically based methods. Statisti-
provided by the sensor manufacturers. Radio- cal methods are mainly based on a variety of
metric noise processes are modeled by adding vegetation indices through regression analysis.
variance to the diagonal entries of the spectral Physical algorithms rely on inverting surface
 1.6 Estimating quantitative variables 31

FIGURE 1.21 Major developmental milestones of quantitative remote sensing. From Liang, S., Liu, Q., Yan, G., Shi, J.,
Kerekes, J.P., 2019. Foreword to the special issue on the recent progress in quantitative land remote sensing: modeling and estimation.
IEEE J. Select. Topics Appl. Earth Obser. Remote Sens. 12, 391e395, ©2019, IEEE.

radiation models. A new trend is to combine sta- the Multiple Adaptive Regression Spline
tistical and physical methods, and Section 1.6.2.4 (MARS). They are mainly used in two respects.
on the direct estimation methods presents such The first is to simplify the complex physical
an example. model to execute the forward simulations in an
inversion procedure. Complex models (such as
1.6.2.1 Statistical analysis and machine atmospheric/surface radiative transfer models)
learning techniques are computationally expensive for repeated for-
Statistical models have proven to be very use- ward simulations. If the inversion process needs
ful in various remote sensing applications. They such a model, the machine learning technique
are usually created using ground measurements. can be used to replace it. The second is directly
Because it is very expensive to collect extensive used for inversion. Based on the remote sensing
ground measurements under various conditions, data and the corresponding surface measure-
the major weakness of models based on ground ments as the input and output data pairs, a ma-
measurements is limited representation. An chine learning model can be established. Such
alternative solution is to simulate remotely input and output data pairs can also be gener-
sensed data using a physically based radiation ated by model simulations.
model that may have been calibrated and vali-
dated by field measurements. The key modeling 1.6.2.1.1 Artificial neural network
components have been discussed in Section ANNs are complex computational models
1.6.1.2. inspired by the human nervous system, which
Different statistical methods can be used to have the ability to learn from the training data
relate inputs and outputs of the model simula- to predict outcomes for new data. ANNs are
tions. Besides the conventional multivariate probably the most widely used machine learning
regression analysis, different machine learning method for estimating land surface variables in
methods have been used, such as artificial neural the past decade, and use of ANNs has been
networks (ANNs), SVM, self-organizing map steadily increasing.
(SOM), decision tree, random forest, case-based There are many different types of ANNS for
reasoning, neuro-fuzzy, genetic algorithm, and estimating land surface biogeophysical
32 1. A systematic view of remote sensing

variables, such as multilayer perceptron (MLP), It is desirable to extract knowledge (e.g.,

adaptive resonance theory, SOMs, radial basis explicit mathematical function) from trained
function (RBF), and recurrent neural network, neural networks for users to gain a better under-
but MLP is mostly used based on the survey standing of how the networks solve the
by Mas and Flores (2008). Although ANNs are nonlinear regression problems. Chan and Chan
considered to be “black boxes,” we still have (2017) developed the piecewise linear artificial
some flexibility, such as selecting input vari- neural network (PWL-ANN) algorithm, which
ables, determining appropriate characteristics is to “open up” the black box of a trained neural
for the training data, and optimizing the archi- network model so that rules in the form of linear
tecture of the network (number of layers and equations are generated by approximating the
nodes) and the method to avoid overtraining. sigmoid activation functions of the hidden neu-
Overfitting is often an issue. Training samples rons in an ANN. Setiono et al. (2002) described
are usually erroneous due to observational un- a method called rule extraction from function
certainties. Overfitting the training samples approximating neural networks (REF-ANN) for
may degrade the prediction ability of an ANN extracting rules from trained neural networks
model. An overtrained network is able to fit for nonlinear function approximation or
the training data precisely but fails for new data- regression.
sets with subtle changes. There are some tech- ANNs that are composed of many layers are
niques to address this issue, such as the referred to deep learning. Deep learning has
EBaLM-OTR (error back propagation and proven to be both a major breakthrough and
LevenbergeMarquardt algorithms for over an extremely powerful tool in applications
training resilience) technique (Wijayasekara where the target function is very complex and
et al., 2011). This method uses k-fold cross- the datasets are large. It has become an increas-
validation to determine the best architecture by ingly important method in remote sensing
dividing all data into three parts for training, (Zhang et al., 2016; Zhu et al., 2017a). The appli-
validating, and testing each network architec- cations have been mostly for image classification
ture. Piotrowski and Napiorkowski (2013) evalu- and data fusion, but some studies on quantita-
ated multiple methods for catchment runoff tive inversion of land surface variables have
modeling, such as the early stopping, the noise also been reported, such as soil moisture upscal-
injection, the weight decay, and optimized ing (Zhang et al., 2017a), land surface tempera-
approximation. These methods could be poten- ture (LST) (Yang et al., 2010),
tially used for estimating land surface evapotranspiration (ET) (Chen et al., 2013), pre-
parameters. cipitation (Tao et al., 2016b), vegetation coverage
Determining network parameters (e.g., num- (Jia et al., 2015), surface net radiation (Jiang et al.,
ber of hidden layers and the number of neutrons 2014), water quality optical parameters (Chen
in hidden layer) is often done by a trial-and-error et al., 2014; Jamet et al., 2012), downward solar
method, which is empirical, is time-consuming, radiation (Tang et al., 2016), soil moisture (Xing
and results in the network configuration that et al., 2017), and so on.
may not be optimal. Some techniques have
been proposed, such as the genetic algorithms 1.6.2.1.2 Support vector machine
(Castillo et al., 2000) and the particle swarm opti- Support vector regression (SVR), which is also
mization (Da and Xiurun, 2005; Liu et al., 2015b), used in regression analysis, is a kind of kernel-
which could be potentially used for remote based algorithm in machine learning. It mini-
sensing inversion. mizes the training error and the complexity of
the model and uses the nonlinear kernel function
 1.6 Estimating quantitative variables 33
that transforms the input data into a high- data and large datasets. Because random forest
dimensional feature space. Compared with other methods are less sensitive to the noise in the
machine learning methods, SVR is more suitable training dataset, it is better in the parameter esti-
for high-dimensional nonlinear problems of mation than the conventional regression tree
small training samples. method.
Mountrakis et al. (2011) reviewed its wide ap- Random forest methods have been widely
plications in remote sensing, although more used for image classification (Belgiu and Dragut,
examples of applications are on image classifica- 2016), but more and more is used in estimating
tion. Application examples of quantitative inver- quantitative parameters, such as grass nutrients
sion include leaf nitrogen concentration in rice and biomass (Ramoelo et al., 2015), nitrogen con-
(Du et al., 2016; Sun et al., 2017a), LAI (Zhu centration in rice (Sun et al., 2017a), LAI (Beck-
et al., 2017b), evapotranspiration (Ke et al., sch€afer et al., 2014; Li et al., 2017; Yuan et al.,
2016), nradiation (Jiang et al., 2016a), water 2017; Zhu et al., 2017b), vegetation coverage
depth and turbidity (Pan et al., 2015), grassland (Halperin et al., 2016), biomass (Lopez-Serrano
biomass (Marabel and Alvarez-Taboada, 2013), et al., 2016; Mutanga et al., 2012; Wang et al.,
leaf nitrogen concentration (Omer et al., 2017; 2016; Xia et al., 2018), surface temperature drop
Wang et al., 2017; Yao et al., 2015), soil moisture (Hutengs and Vohland, 2016), GPP (Tramontana
(Ahmad et al., 2010), and other chemical param- et al., 2015), snow depth (Tinkham et al., 2014),
eters (Axelsson et al., 2013). precipitation (Kuhnlein et al., 2014), forest
coverage and height (Ahmed et al., 2015), forest
1.6.2.1.3 Regression tree aboveground biomass (Karlson et al., 2015;
Regression trees and classification trees are Pflugmacher et al., 2014; Tanase et al., 2014),
often introduced together, but the main differ- and other forest parameters (Garcia-Gutierrez
ence is to separate the type or estimate the value et al., 2015).
of the variable. The tree consists of a root node RF is a robust nonlinear algorithm for predict-
(containing all data), a set of internal nodes ing forest biophysical parameters. However, one
(splits), and a set of terminal nodes (leaves). A drawback of RF regression algorithm when
recursive segmentation algorithm is used to many input predictors are utilized is that it se-
reduce the entropy within the class. The input lects predictors that could be correlated to each
data are hierarchical. The value of the internal other (Omer et al., 2016).
node depends on the predicted average of each
terminal node. Examples of quantitative inver- 1.6.2.1.5 Multiple adaptive regression spline
sion include biomass estimation (Blackard function
et al., 2008), forest structural parameters (Gomez MARS is a nonlinear and nonparametric
et al., 2012; Mora et al., 2010), and forest ground regression model. It is an extension of stepwise
coverage (Donmez et al., 2015). linear regression for adapting nonlinear regres-
sion. It is more flexible in establishing additive
1.6.2.1.4 Random forest relationship or very strong relationship between
Random forest methods have been con- the variables. It can be seen as an extended linear
structed with countless small regression trees model that automatically simulates nonlinearity
to predict the variable. These small regression and the interaction among variables, so it has
tree pruning are based on another random sam- high computational efficiency.
ple subset of the training dataset. Random forest Examples of applications include atmospheric
methods can effectively overcome the problem correction to invert surface reflectance (Kuter
of overfitting in regression, as well as “noise” et al., 2015), soil salinity (Nawar et al., 2014),
34 1. A systematic view of remote sensing

land surface net radiation (Jiang et al., 2016a), approach. One of the major reasons is its compu-
aboveground biomass (Filippi et al., 2014), soil tational cost because it iteratively runs the RT
organic carbon content (Liess et al., 2016), chloro- model in the inversion process. The recent efforts
phyll concentration (Gholizadeh et al., 2015), aim to develop various emulation techniques,
and snow cover fraction (Kuter et al., 2018). i.e., using emulators that are computationally
less expensive to replace the original RT model
1.6.2.2 Optimization algorithms and provide an approximation of the model
output trajectory. The applications of emulators
Optimization inversion is the primary inver-
have emerged in many disciplines (Castelletti
sion method for estimating land surface vari-
et al., 2012; Conti et al., 2009; Lucia et al., 2004;
ables in early days (Liang, 2004). It is based on
Machac et al., 2016). There are two distinct
a physically based model (such as canopy radia-
emulation approaches: dynamic and statistical.
tive transfer model) through the iterative
A dynamic emulator is a simpler representation
method to continuously adjust the model param-
of the original dynamic model. For example,
eters (x) so that the difference between the model
Xiao et al. (2015) replace the land surface
calculated value H(x) based on one or multiple
dynamic model (e.g., dynamic vegetation
radiative transfer (RT) models and satellite ob-
model) that can be very complex by a simple dif-
servations (y) is shrinking. The numerical differ-
ferential equation characterized by LAI for esti-
ence is usually expressed by the following cost
mating multiple land surface parameters
function J(x). The iterative process is to minimize
simultaneously. Statistical emulators relate the
the cost (merit) function:
input parameters to the model output statisti-
JðxÞ ¼ ðHðxÞ
yÞT R
1 ðHðxÞ
yÞ þ J0 (1.1) cally based on extensive model simulations by
varying all key parameters. Verrelst et al.
where R is the covariance matrix of the observa- (2016) explored three machine learning regres-
tion errors, J0 is the constrain term forcing that sion algorithms (kernel ridge regression, neural
the estimates are as close to the background networks, and Gaussian processes regression)
values (or ranges) as possible. These background to approximate the functions of three RT models
values/ranges may come from the climatologies (the leaf RT model PROSPECT-4, the canopy RT
of existing satellite data, field measurements, or model PROSAIL, and the atmospheric RT model
other priori knowledge. MODTRAN5) for global sensitivity analysis. It
There are some examples in using this turns out the sensitivity study taking for over a
approach. It has been widely used for estimating month using the original RT model can be
canopy and soil properties in the early years done in a few minutes. Gomez-Dans et al.
(Goel and Grier, 1987; Liang and Strahler (2016) developed the statistical Gaussian Process
1993a, 1994) and land surface temperature emulators: surrogate functions that accurately
(Liang, 2001). He et al. (2012) applied the optimi- approximate canopy RT models (PROSAIL and
zation method to estimate aerosol optical depth SEMIDISCRETE) and atmospheric RT model
and surface bidirectional reflectance distribution (6S). The developed emulators can provide a
function (BRDF) parameters, which further lead fast and easy route to estimating the Jacobian
to surface albedo calculation, from MODIS of the original model, enabling the use of some
TOA observations. Zhang et al. (2018) further optimization algorithms (e.g., efficient gradient
extended this method and estimated incident so- descent methods). Rivera et al. (2015) developed
lar radiation as well. a statistical emulator toolbox enabling multiout-
Global land surface satellite products have put machine learning regression algorithms
not been generated by using the optimization available in MATLAB to approximate the RT
 1.6 Estimating quantitative variables 35
model called SCOPE. Akbar and Moghaddam Because of its simplicity and easy implemen-
(2015) applied the global optimization method tation, the LUT methods have continued to be
with a regulation term to estimate soil moisture widely used, and the recent examples include
from the combined active radar and passive retrieving LAI (Banskota et al., 2015; Qu et al.,
radiometer microwave data at the same spatial 2014b; Tao et al., 2016a; Yang et al., 2017) and
resolution. The similar scheme is also applied chlorophyll concentration (Darvishzadeh et al.
for estimate soil moisture and roughness (Akbar 2008, 2012), crop biophysical parameters from
et al., 2017). LiDAR data (Ben Hmida et al., 2017), fraction
of vegetation cover (Ding et al., 2016), grassland
1.6.2.3 Look-up table algorithms live fuel moisture content (Quan et al., 2016), soil
moisture (Bai et al., 2017), snow depth (Che et al.,
Optimization algorithms are computationally
2016), surface broadband emissivity (Cheng
expensive and very slow, performing the inver-
et al., 2017), and estimating surface incident solar
sion process with a huge amount of remotely
radiation (Zhang et al., 2014a).
sensed data. The look-up table (LUT) approach
has been used extensively to speed up the inver- 1.6.2.4 Direct estimation methods
sion process. It precomputes the model reflec-
The direct estimation algorithms are very
tance for a large range of combinations of
similar to the LUT algorithms, but the major dif-
parameter values. In this manner, the most
ference is that the table searching is replaced by
computationally expensive aspect can be
regression analysis (Liang, 2003; Liang et al.,
completed before the inversion is attempted,
1999). The first part is almost identical. The for-
and the problem is reduced to searching a LUT
ward RT model is used to simulate TOA reflec-
for the modeled reflectance set that most resem-
tance (radiance), but the inversion part relies
bles the measured set.
on regression analysis that links the inverted pa-
In an ordinary LUT approach, the dimensions
rameters and the simulated reflectance/radiance
of the table must be large enough to achieve high
instead of table searching. It is sometimes called
accuracy, which leads to much slower online
hybrid inversion algorithm because it combines
searching. Moreover, many parameters must be
the RT model simulation (physical) with regres-
fixed in the LUT method. To reduce the dimen-
sion analysis (statistical). Fig. 1.22 illustrates the
sions of the LUTs for rapid table searching,
direct estimation method for estimating land
empirical functions are used to fit the LUT
surface albedo.
values so that a table searching procedure be-
Liang et al. (2005) applied this method to esti-
comes a simple calculation of the local functions,
mate surface snow albedo from MODIS data us-
or a simple linear regression is executed instead
ing linear regression analysis at each angular bin
of table searching.
in the solar illumination and sensor viewing ge-
To reduce the dimensions of the LUTs for
ometry. The linear regression equations were
rapid table searching, Gastellu-Etchegorry et al.
training using the simulated data. This algorithm
(2003) developed empirical functions to fit the
for estimating land surface albedo has been
LUT values so that a table searching procedure
tested for a variety of remotely sensed data,
becomes a simple calculation of the local func-
such as MODIS (Wang et al., 2015c), Landsat
tions. Hedley et al. (2009) developed a so-called
data (He et al., 2018), AVIRIS (He et al., 2014b),
adaptive LUT method. It relies on a postprocess-
and Chinese HJ (He et al., 2015a). In fact, the
ing step to organize the data into a binary space
global land surface albedo products have been
partitioning tree that facilitates an efficient inver-
generated, such as the GLASS albedo product
sion search algorithm.
(Liang et al., 2013c; Qu et al. 2014a, 2016) from
36 1. A systematic view of remote sensing

FIGURE 1.22 Illustration of the direct estimation method for estimating land surface albedo from satellite data.

MODIS data, and the VIIRS albedo product by the sensors. In many cases, the estimation of
(Wang et al., 2013a; Zhou et al., 2016). Compare some variables not directly related to radiance
to the conventional albedo estimation algo- is desirable.
rithms, the direct estimation method does not Given the ill-posed nature of remote sensing
require surface reflectance from atmospheric inversion (the number of unknowns is far greater
correction. Moreover, it also does not need an than the number of observations, and the multi-
accumulation of observations for surface BRDF spectral data are highly correlated) and vast
modeling during a certain period of time that expansion of the amount of observation data,
cannot capture the rapid surface changes. data fusion techniques, which simply register
This approach has also been recently used to and combine datasets together from multiple
estimate surface shortwave net radiation (He sources, may be one solution. The data assimila-
et al., 2015b; Kim and Liang, 2010; Wang et al. tion (DA) method allows use of all available
2014, 2015a,b) and TOA shortwave albedo and information within a time window to estimate
radiative flux (Wang and Liang 2016, 2017). various unknowns of land surface models
(Liang, 2007; Liang et al., 2013a; Liang and Qin,
1.6.2.5 Data assimilation methods 2008). The information that can be incorporated
includes observational data, existing pertinent a
The values of land surface variables, esti-
priori information, and, importantly, a dynamic
mated using the methods from the previous sec-
model that describes the system of interest and
tions from different sources, may not be
encapsulates theoretical understanding.
physically consistent. Most techniques do not
A DA scheme commonly includes the
take advantage of observations acquired at
following components: (1) a forward dynamic
different times and cannot handle observations
model that describes the time evolution of state
with different spatial resolutions together. In
variables such as surface temperature, soil mois-
particular, these techniques estimate only vari-
ture, and carbon stocks; (2) an observation model
ables that significantly affect radiance received
 1.6 Estimating quantitative variables 37
that relates the model estimates of state variables radiative transfer model that are mostly sensitive
to satellite observations and vice versa; (3) an to the observed radiance/reflectance can be
objective function that combines model esti- directly estimated, and a group of other parame-
mates and observations along with any associ- ters (e.g., FAPAR, surface albedo) are calculated
ated prior information and error structure; (4) as an indirect inversion. A series of experiments
an optimization scheme that adjusts forward have been conducted to demonstrate that this
model parameters or state variables to minimize method is robust (Liu et al., 2014; Ma et al.
the discrepancy between model estimates and 2017a,b, 2018; Shi et al. 2016, 2017; Xiao et al.,
satellite observations; and (5) error matrices 2015). However, the disadvantage of this method
that specify the uncertainty of the observations, is computationally expensive and currently
model, and any background information (these cannot be used for generating the global prod-
are usually included in the objective function). ucts. If the TOA observation data are directly
DA for inverting land surface parameters can assimilated (Shi et al., 2016), it can be worse.
integrate not only remote sensing data with Computer emulation techniques may help
different characteristics (multispectral, multian- (Gomez-Dans et al., 2016; Rivera et al., 2015).
gular, and multitemporal) but also various mea-
surement data and a prior knowledge (Lewis 1.6.2.6 Spatial and temporal scaling
et al., 2012; Liang and Qin, 2008). The main dif- Remote sensing data with different spatial
ference with the optimization method described resolutions have different temporal resolutions,
above is that it must rely on a dynamic equation generally speaking, high spatial resolution data
that can be a mechanical physically based model, with low temporal resolution. For example,
such as a crop growth model or a statistical Landsat has multispectral optical data at the
model. The recent applications of DA may spatial resolution of 30 m, but the satellite repeat
include soil moisture (Chen et al., 2015b; Fan cycle is 16 days and cloud contamination leads to
et al., 2015; Han et al., 2015; Qin et al. 2013, much lower effective temporal resolution. In
2015; Yang et al., 2016), hydrological parameters contrast, MODIS data have a higher temporal
(Lei et al., 2014; Xie et al., 2014), heat fluxes resolution but a lower spatial resolution. By
(Wang et al., 2013b; Xu et al. 2014, 2015), carbon combining data at different spatial resolutions,
flux (Liu et al., 2015a), and crop yield (Cheng we will be able to generate new data products
et al., 2016; Huang et al. 2015a,b, 2016). with high spatial and temporal resolutions.
A new remote sensing DA approach has been Gao et al. (2006) proposed to generate a
recently proposed to estimate an improved suite Landsat-like daily reflection at the spatial resolu-
of land products simultaneously. The framework, tion of 30m from high spatial resolution low-
as illustrated in Fig. 1.23, has several unique fea- frequency Landsat data and high-frequency
tures compared to the conventional inversion low spatial resolution MODIS data using a new
methodology: making use of multiple satellite time-space adaptive reflectivity fusion model
data, particularly time series observations; being (STARFM) algorithm. Zhu et al. (2010) proposed
able to incorporate a prior knowledge and various the enhanced version of STARFM (ESTARFM).
constraints; adapting the ensemble of multiple Similar ideas have also been applied to generate
inversion algorithms for estimating the same the LST products at the high spatial and tempo-
land surface variable by taking advantage of the ral resolutions (Moosavi et al., 2015; Weng et al.,
strengths of the individual algorithm. More 2014). These methods are also known in the liter-
importantly, this approach estimates land surface ature as spatial and temporal data fusion
variables using both direct and indirect retrieval methods (Chen et al., 2015a; Wu et al., 2015;
methods. The key parameters of the surface Zhang et al., 2015).
38 1. A systematic view of remote sensing

FIGURE 1.23 A data assimilationebased inversion framework for estimating a suite of land surface variables
simultaneously.

If we need to generate the high-level products associated with a node on a finer or coarser scale.
at multiple spatial resolutions, we need different The multiscale estimation method continues to
methods. The first case is how to estimate from loop up and down. Upward updates the infor-
satellite data with multiple spatial resolutions mation from fine to coarse scales and vice versa
that are acquired simultaneously. In theory, we to transfer information from coarse to fine scales.
can execute inversion separately. However, if The second case is that the inverted satellite
we combine the observations at multiple spatial products of different scales may have systematic
resolutions for joint inversion, relevant informa- errors, spatial discontinuity, and uncertainty.
tion at different scales will be passed to improve Using fusion of multiscale data product, infor-
the inversion accuracy. For example, Jiang et al. mation on different scales is used to improve
(Jiang et al., 2016b) effectively improved the the accuracy and quality at each scale. He et al.
LAI inversion accuracy of ETMþ and MODIS (2014c) integrated land surface albedo products
data using the ensemble multiscale filter. De at three scales using multiresolution tree (MRT)
Vyver and Roulin (2009) used scale recursive method: MISR (1100m), MODIS (500m), and
estimation method to estimate precipitation. ETMþ (30m). They demonstrated that the multi-
These multiscale estimation processes can be scale data fusion method can fill the missing data
represented by a “tree” structure. All grids cover and reduce the systematic error and uncertainty.
the same area, but each grid corresponds to a This approach has also been applied to integrate
different spatial resolution. Each node is the FAPAR products (Tao et al., 2017).
 1.6 Estimating quantitative variables 39
1.6.2.7 Regularization method (2) Integrating data at different spatial
To overcome the “ill-posed” inversion prob- resolutions;
lem, various regularization techniques have (3) Integrating data acquired by sensors with
been recently proposed (Delahaies et al., 2017; different principles.
Laurent et al., 2014) to increase the amount of in- For example, Sun et al. (2011) demonstrated
formation to enhance the stability and accuracy the potential of the combined use of LiDAR
of the solution, such as making use of multi- samples and SAR imagery for forest biomass
source data and a prior knowledge. mapping. A follow-up study further explored
the use of integrated LiDAR, SAR, and multi-
spectral imaging (Landsat) data for achieving
1.6.3 Use of multisource data the accuracy requirements of a global forest
Modern remote sensing technology has biomass mapping mission (Montesano et al.,
evolved to acquire massive amounts of data at 2013). Other forest biomass mapping studies
different spatial, temporal, and spectral resolu- also demonstrated the value of integrating mul-
tions using multiple sensors at various platforms tispectral data and LiDAR data, which can effec-
for the monitoring of the Earth’s environment. tively extend 3D information from discrete
The geostationary satellite data can reveal the LiDAR measurements of forest structure across
diurnal variation of the parameters due to high scales much larger than that of the LiDAR foot-
temporal resolution with relatively low spectral print (Pflugmacher et al., 2014; Zhang et al.,
resolutions (few bands) in many systems, while 2014b) and so on.
the polar-orbiting satellite data usually have
high spectral resolutions and global coverage
but with low temporal resolution. Different 1.6.4 Use of a prior knowledge
sensors have their own advantages in different
Any a prior information that affects the pixel
applications with different band settings,
values can help improve inversion accuracy
observing angles, and temporal and spatial reso-
and stability. Cui et al. (2014) demonstrated
lutions. The integrative use of a variety of
that the combination of a priori knowledge
different satellite data is becoming a new trend
with a regularized inversion can robustly
for increasing the amount of known information
estimate surface albedo through the linear
to improve the accuracy of land parameter
kernel-driven BRDF model for the cases where
inversion.
the number of observations is insufficient, or
Many research projects have shown that the
the angular distribution is poor.
accuracy of the products produced by inte-
Generation of a priori knowledge is the first
grating multisource remote sensing data is
step. A priori knowledge may come from mea-
significantly higher than that of single sensor
surement data, expert knowledge, or a variety
data. Generally, there are three types of integra-
of existing data products. The a priori knowledge
tion according to data characteristics:
for inversion can be generated by analyzing the
(1) Integrating data acquired using the similar surface measurements, the temporal and spatial
principles at the similar spatial resolutions, distribution of the existing satellite data products,
for example, multispectral data from or the numerical model simulations. The a priori
different sensors of polar-orbiting satellites knowledge of the land surface products is mainly
and/or geostationary satellites: coarse composed of statistical values, statistic con-
resolution (MODIS/MERIS/VIIRS/MISR), straints, and intrinsic physical correlations be-
fine resolution (ETMþ/SPOT); tween the land surface variables. The statistical
40 1. A systematic view of remote sensing

values include the annual mean and variance of (Aznay et al., 2011), soil moisture (Owe et al.,
the land surface products and their temporal 2008), snow/ice covers (Husler et al., 2014),
and spatial distributions, for example, the empir- land cover (Broxton et al., 2014), and vegetation
ical relationship between surface temperature, seasonal change (Verger et al., 2016).
emissivity, broadband albedo, and LAI.
There are many mathematical equations that
can be used to characterize the vegetation
growth curve (Tsoularis and Wallace, 2002),
1.6.5 Spaceetime constraints
such as the classical Fourier function to charac- Satellites observe the specific location of the
terize the dynamic changes of NDVI (Hermance, Earth’s surfaces periodically, and the current
2007; Zhou et al., 2015) and surface temperature inversion algorithms estimate surface properties
(Xu and Shen, 2013). mainly using the data acquired at a specific time.
The logistic function has been used for LAI However, many land variables are changing
inversion (Qu et al., 2012), and the double logis- with time, such as LAI, surface temperature,
tic functions are also often used to characterize and short-range surface shortwave radiation.
annual changes in vegetation indices (Beck Inversion relying on temporally discrete satellite
et al., 2006; Fisher, 1994; Zhang et al., 2003). observations cannot effectively capture the tem-
Bayesian principles can be used for effective poral variations. Therefore, one of the solutions
incorporation of a priori knowledge into the in developing advanced inversion algorithms is
inversion process (Qu et al., 2014b; Shiklomanov to invert a temporal sequence of observations
et al., 2016; Varvia et al., 2018). Bayesian network by using the coupled surface dynamic process
method has been used to invert LAI and other model that describes the temporal variations of
parameters (Qu et al., 2014b; Quan et al., 2015; land surface variables with the radiative transfer
Zhang et al., 2012). Most of these studies using model that links the surface state variables with
the Bayesian network approach assume that radiance observed by the satellite sensors.
the inverted variables are independent, but The time constraints apply to inverting the
Quan et al. (2015) take into account the correla- surface parameters from a time series of remote
tion between variables. Laurent et al. (2014) sensing data, assuming that some parameters
estimated LAI and other parameters using are constant and some are temporally variable
Bayesian inversion algorithm by incorporating (Houborg et al., 2007 5397; Lauvernet et al.,
a priori knowledge and a coupled vegetation- 2008 5398) or assuming that the parameters
atmosphere radiative transfer model. (such as LAI) follow a certain seasonal variation
Many satellite products have missing values patterns or are temporally smooth (Quaife and
and “noises” that need to be filled and smoothed Lewis, 2010).
before the statistical analysis. There have been The spatial constraints apply to a sequence of
studies that make use of a priori knowledge data at a small neighborhood. Assuming that
based on multiple satellite data products. For some parameters are spatially uniform (e.g.,
example, albedo (Fang et al., 2007; He et al., same cover type), grouping these pixels can
2014a; Moody et al., 2005; Zhang et al., 2010), reduce the number of unknown variables (Atz-
LAI (Fang et al. 2008b, 2013), FAPAR and frac- berger, 2004; Laurent et al., 2013). For example,
tional vegetation coverage (Verger et al., 2015 the 33 window was used to constrain the
5350), land surface temperature (Bechtel, 2015), crop LAI inversion (Atzberger and Richter,
sea surface temperature (Banzon et al., 2014), 2012). In atmospheric correction, the entire
ground radiation (Krahenmann et al., 2013; image is often assumed to be a window with
Posselt et al., 2014), aerosol optical properties one set of atmospheric optical parameters, and
 1.7 Production, archiving, and distribution of high-level products 41
sometimes the entire image is divided into mul- of data information systems is essential. For
tiple windows, each with the same atmospheric example, NASA’s principal Earth Science infor-
parameters to be estimated. Spatial smoothness mation system is the Earth Observing System
is also used to constrain parameter inversion, Data and Information System (EOSDIS), which
as in the case of temporal smoothness constraints has been operational since August 1994. EOSDIS
(Wang et al., 2008). acquires, processes, archives, and distributes
Earth Science data and information products
created from satellite data that arrive at a rate
1.6.6 Algorithm ensemble of more than four trillion bytes (4 terabytes)
per day. More and more information systems
At present, the majority of satellite products
are supported by high-performance computing
have been generated by one “best” algorithm
capabilities.
for each product. As different algorithms often
The various levels of data used by the EOSDIS
have different characteristics, it is difficult to
are defined below. For some instruments, there
determine the “best” algorithm. As a result, there
will be no Level 1B product that is distinct
often exist systemic biases, and the accuracies of
from the Level 1A product. In these cases, the
these products vary regionally under different
reference to Level 1B data can be assumed to
conditions. Algorithm ensemble integrates the
refer to Level 1A data. Brief definitions follow:
outputs of a group of algorithms and can greatly
improve the inversion accuracy and stability. • Level 0: Reconstructed, unprocessed
Such an approach has been first applied to the instrument/payload data at full resolution;
development of the GLASS products (Liang any and all communications artifacts (e.g.,
et al., 2013c). GLASS land surface albedo prod- synchronization frames, communications
uct is generated in the integration of two estima- headers, duplicate data removed). In most
tion algorithms (Liu et al. 2013a,b). The GLASS cases these data are provided by EDOS to a
ET product is based on integrating five algo- Distributed Active Archive Center (DAAC) as
rithms (Yao et al., 2014). production datasets to produce higher level
The Bayesian Model Average (BMA) method products.
has been proposed to integrate multiple down- • Level 1A: Reconstructed, unprocessed
ward and longwave radiation models (Wu instrument data at full resolution, time-
et al., 2012).There are many other examples referenced, and annotated with ancillary
(Chen et al., 2015d; Kim et al., 2015; Shi and information, including radiometric and
Liang, 2013a,b; Shi and Liang, 2014). geometric calibration coefficients, and
georeferencing parameters (e.g., platform
ephemeris, computed and appended but not
applied to the Level 0 data).
1.7 Production, archiving, and • Level 1B: Level 1A data that have been
distribution of high-level products processed to sensor units (not all instruments
will have a Level 1B equivalent).
Satellite observations can be converted into
• Level 2: Derived geophysical variables at the
high-level bio/geophysical products using the
same resolution and location as the Level 1
inversion methods outlined above through a
source data.
production system. Given satellite observations
• Level 3: Variables mapped on uniform
and the inversion algorithm, production of
spaceetime grid scales, usually with some
high-level products is not straightforward
completeness and consistency.
because of the huge amount of data. Creation
42 1. A systematic view of remote sensing

FIGURE 1.24 Distribution of the NASA DAAC, downloaded from NASA website NASA’s Global Imagery Browse Ser-
vices (GIBS), provides quick access to over 800 satellite imagery products, covering every part of the world. Most imagery
is available within a few hours after satellite overpass, some products span almost 30 years, and the imagery can be rendered
in the web client or GIS application.

• Level 4: Model output or results from analyses from several NASA instruments, including
of lower level data (e.g., variables derived AIRS, AMSR2, LIS (ISS), MISR, MLS, MODIS,
from multiple measurements). MOPITT, OMI, and VIIRS, are available in less
than 3 h from satellite observations.
Archiving and distribution of the huge
ESA distributes Earth observation data from
amount of data and products are also chal-
ESA EO Missions, third party missions, ESA
lenging. As a user, we have several ways to
campaigns, the Copernicus space component,
search for EO data of interest.
as well as sample and auxiliary data from a num-
NASA Earth Science information is archived
ber of missions and instruments (https://earth.
at the DAACs located across the United States
esa.int/web/guest/data-access), under different
(Fig. 1.24). The DAACs specialize by topic area
data policies and by various access mechanisms.
and make their data available to researchers
China’s satellite data are archived and distrib-
around the world. Almost all data in EOSDIS
uted through several agencies, such as the Chi-
are held online and accessed via ftp and https.
nese Meteorological Administration (http://
NASA’s Land, Atmosphere Near real-time
www.cma.gov.cn/en2014/satellites/), mainly
Capability for EOS (LANCE) provides near
for meteorological satellite data, and China
real-time imagery, and high-level products
 1.7 Production, archiving, and distribution of high-level products 43
Center for Resources Satellite Data and Applica- an international effort developed to assist re-
tion (http://www.cresda.com/EN/) mainly for searchers in locating information on available
other land satellite remote sensing data. The datasets.
China National Data Sharing Infrastructure of In addition to the space agencies that usually
Earth System Science (www.geodata.cn) distrib- produce the high-level products from a specific
utes data products generated from major satellite data with a limited life span, some uni-
research projects in China, including surface versities and research institutes led research
measurements, high-level products from teams are also generating the high-level satellite
remotely sensed observations, and model simu- products from multiple satellite data. For
lations. It was created in 2004 and is being example, the GLASS product suite initially
contributed to by more than 40 Chinese insti- included five products (Liang et al. 2013b,c),
tutes including various local data centers. but it has been recently expanded into 12 prod-
CEOS International Directory Network (IDN) ucts (see Table 1.8) from AVHRR, MODIS, and
serves as a gateway to the world of Earth science other satellite data. The GLASS products can be
data (https://idn.ceos.org/). The CEOS IDN is

TABLE 1.8 The list of the GLASS products and their characteristics as of 2019.

Temporal
NoProduct Temporal range Spatial range resolution Spatial resolution

1 Leaf area index 1981e2018 Global land 8 days 5 km before 2000, 1 km

surface after 2000

2 Surface broadband albedo 1981e2018 Global 8 days 5 km before 2000, 1 km

after 2000

3 Broadband emissivity 1981e2018 Global land 8 days 5 km before 2000, 1 km

surface after 2000
4 Photosynthetically active radiation 2000e2018 Global land Daily 5 km
surface
5 Downward shortwave radiation 2000e2018 Global land Daily 5 km
surface
6 Surface longwave net radiation 1983, 1993, 2003, Global land 8 days 5 km before 2000, 1 km
2013 surface after 2000
7 Surface all-wave net radiation 2000e2018 Global land 8 days 5 km
surface
S Land surface temperature 1983, 1993, 2003, Global land Instantaneous 5 km before 2000, 1 km
2013 surface after 2000
9 Fraction of absorbed photosynthetically 1981e2018 Global land 8 days 5 km before 2000, 1 km
active radiation surface after 2000
10 Fractional vegetation cover 1981e2018 Global land 8 days 5 km before 2000, 1 km
surface after 2000
11 Evapotranspiration 1981e2018 Global land 8 days 5 km before 2000, 1 km
surface after 2000

12 Gross primary production 1981e2018 Global land 8 days 5 km before 2000, 1 km

surface after 2000
44 1. A systematic view of remote sensing

freely downloaded from www.geodata.cn or instruments, measured data, and results offer vital
www.glass.umd.edu. ways that will lead to success and progress. One of
More details on data, production, and distri- the critical issues is the mismatch between ground
bution systems are presented in Chapter 22. “point” measurements and kilometer-scale pixel
values over the heterogeneous landscape. Upscal-
ing “point” measurements using high-resolution
1.8 Product validation remotely sensed data is the key to addressing
this issue. Loew et al. (2017) provided a compre-
One of the key components of information hensive review of state-of-the-art methods of sat-
extraction is validation. Without a known accu- ellite validation and documents their similarities
racy, the high-level product cannot be used reli- and differences, and more details on validating
ably and, therefore, has limited applicability. different high-level land products are discussed
With various land products available, users need in the relevant chapters.
quantitative information on product uncertainties
to select the most suitable product, or combination
of products, for their specific needs. As remote
sensing observations are generally merged with
1.9 Remote sensing applications
other sources of information or assimilated within
Remote sensing has generated comprehen-
process models, evaluation of product accuracy is
sive, near-real-time environmental data, infor-
required. Making quantified accuracy information
mation, and analyses. This important
available to the user can ultimately provide devel-
technology serves a wide range of users and em-
opers the necessary feedback for improving the
powers decision-makers to respond more effec-
products and can possibly provide methods for
tively to the many environmental challenges
their fusion to construct a consistent long-term se-
facing us today. Remotely sensed information
ries of surface status.
can be applied in different ways (Balsamo
Land product validation has to rely on ground
et al., 2018; Pasetto et al., 2018). We can group
measurements, which may be both time-
this information into the following categories,
consuming and very expensive. Because of its
as illustrated in Fig. 1.25:
importance, such product validation must involve
the efforts of the entire remote sensing commu- (1) Detection and monitoring. The long-term
nity. Sharing the validation methodologies, individual or combined high-level products

FIGURE 1.25 Different ways of applications of the high-level satellite products.

 1.10 Conclusion 45
can be used for characterizing the surface other in situ data to societal benefits. Data can
dynamics resulted from natural variabilities be used for driving, calibrating, and validating
and human activities, for example, LAI for models and decision support tools. The last
the global “greening” trends (Zhu et al., four chapters of this book illustrate how different
2016); snow cover and surface albedo for remote sensing data and products can be used
regional and global climate change (Chen for monitoring land cover and land use changes
et al., 2015c; He et al. 2013, 2014a). and assessing their environmental impacts.
(2) Driving numerical models and decision GEO identified nine societal benefit areas in
support systems. The satellite products have which there was recognition that clear societal
been used as inputs to various models, such benefits could be derived from a coordinated
as ecological models (Pasetto et al., 2018), global observation system. As illustrated in
hydrological models (McCabe et al., 2017), Fig. 1.27, the nine societal benefit areas include,
numerical weather prediction models (Fang for example,
et al., 2018), Earth system models (Balsamo
• Understanding environmental factors
et al., 2018; Simmons et al., 2016), or decision
affecting human health and well-being.
support systems (Mohammed et al., 2018;
• Improving management of energy resources.
Murray et al., 2018; Rahman and Di, 2017).
• Understanding, assessing, predicting,
(3) Validating model simulations. Numerical
mitigating, and adapting to climate
models may not actually use satellite
variability and change.
products, but their simulation results are
• Improving water resource management
validated by satellite products, such as soil
through better understanding of the water
moisture (Gu et al., 2019), precipitation
cycle.
(Tapiador et al., 2018), temperature (Ouyang
• Improving weather information, forecasting,
et al., 2018), and water storage (Zhang et al.,
and warning.
2017b).
(4) Assimilation into numerical models. The Many of these societal benefit areas are them-
numerical models usually have many selves complex clusters of issues, with multiple
parameters that may be spatially variable and varied stakeholders. In each area there are
that are difficult to predefine. DA approach observational needs for numerous variables,
can be used to adjust these parameters with requirements for their accuracy, spatial
iteratively by matching the model simulated and temporal resolution, and speed of delivery
values with satellite products (Liang et al., to the user. These societal benefit areas are now
2013a; Liang and Qin, 2008). Thus, the at widely varying levels of maturity with respect
models can generate the simulations with to establishing user needs, defining the observa-
spatial-temporal continuity (satellite tion requirements, and implementing coordi-
products are often irregular), such as soil nated systems.
moisture (Qin et al., 2009), or simulate the
variables that cannot be directly observed or
retrieved from remotely sensed data, such as 1.10 Conclusion
carbon fluxes (Scholze et al., 2017) and crop
yield (Fang et al., 2008a; Huang et al., 2019; Significant progress in engineering issues has
Jin et al., 2018). enabled remarkable improvements in platform
and sensor systems that are evidenced by
Fig. 1.26 depicts the linkage and flow of infor-
many indicators, such as signal-to-noise ratios,
mation from remote sensing observations and
resolutions, pointing accuracies, geometric and
46 1. A systematic view of remote sensing

Predictions
Earth System Models
Weather
Climate
Atmosphere Policy
Other... Decisions
Decision Support Societal
Benefits
Management
Decisions
DATA
Earth Observation Personal
Systems Decisions
Observations
Remotely-sensed
In situ

On-going feedback to optimize

value and reduce gaps

FIGURE 1.26 Linking earth observations with societal benefits (CENR/IWGEO, 2005).

FIGURE 1.27 GEOSS and applications (GEO, 2005).

spectroradiometric integrity, and calibration. surface radiation regimes, and some of these
The immense amounts of data available from models have been incorporated into useful algo-
satellite observations present great challenges. rithms for estimating land surface variables from
Considerable investments have been made in satellite observations. However, development of
developing physical models to understand realistic and computationally simplified surface
 References 47
radiation models mostly suitable for inversion of Axelsson, C., Skidmore, A.K., Schlerf, M., Fauzi, A.,
land surface variables from satellite data are still Verhoef, W., 2013. Hyperspectral analysis of mangrove
foliar chemistry using PLSR and support vector
urgently required. Inversion of land surface pa- regression. Int. J. Remote Sens. 34, 1724e1743.
rameters is generally a nonlinear ill-posed prob- Aznay, O., Zagolski, F., Santer, R., 2011. A new climatology
lem, and the use of more spatial and temporal for remote sensing over land based on the inherent optical
constraints by incorporating a priori knowledge properties. Int. J. Remote Sens. 32, 2851e2885.
and integrating multiple-source data deserves Bai, X.J., He, B.B., Li, X., Zeng, J.Y., Wang, X., Wang, Z.L.,
Zeng, Y.J., Su, Z.B., 2017. First assessment of Sentinel-
further research. 1A data for surface soil moisture estimations using a
Although remote sensing data products have coupled water cloud model and advanced integral equa-
been widely used, significant disconnects be- tion model over the Tibetan plateau. Remote Sens. 9, 20.
tween remote sensing development and applica- Balsamo, G., Agusti-Panareda, A., Albergel, C., Arduini, G.,
tions continue to exist. Some products developed Beljaars, A., Bidlot, J., Bousserez, N., Boussetta, S.,
Brown, A., Buizza, R., Buontempo, C., Chevallier, F.,
by remote sensing scientists have not been Choulga, M., Cloke, H., Cronin, M., Dahoui, M., De
widely used, and many variables required by Rosnay, P., Dirmeyer, P., Drusch, M., Dutra, E., Ek, M.,
land process models and decision support sys- Gentine, P., Hewitt, H., Keeley, S., Kerr, Y., Kumar, S.,
tems have not been generated. The product accu- Lupu, C., Mahfouf, J.-F., McNorton, J., Mecklenburg, S.,
racy and application requirements may not Mogensen, K., Mu~ noz-Sabater, J., Orth, R., Rabier, F.,
Reichle, R., Ruston, B., Pappenberger, F., Sandu, I.,
always be consistent. Successful applications Seneviratne, S., Tietsche, S., Trigo, I., Uijlenhoet, R.,
are not static but evolve as new sensor, data pro- Wedi, N., Woolway, R., Zeng, X., 2018. Satellite and
cessing, and network technologies emerge. The in situ observations for advancing global earth surface
improved linking of remote sensing science modelling: a review. Remote Sens. 10, 2038.
with applications would be most advantageous. Banskota, A., Serbin, S.P., Wynne, R.H., Thomas, V.A.,
Falkowski, M.J., Kayastha, N., Gastellu-Etchegorry, J.-P.,
Townsend, P.A., 2015. An LUT-based inversion of
DART model to estimate forest LAI from hyperspectral
References data. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens. 8,
3147e3160.
Ahmad, S., Kalra, A., Stephen, H., 2010. Estimating soil mois- Banzon, V.F., Reynolds, R.W., Stokes, D., Xue, Y., 2014. A 1/4
ture using remote sensing data: a machine learning degrees-Spatial-Resolution daily sea surface temperature
approach. Adv. Water Resour. 33, 69e80. climatology based on a blended satellite and in situ
Ahmed, O.S., Franklin, S.E., Wulder, M.A., White, J.C., 2015. analysis. J. Clim. 27, 8221e8228.
Characterizing stand-level forest canopy cover and height Bechtel, B., 2015. A new global climatology of annual land
using Landsat time series, samples of airborne LiDAR, surface temperature. Remote Sens. 7, 2850e2870.
and the Random Forest algorithm. ISPRS J. Photogram- Beck, P.S.A., Atzberger, C., Hogda, K.A., Johansen, B.,
metry Remote Sens. 101, 89e101. Skidmore, A.K., 2006. Improved monitoring of vegetation
Akbar, R., Cosh, M.H., O’Neill, P.E., Entekhabi, D., dynamics at very high latitudes: a new method using
Moghaddam, M., 2017. Combined radar-radiometer sur- MODIS NDVI. Remote Sens. Environ. 100, 321e334.
face soil moisture and roughness estimation. IEEE Trans. Becksch€afer, P., Fehrmann, L., Harrison, R.D., Xu, J.,
Geosci. Remote Sens. 55, 4098e4110. Kleinn, C., 2014. Mapping Leaf Area Index in subtropical
Akbar, R., Moghaddam, M., 2015. A combined active-passive upland ecosystems using RapidEye imagery and the ran-
soil moisture estimation algorithm with adaptive regula- domForest algorithm. iFor. Biogeosci. For. 7, 1.
rization in support of SMAP. IEEE Trans. Geosci. Remote Belgiu, M., Dragut, L., 2016. Random forest in remote
Sens. 53, 3312e3324. sensing: a review of applications and future directions.
Atzberger, C., 2004. Object-based retrieval of biophysical can- ISPRS J. Photogrammetry Remote Sens. 114, 24e31.
opy variables using artificial neural nets and radiative Ben Hmida, S., Kallel, A., Gastellu-Etchegorry, J.P.,
transfer models. Remote Sens. Environ. 93, 53e67. Roujean, J.L., 2017. Crop biophysical properties estima-
Atzberger, C., Richter, K., 2012. Spatially constrained inver- tion based on LiDAR full-waveform inversion using the
sion of radiative transfer models for improved LAI map- DART RTM. IEEE J. Sel. Top. Appl. Earth Obs. Remote
ping from future Sentinel-2 imagery. Remote Sens. Sens. 10, 4853e4868.
Environ. 120, 208e218.
48 1. A systematic view of remote sensing

Blackard, J.A., Finco, M.V., Helmer, E.H., Holden, G.R., 2015d. Using Bayesian model averaging to estimate
Hoppus, M.L., Jacobs, D.M., Lister, A.J., Moisen, G.G., terrestrial evapotranspiration in China. J. Hydrol. 528,
Nelson, M.D., Riemann, R., Ruefenacht, B., Salajanu, D., 537e549.
Weyermann, D.L., Winterberger, K.C., Brandeis, T.J., Chen, Z.Q., Shi, R.H., Zhang, S.P., 2013. An artificial neural
Czaplewski, R.L., McRoberts, R.E., Patterson, P.L., network approach to estimate evapotranspiration from
Tymcio, R.P., 2008. Mapping US forest biomass using remote sensing and AmeriFlux data. Front. Earth Sci. 7,
nationwide forest inventory data and moderate resolu- 103e111.
tion information. Remote Sens. Environ. 112, 1658e1677. Cheng, J., Liu, H., Liang, S., Nie, A., Liu, Q., Guo, Y., 2017.
Borel, C.C., Gerstl, S.A.W., Powers, B.J., 1991. The radiosity A framework for estimating the 30-m thermal-infrared
method in optical remote sensing of structured 3-D broadband emissivity from Landsat surface-reflectance
surfaces. Remote Sens. Environ. 36, 13e44. data. J. Geophys. Res. Atmospheres 122, 11405e11421.
Breaker, L.C., 1990. Estimating and removing sensor-induced Cheng, Z.Q., Meng, J.H., Wang, Y.M., 2016. Improving
correlation from advanced very high resolution radiom- spring maize yield estimation at field scale by assimi-
eter satellite data. J. Geophys. Res. 95, 9701e9711. lating time-series HJ-1 CCD data into the WOFOST model
Broxton, P.D., Zeng, X.B., Sulla-Menashe, D., Troch, P.A., using a new method with fast algorithms. Remote Sens. 8.
2014. A global land cover climatology using MODIS Conti, S., Gosling, J.P., Oakley, J.E., O’Hagan, A., 2009.
data. J. Appl. Meteorol. Climatol. 53, 1593e1605. Gaussian process emulation of dynamic computer
Castelletti, A., Galelli, S., Ratto, M., Soncini-Sessa, R., codes. Biometrika 96, 663e676.
Young, P.C., 2012. A general framework for Dynamic Cui, S.C., Yang, S.Z., Zhu, C.J., Wen, N., 2014. Remote
Emulation Modelling in environmental problems. Envi- sensing of surface reflective properties: role of regulariza-
ron. Model. Softw 34, 5e18. tion and a priori knowledge. Optik 125, 7106e7112.
Castillo, P.A., Merelo, J., Prieto, A., Rivas, V., Romero, G., Da, Y., Xiurun, G., 2005. An improved PSO-based ANN with
2000. G-Prop: global optimization of multilayer percep- simulated annealing technique. Neurocomputing 63,
trons using GAs. Neurocomputing 35, 149e163. 527e533.
CENR/IWGEO, 2005. Strategic plan for the U.S. integrated Darvishzadeh, R., Matkan, A.A., Ahangar, A.D., 2012. Inver-
Earth observation system. In: N.S.a.T.C.C.o.E.a.n. sion of a radiative transfer model for estimation of rice
Resources. National Science and Technology Council canopy chlorophyll content using a lookup-table
Committee on Environment and natural Resources, Approach. IEEE J. Sel. Top. Appl. Earth Obs. Remote
Washington DC, p. 149. Sens. 5, 1222e1230.
Chan, V., Chan, C., 2017. Towards developing the piece-wise Darvishzadeh, R., Skidmore, A., Schlerf, M., Atzberger, C.,
linear neural network algorithm for rule extraction. Int. J. 2008. Inversion of a radiative transfer model for esti-
Cogn. Inf. Nat. Intell. 11, 17. mating vegetation LAI and chlorophyll in a heteroge-
Che, T., Dai, L.Y., Zheng, X.M., Li, X.F., Zhao, K., 2016. Esti- neous grassland. Remote Sens. Environ. 112, 2592e2604.
mation of snow depth from passive microwave bright- Dash, J., Ogutu, B.O., 2016. Recent advances in space-borne
ness temperature data in forest regions of northeast optical remote sensing systems for monitoring global
China. Remote Sens. Environ. 183, 334e349. terrestrial ecosystems. Prog. Phys. Geogr. 40, 322e351.
Chen, B., Huang, B., Xu, B., 2015a. Comparison of spatiotem- Delahaies, S., Roulstone, I., Nichols, N., 2017. Constraining
poral fusion models: a review. Remote Sens. 7, DALECv2 using multiple data streams and ecological
1798e1835. constraints: analysis and application. Geosci. Model
Chen, J., Quan, W.T., Cui, T.W., Song, Q.J., Lin, C.S., 2014. Dev. (GMD) 10, 2635e2650.
Remote sensing of absorption and scattering coefficient de Vyver, H.V., Roulin, E., 2009. Scale-recursive estimation
using neural network model: development, validation, for merging precipitation data from radar and microwave
and application. Remote Sens. Environ. 149, 213e226. cross-track scanners. J. Geophys. Res. Atmospheres 114,
Chen, W.J., Huang, C.L., Shen, H.F., Li, X., 2015b. Compari- 14.
son of ensemble-based state and parameter estimation Ding, Y.L., Zhang, H.Y., Li, Z.W., Xin, X.P., Zheng, X.M.,
methods for soil moisture data assimilation. Adv. Water Zhao, K., 2016. Comparison of fractional vegetation cover
Resour. 86, 425e438. estimations using dimidiate pixel models and look- up ta-
Chen, X., Liang, S., Cao, Y., He, T., Wang, D., 2015c. ble inversions of the PROSAIL model from Landsat 8 OLI
Observed contrast changes in snow cover phenology in data. J. Appl. Remote Sens. 10, 15.
northern middle and high latitudes from 2001e2014. Disney, M., Lewis, P., Saich, P., 2006. 3D modelling of forest
Sci. Rep. 5, 16820. canopy structure for remote sensing simulations in the
Chen, Y., Yuan, W.P., Xia, J.Z., Fisher, J.B., Dong, W.J., optical and microwave domains. Remote Sens. Environ.
Zhang, X.T., Liang, S., Ye, A.Z., Cai, W.W., Feng, J.M., 100, 114e132.
 References 49
Donmez, C., Berberoglu, S., Erdogan, M.A., Tanriover, A.A., Garcia-Gutierrez, J., Martinez-Alvarez, F., Troncoso, A.,
Cilek, A., 2015. Response of the regression tree model to Riquelme, J.C., 2015. A comparison of machine learning
high resolution remote sensing data for predicting percent regression techniques for LiDAR-derived estimation of
tree cover in a Mediterranean ecosystem. Environ. Monit. forest variables. Neurocomputing 167, 24e31.
Assess. 187, 12. Gastellu-Etchegorry, J.P., 2008. 3D modeling of satellite spec-
Du, L., Shi, S., Yang, J., Sun, J., Gong, W., 2016. Using tral images, radiation budget and energy budget of urban
different regression methods to estimate leaf nitrogen landscapes. Meteorol. Atmos. Phys. 102, 187e207.
content in rice by fusing hyperspectral LiDAR data and Gastellu-Etchegorry, J.P., Gascon, F., Esteve, P., 2003. An
laser-induced chlorophyll fluorescence data. Remote interpolation procedure for generalizing a look-up table
Sens. 8, 14. inversion method. Remote Sens. Environ. 87, 55e71.
Fan, L., Xiao, Q., Wen, J.G., Liu, Q., Jin, R., You, D.Q., Gastellu-Etchegorry, J.P., Yin, T.G., Lauret, N., Cajgfinger, T.,
Li, X.W., 2015. Mapping high-resolution soil moisture Gregoire, T., Grau, E., Feret, J.B., Lopes, M., Guilleux, J.,
over heterogeneous cropland using multi-resource Dedieu, G., Malenovsky, Z., Cook, B.D., Morton, D.,
remote sensing and ground observations. Remote Sens. Rubio, J., Durrieu, S., Cazanave, G., Martin, E.,
7, 13273e13297. Ristorcelli, T., 2015. Discrete anisotropic radiative transfer
Fang, H., Jiang, C., Li, W., Wei, S., Baret, F., Chen, J.M., (DART 5) for modeling airborne and satellite spectroradi-
Garcia-Haro, J., Liang, S., Liu, R., Myneni, R.B., ometer and LIDAR acquisitions of natural and urban
Pinty, B., Xiao, Z., Zhu, Z., 2013. Characterization and landscapes. Remote Sens. 7, 1667e1701.
intercomparison of global moderate resolution leaf area GEO, 2005. GEOSS 10-Year Implementation Plan Reference
index (LAI) products: analysis of climatologies and theo- Document, p. 209.
retical uncertainties. J. Geophys. Res. Biogeosci. 118 Gholizadeh, H., Robeson, S.M., Rahman, A.F., 2015.
https://doi.org/10.1002/jgrg.20051. Comparing the performance of multispectral vegetation
Fang, H., Kim, H., Liang, S., Schaaf, C., Strahler, A., indices and machine-learning algorithms for remote esti-
Townshend, G.R.G., Dickinson, R., 2007. Developing a mation of chlorophyll content: a case study in the Sundar-
spatially continuous 1 km surface albedo data set over bans mangrove forest. Int. J. Remote Sens. 36, 3114e3133.
North America from Terra MODIS products. Goel, N.S., Grier, T., 1987. Estimation of canopy parameters
J. Geophys. Res. 112, D20206, 20210.21029/ of row planted vegetation canopies using reflectance
22006JD008377. data for only four directions. Remote Sens. Environ. 21,
Fang, H., Liang, S., Hoogenboom, G., Teasdale, J., 37e51.
Cavigelli, M., 2008a. Crop yield estimation through Gomez-Dans, J.L., Lewis, P.E., Disney, M., 2016. Efficient
assimilation of remotely sensed data into DSSAT- emulation of radiative transfer codes using Gaussian pro-
CERES. Int. J. Remote Sens. 29, 3011e3032. cesses and application to land surface parameter
Fang, H., Liang, S., Townshend, J., Dickinson, R., 2008b. inferences. Remote Sens. 8, 32.
Spatially and temporally continuous LAI data sets based Gomez, C., Wulder, M.A., Montes, F., Delgado, J.A., 2012.
on an new filtering method: examples from North Modeling forest structural parameters in the mediterra-
America. Remote Sens. Environ. 112, 75e93. nean pines of Central Spain using QuickBird-2 imagery
Fang, L., Zhan, X.W., Hain, C.R., Liu, J.C., 2018. Impact of us- and classification and regression tree analysis (CART).
ing near real-time green vegetation fraction in Noah land Remote Sens. 4, 135e159.
surface model of NOAA NCEP on numerical weather Gu, X.H., Li, J.F., Chen, Y.D., Kong, D.D., Liu, J.Y., 2019. Con-
predictions. Adv. Meteorol. 12. sistency and discrepancy of global surface soil moisture
Filippi, A.M., Guneralp, I., Randall, J., 2014. Hyperspectral changes from multiple model-based data sets against sat-
remote sensing of aboveground biomass on a river ellite observations. J. Geophys. Res. Atmospheres 124,
meander bend using multivariate adaptive regression 1474e1495.
splines and stochastic gradient boosting. Remote Sens. Gui, S., Liang, S., Wang, K., Li, L., Zhang, X., 2010. Assess-
Lett. 5, 432e441. ment of three satellite-estimated land surface down-
Fisher, A., 1994. A model for the seasonal variations of vege- welling shortwave irradiance data sets. IEEE Geosci.
tation indices in coarse resolution data and its inversion to Remote Sens. Lett. 7, 776e780.
extract crop parameters. Remote Sens. Environ. 48, Halperin, J., LeMay, V., Coops, N., Verchot, L., Marshall, P.,
220e230. Lochhead, K., 2016. Canopy cover estimation in miombo
Gao, F., Masek, J., Schwaller, M., Hall, F., 2006. On the woodlands of Zambia: comparison of Landsat 8 OLI
blending of the Landsat and MODIS surface reflectance: versus RapidEye imagery using parametric, nonpara-
predicting daily Landsat surface reflectance. IEEE Trans. metric, and semiparametric methods. Remote Sens. Envi-
Geosci. Remote Sens. 44, 2207e2218. ron. 179, 170e182.
50 1. A systematic view of remote sensing

Han, X.J., Li, X., Rigon, R., Jin, R., Endrizzi, S., 2015. Soil mois- reflectance over complex vegetated scenes. Remote Sens.
ture estimation by assimilating L-band microwave bright- Environ. 132, 221e237.
ness temperature with geostatistics and observation Huang, J., Ma, H., Sedano, F., Lewis, P., Liang, S., Wu, Q.,
localization. PLoS One 10, 20. Su, W., Zhang, X., Zhu, D., 2019. Evaluation of regional
He, T., Liang, S., Song, D.-X., 2014a. Analysis of global land estimates of winter wheat yield by assimilating three
surface albedo climatology and spatial-temporal variation remotely sensed reflectance datasets into the coupled
during 1981e2010 from multiple satellite products. WOFOSTePROSAIL model. Eur. J. Agron. 102, 1e13.
J. Geophys. Res. Atmospheres 119, 10,281-210,298. Huang, J.X., Ma, H.Y., Su, W., Zhang, X.D., Huang, Y.B.,
He, T., Liang, S., Wang, D., Cao, Y., Gao, F., Yu, Y., Feng, M., Fan, J.L., Wu, W.B., 2015a. Jointly assimilating MODIS
2018. Evaluating land surface albedo estimation from LAI and ET products into the SWAP model for winter
Landsat MSS, TM, ETM þ, and OLI data based on the uni- wheat yield estimation. IEEE J. Sel. Top. Appl. Earth
fied direct estimation approach. Remote Sens. Environ. Obs. Remote Sens. 8, 4060e4071.
204, 181e196. Huang, J.X., Sedano, F., Huang, Y.B., Ma, H.Y., Li, X.L.,
He, T., Liang, S., Wang, D., Chen, X., Song, D.-X., Jiang, B., Liang, S., Tian, L.Y., Zhang, X.D., Fan, J.L., Wu, W.B.,
2015a. Land surface albedo estimation from Chinese HJ 2016. Assimilating a synthetic Kalman filter leaf area in-
satellite data based on the direct estimation approach. dex series into the WOFOST model to improve regional
Remote Sens. 7, 5495e5510. winter wheat yield estimation. Agric. For. Meteorol. 216,
He, T., Liang, S., Wang, D., Shi, Q., Goulden, M.L., 2015b. 188e202.
Estimation of high-resolution land surface net shortwave Huang, J.X., Tian, L.Y., Liang, S., Ma, H.Y., Becker-Reshef, I.,
radiation from AVIRIS data: algorithm development and Huang, Y.B., Su, W., Zhang, X.D., Zhu, D.H., Wu, W.B.,
preliminary results. Remote Sens. Environ. 167, 20e30. 2015b. Improving winter wheat yield estimation by
He, T., Liang, S., wang, D., Shi, Q., Tao, X., 2014b. Estimation assimilation of the leaf area index from Landsat TM and
of high-resolution land surface shortwave albedo from MODIS data into the WOFOST model. Agric. For. Mete-
AVIRIS data. IEEE J. Special Top. Appl. Earth Obs. orol. 204, 106e121.
Remote Sens. 7, 4919e4928. Husler, F., Jonas, T., Riffler, M., Musial, J.P., Wunderle, S.,
He, T., Liang, S., Wang, D., Wu, H., Yu, Y., Wang, J., 2012. 2014. A satellite-based snow cover climatology (1985-
Estimation of surface albedo and reflectance from moder- 2011) for the European Alps derived from AVHRR data.
ate resolution imaging spectroradiometer observations. Cryosphere 8, 73e90.
Remote Sens. Environ. 119, 286e300. Hutengs, C., Vohland, M., 2016. Downscaling land surface
He, T., Liang, S., Wang, D.D., Shuai, Y.M., Yu, Y.Y., 2014c. temperatures at regional scales with random forest
Fusion of satellite land surface albedo products across regression. Remote Sens. Environ. 178, 127e141.
scales using a multiresolution tree method in the North Jamet, C., Loisel, H., Dessailly, D., 2012. Retrieval of the spec-
Central United States. IEEE Trans. Geosci. Remote Sens. tral diffuse attenuation coefficient K-d(lambda) in open
52, 3428e3439. and coastal ocean waters using a neural network
He, T., Liang, S., Yu, Y., Liu, Q., Gao, F., 2013. Greenland sur- inversion. J. Geophys. Res.-Oceans 117, 14.
face albedo changes 1981-2012 from satellite observations. Jia, K., Liang, S., Liu, S.H., Li, Y.W., Xiao, Z.Q., Yao, Y.J.,
Environ. Res. Lett. 8, 044043. https://doi.org/044010. Jiang, B., Zhao, X., Wang, X.X., Xu, S., Cui, J., 2015. Global
041088/041748-049326/044048/044044/044043. land surface fractional vegetation cover estimation using
Hedley, J., Roelfsema, C., Phinn, S.R., 2009. Efficient radiative general regression neural networks from MODIS surface
transfer model inversion for remote sensing applications. reflectance. IEEE Trans. Geosci. Remote Sens. 53,
Remote Sens. Environ. 113, 2527e2532. 4787e4796.
Hermance, J.F., 2007. Stabilizing high-order, non-classical Jiang, B., Liang, S., Ma, H., Zhang, X., Xiao, Z., Zhao, X.,
harmonic analysis of NDVI data for average annual Jia, K., Yao, Y., Jia, A., 2016a. GLASS daytime all-wave
models by damping model roughness. Int. J. Remote net radiation product: algorithm development and pre-
Sens. 28, 2801e2819. liminary validation. Remote Sens. 8, 222.
Houborg, R., Soegaard, H., Boegh, E., 2007. Combining vege- Jiang, B., Zhang, Y., Liang, S., Zhang, X., Xiao, Z., 2014. Sur-
tation index and model inversion methods for the extrac- face daytime net radiation estimation using artificial neu-
tion of key vegetation biophysical parameters using Terra ral networks. Remote Sens. 6, 11031e11050.
and Aqua MODIS reflectance data. Remote Sens. Environ. Jiang, J., Xiao, Z., Wang, J., Song, J., 2016b. Multiscale estima-
106, 39e58. tion of leaf area index from satellite observations based on
Huang, H.G., Qin, W.H., Liu, Q.H., 2013. RAPID: a Radiosity an ensemble multiscale filter. Remote Sens. 8, 229.
Applicable to Porous IndiviDual Objects for directional
 References 51
Jin, X.L., Kumar, L., Li, Z.H., Feng, H.K., Xu, X.G., Yang, G.J., Lauvernet, C., Baret, F., Hascoët, L., Buis, S., Le Dimet, F.-X.,
Wang, J.H., 2018. A review of data assimilation of remote 2008. Multitemporal-patch ensemble inversion of coupled
sensing and crop models. Eur. J. Agron. 92, 141e152. surfaceeatmosphere radiative transfer models for land
Karlson, M., Ostwald, M., Reese, H., Sanou, J., Tankoano, B., surface characterization. Remote Sens. Environ. 112,
Mattsson, E., 2015. Mapping tree canopy cover and above- 851e861.
ground biomass in Sudano-Sahelian woodlands using Lei, F.N., Huang, C.L., Shen, H.F., Li, X., 2014. Improving the
Landsat 8 and random forest. Remote Sens. 7, 10017e10041. estimation of hydrological states in the SWAT model via
Ke, Y.H., Im, J., Park, S., Gong, H.L., 2016. Downscaling of the ensemble Kalman smoother: synthetic experiments
MODIS one kilometer evapotranspiration using for the Heihe River Basin in northwest China. Adv. Water
Landsat-8 data and machine learning approaches. Resour. 67, 32e45.
Remote Sens. 8, 26. Lewis, P., 1999. Three-dimensional plant modelling for
Kerekes, J.P., Baum, J.E., 2005. Full-spectrum spectral imag- remote sensing simulation studies using the Botanical
ing system analytical model. IEEE Trans. Geosci. Remote Plant Modelling System. Agronomie 19, 185e210.
Sens. 43, 571e580. Lewis, P., Gomez-Dans, J., Kaminski, T., Settle, J., Quaife, T.,
Kim, H., Liang, S., 2010. Development of a new hybrid method Gobron, N., Styles, J., Berger, M., 2012. An earth observa-
for estimating land surface shortwave net radiation from tion land data assimilation system (EO-LDAS). Remote
MODIS data. Remote Sens. Environ. 114, 2393e2402. Sens. Environ. 120, 219e235.
Kim, J., Mohanty, B.P., Shin, Y., 2015. Effective soil moisture Li, X., Strahler, A., 1985. Geometric-optical modeling of a
estimate and its uncertainty using multimodel simulation coniferous forest canopy. IEEE Trans. Geosci. Remote
based on Bayesian Model Averaging. J. Geophys. Res. At- Sens. 23, 705e721.
mospheres 120, 8023e8042. Li, X., Strahler, A., 1986. Geometric-optical bi-directional
Krahenmann, S., Obregon, A., Muller, R., Trentmann, J., reflectance modeling of a coniferous forest canopy. IEEE
Ahrens, B., 2013. A satellite-based surface radiation clima- Trans. Geosci. Remote Sens. 24, 906e919.
tology derived by combining climate data records and Li, Z.W., Xin, X.P., Tang, H., Yang, F., Chen, B.R.,
near-real-time data. Remote Sens. 5, 4693e4718. Zhang, B.H., 2017. Estimating grassland LAI using the
Kuhnlein, M., Appelhans, T., Thies, B., Nauss, T., 2014. Random Forests approach and Landsat imagery in the
Improving the accuracy of rainfall rates from optical sat- meadow steppe of Hulunber, China. J. Integr. Agric. 16,
ellite sensors with machine learning e a random forests- 286e297.
based approach applied to MSG SEVIRI. Remote Sens. Liang, S., 2001. An optimization algorithm for separating
Environ. 141, 129e143. land surface temperature and emissivity from multispec-
Kuter, S., Akyurek, Z., Weber, G.W., 2018. Retrieval of frac- tral thermal infrared imagery. IEEE Trans. Geosci. Remote
tional snow covered area from MODIS data by multivar- Sens. 39, 264e274.
iate adaptive regression splines. Remote Sens. Environ. Liang, S., 2003. A direct algorithm for estimating land surface
205, 236e252. broadband albedos from MODIS imagery. IEEE Trans.
Kuter, S., Weber, G.W., Akyurek, Z., Ozmen, A., 2015. Inver- Geosci. Remote Sens. 41, 136e145.
sion of top of atmospheric reflectance values by conic Liang, S., 2004. Quantitative Remote Sensing of Land
multivariate adaptive regression splines. Inverse Prob. Surfaces. John Wiley & Sons, Inc, New York.
Sci. Eng. 23, 651e669. Liang, S., 2007. Recent developments in estimating land sur-
Kuusk, A., 1995. A fast invertible canopy reflectance model. face biogeophysical variables from optical remote
Remote Sens. Environ. 51, 342e350. sensing. Prog. Phys. Geogr. 31, 501e516.
Kuusk, A., 2018. Canopy radiative transfer modeling. In: Liang, S., Fang, H., Chen, M., Shuey, C., Walthall, C.,
Chen, J. (Ed.), Terrestrial Ecosystems. In: Liang, S. (Ed.), Daughtry, C., 2002. Atmospheric correction of Landsat
Comprehensive Remote Sensing, vol. 3. Elsevier, pp. 9e22. ETMþ land surface imagery: II. Validation and
Laurent, V.C.E., Schaepman, M.E., Verhoef, W., applications. IEEE Trans. Geosci. Remote Sens. 40,
Weyermann, J., Chavez, R.O., 2014. Bayesian object- 2736e2746.
based estimation of LAI and chlorophyll from a simulated Liang, S., Li, X., Xie, X., 2013a. Land Surface Observation,
Sentinel-2 top-of-atmosphere radiance image. Remote Modeling and Data Assimilation. World Scientific.
Sens. Environ. 140, 318e329. Liang, S., Liu, Q., Yan, G., Shi, J., Kerekes, J.P., 2019. Fore-
Laurent, V.C.E., Verhoef, W., Damm, A., Schaepman, M.E., word to the special issue on the recent progress in quan-
Clevers, J., 2013. A Bayesian object-based approach for titative land remote sensing: modeling and estimation.
estimating vegetation biophysical and biochemical vari- IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens. 12,
ables from APEX at-sensor radiance data. Remote Sens. 391e395.
Environ. 139, 6e17.
52 1. A systematic view of remote sensing

Liang, S., Qin, J., 2008. Data assimilation methods for land Liu, Q., Wang, L., Qu, Y., Liu, N., Liu, S., Tang, H., Liang, S.,
surface variable estimation. In: Liang, S. (Ed.), Advances 2013b. Priminary evaluation of the long-term GLASS al-
in Land Remote Sensing: System, Modeling, Inversion bedo product. Int. J. Digital Earth 6, 69e95. https://
and Application. Springer, New York, pp. 313e339. doi.org/10.1080/17538947.17532013.17804601.
Liang, S., Shi, J., Yan, G., 2018. Recent progress in quantitative Liu, Y.M., Niu, B., Luo, Y.F., 2015b. Hybrid learning particle
land remote sensing in China. Remote Sens. 10, 1490. swarm optimizer with genetic disturbance. Neurocom-
Liang, S., Strahler, A., Walthall, C., 1999. Retrieval of land puting 151, 1237e1247.
surface albedo from satellite observations: a simulation Loew, A., Bell, W., Brocca, L., Bulgin, C.E., Burdanowitz, J.,
study. J. Appl. Meteorol. 38, 712e725. Calbet, X., Donner, R.V., Ghent, D., Gruber, A.,
Liang, S., Strahler, A.H., 1993a. An analytic BRDF model of Kaminski, T., Kinzel, J., Klepp, C., Lambert, J.C.,
canopy radiative transfer and its inversion. IEEE Trans. Schaepman-Strub, G., Schroder, M., Verhoelst, T., 2017.
Geosci. Remote Sens. 31, 1081e1092. Validation practices for satellite-based Earth observation
Liang, S., Strahler, A.H., 1993b. The calculation of the radi- data across communities. Rev. Geophys. 55, 779e817.
ance distribution of the coupled atmosphere-canopy. Lopez-Serrano, P.M., Lopez-Sanchez, C.A., Alvarez-
IEEE Trans. Geosci. Remote Sens. 31, 491e502. Gonzalez, J.G., Garcia-Gutierrez, J., 2016. A comparison
Liang, S., Strahler, A.H., 1994. Retrieval of surface BRDF from of machine learning techniques applied to Landsat-5 TM
multiangle remotely sensed data. Remote Sens. Environ. spectral data for biomass estimation. Can. J. Remote
50, 18e30. Sens. 42, 690e705.
Liang, S., Stroeve, J., Box, J.E., 2005. Mapping daily snow/ice Lu, D., Weng, Q., 2007. A survey of image classification
shortwave broadband albedo from Moderate Resolution methods and techniques for improving classification
Imaging Spectroradiometer (MODIS): the improved performance. Int. J. Remote Sens. 28, 823e870.
direct retrieval algorithm and validation with Greenland Lucia, D.J., Beran, P.S., Silva, W.A., 2004. Reduced-order
in situ measurement. J. Geophys. Res. Atmospheres 110. modeling: new approaches for computational physics.
Art. No. D10109. Prog. Aerosp. Sci. 40, 51e117.
Liang, S., Townshend, J.R.G., 1996. A modified Hapke model Ma, H., Liang, S., Xiao, Z., Shi, H., 2017a. Simultaneous inver-
for soil bidirectional reflectance. Remote Sens. Environ. sion of multiple land surface parameters from MODIS
55, 1e10. opticalethermal observations. ISPRS J. Photogrammetry
Liang, S., Zhang, X., Xiao, Z., Cheng, J., Liu, Q., Zhao, X., Remote Sens. 128, 240e254.
2013b. Global LAnd Surface Satellite (GLASS) Products: Ma, H., Liang, S., Xiao, Z., Wang, D., 2018. Simultaneous esti-
Algorithms, Validation and Analysis. Springer. mation of multiple land surface parameters from VIIRS
Liang, S., Zhao, X., Yuan, W., Liu, S., Cheng, X., Xiao, Z., optical-thermal data. IEEE Geosci. Remote Sens. Lett. 15,
Zhang, X., Liu, Q., Cheng, J., Tang, H., Qu, Y.H., Bo, Y., 151e160.
Qu, Y., Ren, H., Yu, K., Townshend, J., 2013c. A long- Ma, H., Liu, Q., Liang, S., Xiao, Z., 2017b. Simultaneous esti-
term global LAnd surface satellite (GLASS) dataset for mation of leaf area index, fraction of absorbed photosyn-
environmental studies. Int. J. Digital Earth 6, 5e33. thetically active radiation and surface albedo from
Liess, M., Schmidt, J., Glaser, B., 2016. Improving the spatial multiple-satellite data. IEEE Trans. Geosci. Remote Sens.
prediction of soil organic carbon stocks in a complex trop- 55, 4334e4354 doi:4310.1109/TGRS.2017.2691542.
ical mountain landscape by methodological specifications Machac, D., Reichert, P., Albert, C., 2016. Emulation of dy-
in machine learning approaches. PLoS One 11, 22. namic simulators with application to hydrology.
Liu, M., He, H.L., Ren, X.L., Sun, X.M., Yu, G.R., Han, S.J., J. Comput. Phys. 313, 352e366.
Wang, H.M., Zhou, G.Y., 2015a. The effects of constrain- Manfreda, S., McCabe, M.E., Miller, P.E., Lucas, R.,
ing variables on parameter optimization in carbon and Madrigal, V.P., Mallinis, G., Dor, E., Helman, D.,
water flux modeling over different forest ecosystems. Estes, L., Ciraolo, G., Mullerova, J., Tauro, F., de
Ecol. Model. 303, 30e41. Lima, M.I., del Lima, J., Maltese, A., Frances, F.,
Liu, N., Liu, Q., Wang, L., Liang, S., Wen, J., Qu, Y., Liu, S., Caylor, K., Kohv, M., Perks, M., Ruiz-Perez, G., Su, Z.,
2013a. A statistics-based temporal filter algorithm to map Vico, G., Toth, B., 2018. On the use of unmanned aerial
spatiotemporally continuous shortwave albedo from systems for environmental monitoring. Remote Sens. 10,
MODIS data. Hydrol. Earth Syst. Sci. 17, 2121e2129. 28.
https://doi.org/2110.5194/hess-2117-2121-2013. Marabel, M., Alvarez-Taboada, F., 2013. Spectroscopic deter-
Liu, Q., Liang, S., Xiao, Z.Q., Fang, H.L., 2014. Retrieval of mination of aboveground biomass in grasslands using
leaf area index using temporal, spectral, and angular spectral transformations, support vector machine and
information from multiple satellite data. Remote Sens. partial least squares regression. Sensors 13, 10027.
Environ. 145, 25e37.
 References 53
Mas, J.F., Flores, J.J., 2008. The application of artificial neural North, P.R.J., 1996. Three-dimensional forest light interaction
networks to the analysis of remotely sensed data. Int. J. model using a Monte Carlo method. IEEE Trans. Geosci.
Remote Sens. 29, 617e663. Remote Sens. 34, 946e956.
McCabe, M.F., Rodell, M., Alsdorf, D.E., Miralles, D.G., Omer, G., Mutanga, O., Abdel-Rahman, E.M., Adam, E.,
Uijlenhoet, R., Wagner, W., Lucieer, A., Houborg, R., 2016. Empirical prediction of leaf area index (LAI) of en-
Verhoest, N.E.C., Franz, T.E., Shi, J.C., Gao, H.L., dangered tree species in intact and fragmented indige-
Wood, E.F., 2017. The future of Earth observation in nous forests ecosystems using WorldView-2 data and
hydrology. Hydrol. Earth Syst. Sci. 21, 3879e3914. two robust machine learning algorithms. Remote Sens.
Mohammed, I.N., Bolten, J.D., Srinivasan, R., Lakshmi, V., 8, 26.
2018. Improved hydrological decision support system Omer, G., Mutanga, O., Abdel-Rahman, E.M., Peerbhay, K.,
for the lower Mekong river basin using satellite-based Adam, E., 2017. Mapping leaf nitrogen and carbon con-
earth observations. Remote Sens. 10, 17. centrations of intact and fragmented indigenous forest
Montesano, P.M., Cook, B.D., Sun, G., Simard, M., ecosystems using empirical modeling techniques and
Nelson, R.F., Ranson, K.J., Zhang, Z., Luthcke, S., 2013. WorldView-2 data. ISPRS J. Photogrammetry Remote
Achieving accuracy requirements for forest biomass map- Sens. 131, 26e39.
ping: a spaceborne data fusion method for estimating for- Ouyang, X.Y., Chen, D.M., Lei, Y.H., 2018. A generalized
est biomass and LiDAR sampling error. Remote Sens. evaluation scheme for comparing temperature products
Environ. 130, 153e170. from satellite observations, numerical weather model,
Moody, E.G., King, M.D., Platnick, S., Schaaf, C.B., Gao, F., and ground measurements over the Tibetan plateau.
2005. Spatially complete global spectral surface albedos: IEEE Trans. Geosci. Remote Sens. 56, 3876e3894.
value-added datasets derived from terra MODIS land Owe, M., de Jeu, R., Holmes, T., 2008. Multisensor historical
products. IEEE Trans. Geosci. Remote Sens. 43, 144e158. climatology of satellite-derived global land surface
Moosavi, V., Talebi, A., Mokhtari, M.H., Shamsi, S.R.F., moisture. J. Geophys. Res. 113, F01002.
Niazi, Y., 2015. A wavelet-artificial intelligence fusion Pan, Z.G., Glennie, C., Legleiter, C., Overstreet, B., 2015. Esti-
approach (WAIFA) for blending Landsat and MODIS sur- mation of water depths and turbidity from hyperspectral
face temperature. Remote Sens. Environ. 169, 243e254. imagery using support vector regression. IEEE Geosci.
Mora, B., Wulder, M.A., White, J.C., 2010. Segment- Remote Sens. Lett. 12, 2165e2169.
constrained regression tree estimation of forest stand Pasetto, D., Arenas-Castro, S., Bustamante, J., Casagrandi, R.,
height from very high spatial resolution panchromatic im- Chrysoulakis, N., Cord, A.F., Dittrich, A., Domingo-
agery over a boreal environment. Remote Sens. Environ. Marimon, C., Serafy, G., Karnieli, A., Kordelas, G.A.,
114, 2474e2484. Manakos, I., Mari, L., Monteiro, A., Palazzi, E.,
Mountrakis, G., Im, J., Ogole, C., 2011. Support vector ma- Poursanidis, D., Rinaldo, A., Terzago, S., Ziemba, A.,
chines in remote sensing: a review. ISPRS J. Photogram- Ziv, G., 2018. Integration of satellite remote sensing data
metry Remote Sens. 66, 247e259. in ecosystem modelling at local scales: practices and
Murray, N.J., Keith, D.A., Bland, L.M., Ferrari, R., trends. Methods Ecol. Evol. 9, 1810e1821.
Lyons, M.B., Lucas, R., Pettorelli, N., Nicholson, E., Pflugmacher, D., Cohen, W.B., Kennedy, R.E., Yang, Z., 2014.
2018. The role of satellite remote sensing in structured Using Landsat-derived disturbance and recovery history
ecosystem risk assessments. Sci. Total Environ. 619, and lidar to map forest biomass dynamics. Remote
249e257. Sens. Environ. 151, 124e137.
Mutanga, O., Adam, E., Cho, M.A., 2012. High density Piotrowski, A.P., Napiorkowski, J.J., 2013. A comparison of
biomass estimation for wetland vegetation using methods to avoid overfitting in neural networks training
WorldView-2 imagery and random forest regression in the case of catchment runoff modelling. J. Hydrol. 476,
algorithm. Int. J. Appl. Earth Obs. Geoinf. 18, 399e406. 97e111.
Myneni, R.B., Ross, J., Asrar, G., 1989. A review on the theory Posselt, R., Mueller, R., Trentrnann, J., Stockli, R.,
of photon transport in leaf canopies. Agric. For. Meteorol. Liniger, M.A., 2014. A surface radiation climatology
45, 1e153. across two Meteosat satellite generations. Remote Sens.
Nawar, S., Buddenbaum, H., Hill, J., Kozak, J., 2014. Environ. 142, 103e110.
Modeling and mapping of soil salinity with reflectance Qi, J., Xie, D., Yin, T., Yan, G., Gastellu-Etchegorry, J.-P.,
spectroscopy and Landsat data using two quantitative Li, L., Zhang, W., Mu, X., Norford, L.K., 2019. LESS:
methods (PLSR and MARS). Remote Sens. 6, LargE-Scale remote sensing data and image simulation
10813e10834. framework over heterogeneous 3D scenes. Remote Sens.
Environ. 221, 695e706.
54 1. A systematic view of remote sensing

Qin, J., Liang, S., Yang, K., Kaihotsu, I., Liu, R.G., Koike, T., Rivera, J.P., Verrelst, J., Gomez-Dans, J., Munoz-Mari, J.,
2009. Simultaneous estimation of both soil moisture and Moreno, J., Camps-Valls, G., 2015. An emulator toolbox
model parameters using particle filtering method through to approximate radiative transfer models with statistical
the assimilation of microwave signal. J. Geophys. Res. At- learning. Remote Sens. 7, 9347e9370.
mospheres 114, 13. Scholze, M., Buchwitz, M., Dorigo, W., Guanter, L.,
Qin, J., Yang, K., Lu, N., Chen, Y.Y., Zhao, L., Han, M.L., Quegan, S., 2017. Reviews and syntheses: systematic
2013. Spatial upscaling of in-situ soil moisture measure- Earth observations for use in terrestrial carbon cycle
ments based on MODIS-derived apparent thermal data assimilation systems. Biogeosci. Discuss. 1e49.
inertia. Remote Sens. Environ. 138, 1e9. Setiono, R., Wee Kheng, L., Zurada, J.M., 2002. Extraction of
Qin, J., Zhao, L., Chen, Y.Y., Yang, K., Yang, Y.P., Chen, Z.Q., rules from artificial neural networks for nonlinear
Lu, H., 2015. Inter-comparison of spatial upscaling regression. IEEE Trans. Neural Netw. 13, 564e577.
methods for evaluation of satellite-based soil moisture. Shi, H., Xiao, Z., Liang, S., Ma, H., 2017. A method for consis-
J. Hydrol. 523, 170e178. tent estimation of multiple land surface parameters from
Qin, W.H., Gerstl, S.A.W., 2000. 3-D scene modeling of semi- MODIS top-of-atmosphere time series data. IEEE Trans.
desert vegetation cover and its radiation regime. Remote Geosci. Remote Sens. 55, 5158e5173.
Sens. Environ. 74, 145e162. Shi, H., Xiao, Z., Liang, S., Zhang, X., 2016. Consistent estima-
Qu, Y., Liang, S., Liu, Q., Li, X., Feng, Y., Liu, S., 2016. Esti- tion of multiple parameters from MODIS top of atmo-
mating Arctic sea-ice shortwave albedo from MODIS sphere reflectance data using a coupled soil-canopy-
data. Remote Sens. Environ. 186, 32e46. atmosphere radiative transfer model. Remote Sens. Envi-
Qu, Y., Liu, Q., Liang, S., Wang, L., Liu, N., Liu, S., 2014a. ron. 184, 40e57.
Improved direct-estimation algorithm for mapping daily Shi, Q., Liang, S., 2013a. Characterizing the surface radiation
land-surface broadband albedo from MODIS data. IEEE budget over the Tibetan Plateau with ground-measured,
Trans. Geosci. Remote Sens. 52, 907e919. reanalysis, and remote sensing data sets: 1.
Qu, Y., Zhang, Y., Xue, H., 2014b. Retrieval of 30-m-resolution Methodology. J. Geophys. Res. Atmospheres 118,
leaf area index from China HJ-1 CCD Data and MODIS 9642e9657.
Products through a dynamic bayesian network. IEEE J. Shi, Q., Liang, S., 2013b. Characterizing the surface radiation
Sel. Top. Appl. Earth Obs. Remote Sens. 7, 222e228. budget over the Tibetan Plateau with ground-measured,
Qu, Y.H., Zhang, Y.Z., Wang, J.D., 2012. A dynamic Bayesian reanalysis, and remote sensing data sets: 2. Spatiotem-
network data fusion algorithm for estimating leaf area in- poral analysis. J. Geophys. Res.: Atmospheres 118,
dex using time-series data from in situ measurement to 8921e8934.
remote sensing observations. Int. J. Remote Sens. 33, Shi, Q., Liang, S., 2014. Surface sensible and latent heat fluxes
1106e1125. over the Tibetan Plateau from ground measurements,
Quaife, T., Lewis, P., 2010. Temporal constraints on linear reanalysis, and satellite data. Atmos. Chem. Phys. 14,
BRDF model parameters. IEEE Trans. Geosci. Remote 5659e5677.
Sens. 48, 2445e2450. Shiklomanov, A.N., Dietze, M.C., Viskari, T.,
Quan, X.W., He, B.B., Li, X., 2015. A bayesian network-based Townsend, P.A., Serbin, S.P., 2016. Quantifying the influ-
method to alleviate the ill-posed inverse problem: a case ences of spectral resolution on uncertainty in leaf trait es-
study on leaf area index and canopy water content timates through a Bayesian approach to RTM inversion.
retrieval. IEEE Trans. Geosci. Remote Sens. 53, Remote Sens. Environ. 183, 226e238.
6507e6517. Simmons, A., Fellous, J.L., Ramaswamy, V., Trenberth, K.,
Quan, X.W., He, B.B., Li, X., Liao, Z.M., 2016. Retrieval of Study Team Comm Space, R., 2016. Observation and inte-
grassland live fuel moisture content by parameterizing grated Earth-system science: a roadmap for 2016e2025.
radiative transfer model with interval estimated LAI. Adv. Space Res. 57, 2037e2103.
IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens. 9, Strahler, A.H., Woodcock, C.E., Smith, J.A., 1986. On the na-
910e920. ture of models in remote sensing. Remote Sens. Environ.
Rahman, M.S., Di, L.P., 2017. The state of the art of space- 20, 121e139.
borne remote sensing in flood management. Nat. Hazards Sun, G., Ranson, K.J., Guo, Z., Zhang, Z., Montesano, P.,
85, 1223e1248. Kimes, D., 2011. Forest biomass mapping from lidar and
Ramoelo, A., Cho, M.A., Mathieu, R., Madonsela, S., van de radar synergies. Remote Sens. Environ. 115, 2906e2916.
Kerchove, R., Kaszta, Z., Wolff, E., 2015. Monitoring grass Sun, J., Yang, J., Shi, S., Chen, B.W., Du, L., Gong, W.,
nutrients and biomass as indicators of rangeland quality Song, S.L., 2017a. Estimating rice leaf nitrogen concentra-
and quantity using random forest modelling and World tion: influence of regression algorithms based on passive
View-2 data. Int. J. Appl. Earth Obs. Geoinf. 43, 43e54. and active leaf reflectance. Remote Sens. 9, 15.
 References 55
Sun, L., Mi, X.T., Wei, J., Wang, J., Tian, X.P., Yu, H.Y., Verger, A., Filella, I., Baret, F., Pe~
nuelas, J., 2016. Vegetation
Gan, P., 2017b. A cloud detection algorithm-generating baseline phenology from kilometric global LAI satellite
method for remote sensing data at visible to short-wave products. Remote Sens. Environ. 178, 1e14.
infrared wavelengths. ISPRS J. Photogrammetry Remote Verhoef, W., 1984. Light scattering by leaf layers with appli-
Sens. 124, 70e88. cation to canopy reflectance modeling: the SAIL model.
Tanase, M.A., Panciera, R., Lowell, K., Tian, S., Hacker, J.M., Remote Sens. Environ. 16, 125e141.
Walker, J.P., 2014. Airborne multi-temporal L-band polar- Vermote, E., Kaufman, Y.J., 1995. Absolute calibration of
imetric SAR data for biomass estimation in semi-arid AVHRR visible and near-infrared channels using ocean
forests. Remote Sens. Environ. 145, 93e104. and cloud views. Int. J. Remote Sens. 16, 2317e2340.
Tang, W.J., Qin, J., Yang, K., Liu, S.M., Lu, N., Niu, X.L., 2016. Verrelst, J., Sabater, N., Rivera, J.P., Munoz-Mari, J., Vicent, J.,
Retrieving high-resolution surface solar radiation with Camps-Valls, G., Moreno, J., 2016. Emulation of leaf, can-
cloud parameters derived by combining MODIS and opy and atmosphere radiative transfer models for fast
MTSAT data. Atmos. Chem. Phys. 16, 2543e2557. global sensitivity analysis. Remote Sens. 8, 27.
Tao, L.L., Li, J., Jiang, J.B., Chen, X., 2016a. Leaf area index Wang, D., Liang, S., 2016. Estimating high-resolution top of
inversion of winter wheat using modified water-cloud atmosphere albedo from Moderate Resolution Imaging
model. IEEE Geosci. Remote Sens. Lett. 13, 816e820. Spectroradiometer data. Remote Sens. Environ. 178,
Tao, X., Liang, S., He, T., Wang, D., 2017. Integration of satel- 93e103.
lite fraction of absorbed photosynthetically active radia- Wang, D., Liang, S., 2017. Estimating top-of-atmosphere
tion products: method and validation. IEEE Trans. daily reflected shortwave radiation flux over land from
Geosci. Remote Sens. (revised). MODIS data. IEEE Trans. Geosci. Remote Sens. 55,
Tao, Y.M., Gao, X.G., Hsu, K.L., Sorooshian, S., Ihler, A., 4022e4031.
2016b. A deep neural network modeling framework to Wang, D., Liang, S., He, T., Cao, Y., Jiang, B., 2015a. Surface
reduce bias in satellite precipitation products. shortwave net radiation estimation from FengYun-3
J. Hydrometeorol. 17, 931e945. MERSI data. Remote Sens. 7, 6224e6239.
Tapiador, F.J., Navarro, A., Jimenez, A., Moreno, R., Garcia- Wang, D., Liang, S., He, T., Shi, Q., 2015b. Estimation of daily
Ortega, E., 2018. Discrepancies with satellite observations surface shortwave net radiation from the combined
in the spatial structure of global precipitation as derived MODIS data. IEEE Trans. Geosci. Remote Sens. 53,
from global climate models. Q. J. R. Meteorol. Soc. 144, 5519e5529.
419e435. Wang, D., Liang, S., He, T., Yu, Y., 2013a. Direct estimation of
Tinkham, W.T., Smith, A.M.S., Marshall, H.P., Link, T.E., land surface albedo from VIIRS data: algorithm improve-
Falkowski, M.J., Winstral, A.H., 2014. Quantifying spatial ment and preliminary validation. J. Geophys. Res. 118,
distribution of snow depth errors from LiDAR using 12577e12586.
Random Forest. Remote Sens. Environ. 141, 105e115. Wang, D., Liang, S., He, T., Yu, Y.Y., Schaaf, C., Wang, Z.S.,
Toth, C., Jozkow, G., 2016. Remote sensing platforms and 2015c. Estimating daily mean land surface albedo from
sensors: a survey. ISPRS J. Photogrammetry Remote MODIS data. J. Geophys. Res. Atmospheres 120,
Sens. 115, 22e36. 4825e4841.
Tramontana, G., Ichii, K., Camps-Valls, G., Tomelleri, E., Wang, D., Liang, S., Tao, H., 2014. Mapping high-resolution
Papale, D., 2015. Uncertainty analysis of gross primary surface shortwave net radiation from Landsat data.
production upscaling using Random Forests, remote IEEE Geosci. Remote Sens. Lett. 11, 459e463.
sensing and eddy covariance data. Remote Sens. Environ. Wang, K., Tang, R.L., Li, Z.L., 2013b. Comparison of inte-
168, 360e373. grating LAS/MODIS data into a land surface model for
Tsoularis, A., Wallace, J., 2002. Analysis of logistic growth improved estimation of surface variables through data
models. Math. Biosci. 179, 21e55. assimilation. Int. J. Remote Sens. 34, 3193e3207.
Varvia, P., Rautiainen, M., Sepp€anen, A., 2018. Bayesian esti- Wang, L.A., Zhou, X.D., Zhu, X.K., Dong, Z.D., Guo, W.S.,
mation of seasonal course of canopy leaf area index from 2016. Estimation of biomass in wheat using random forest
hyperspectral satellite data. J. Quant. Spectrosc. Radiat. regression algorithm and remote sensing data. Crop J. 4,
Transf. 208, 19e28. 212e219.
Verger, A., Baret, F., Weiss, M., Filella, I., Pe~
nuelas, J., 2015. Wang, L.A., Zhou, X.D., Zhu, X.K., Guo, W.S., 2017. Estima-
GEOCLIM: a global climatology of LAI, FAPAR, and tion of leaf nitrogen concentration in wheat using the MK-
FCOVER from VEGETATION observations for SVR algorithm and satellite remote sensing data. Comput.
1999e2010. Remote Sens. Environ. 166, 126e137. Electron. Agric. 140, 327e337.
56 1. A systematic view of remote sensing

Wang, Y., Yang, C., Li, X., 2008. Regularizing kernel-based Satellites. J. Geophys. Res. Atmospheres 119, 10,780-
BRDF model inversion method for ill-posed land surface 710,798, doi:710.1002/2014JD021814.
parameter retrieval using smoothness constraint. Xu, Y.M., Shen, Y., 2013. Reconstruction of the land surface
J. Geophys. Res. Atmospheres 113. temperature time series using harmonic analysis. Com-
Weng, Q.H., Fu, P., Gao, F., 2014. Generating daily land sur- put. Geosci. 61, 126e132.
face temperature at Landsat resolution by fusing Landsat Yang, B., Knyazikhin, Y., Mottus, M., Rautiainen, M.,
and MODIS data. Remote Sens. Environ. 145, 55e67. Stenberg, P., Yan, L., Chen, C., Yan, K., Choi, S.,
Wijayasekara, D., Manic, M., Sabharwall, P., Utgikar, V., Park, T., Myneni, R.B., 2017. Estimation of leaf area index
2011. Optimal artificial neural network architecture selec- and its sunlit portion from DSCOVR EPIC data: theoret-
tion for performance prediction of compact heat ical basis. Remote Sens. Environ. 198, 69e84.
exchanger with the EBaLM-OTR technique. Nucl. Eng. Yang, G.J., Pu, R.L., Huang, W.J., Wang, J.H., Zhao, C.J., 2010.
Des. 241, 2549e2557. A novel method to estimate subpixel temperature by
Wu, H., Zhang, X., Liang, S., Yang, H., Zhou, H., 2012. Esti- fusing solar-reflective and thermal-infrared remote-
mation of clear-sky land surface longwave radiation sensing data with an artificial neural network. IEEE
from MODIS data products by merging multiple Trans. Geosci. Remote Sens. 48, 2170e2178.
models. J. Geophys. Res. 117, D22107 doi:22110.21029/ Yang, K., Zhu, L., Chen, Y.Y., Zhao, L., Qin, J., Lu, H.,
22012JD017567. Tang, W.J., Han, M.L., Ding, B.H., Fang, N., 2016. Land
Wu, P.H., Shen, H.F., Zhang, L.P., Gottsche, F.M., 2015. Inte- surface model calibration through microwave data assim-
grated fusion of multi-scale polar-orbiting and geostation- ilation for improving soil moisture simulations. J. Hydrol.
ary satellite observations for the mapping of high spatial 533, 266e276.
and temporal resolution land surface temperature. Yao, X., Huang, Y., Shang, G.Y., Zhou, C., Cheng, T.,
Remote Sens. Environ. 156, 169e181. Tian, Y.C., Cao, W.X., Zhu, Y., 2015. Evaluation of six al-
Xia, J.Z., Ma, M.N., Liang, T.G., Wu, C.Y., Yang, Y.H., gorithms to monitor wheat leaf nitrogen concentration.
Zhang, L., Zhang, Y.J., Yuan, W.P., 2018. Estimates of Remote Sens. 7, 14939e14966.
grassland biomass and turnover time on the Tibetan Yao, Y., Liang, S., Li, X., Hong, Y., Fisher, J.B., Zhang, N.,
Plateau. Environ. Res. Lett. 13, 12. Chen, J., Cheng, J., Zhao, S., Zhang, X., Jiang, B.,
Xiao, Z., Liang, S., Wang, J., Xie, D., Song, J., Fensholt, R., Sun, L., Jia, K., Wang, K., Chen, Y., Mu, Q., Feng, F.,
2015. A framework for the simultaneous estimation of 2014. Bayesian multimodel estimation of global terrestrial
leaf area index, fraction of absorbed photosynthetically latent heat flux from eddy covariance, meteorological,
active radiation and albedo from MODIS time series and satellite observations. J. Geophys. Res. Atmospheres
data. IEEE Trans. Geosci. Remote Sens. 53, 3178e3197. 119, 2013JD020864.
Xie, X., Meng, S., Liang, S., Yao, Y., 2014. Improving stream- Yuan, H., Yang, G., Li, C., Wang, Y., Liu, J., Yu, H., Feng, H.,
flow predictions at ungauged locations with real-time Xu, B., Zhao, X., Yang, X., 2017. Retrieving soybean leaf
updating: application of an EnKF-based state-parameter area index from unmanned aerial vehicle hyperspectral
estimation strategy. Hydrol. Earth Syst. Sci. 18, 3923e3936. remote sensing: analysis of RF, ANN, and SVM regression
Xing, C.J., Chen, N.C., Zhang, X., Gong, J.Y., 2017. A machine models. Remote Sens. 9, 309.
learning based reconstruction method for satellite remote Zhang, P., Li, J., Olson, E., Schmit, T.J., Li, J.L., Menzel, W.P.,
sensing of soil moisture images with in situ observations. 2006. Impact of point spread function on infrared radi-
Remote Sens. 9, 24. ances from geostationary satellites. IEEE Trans. Geosci.
Xiong, X.X., Angal, A., Barnes, W.L., Chen, H., Chiang, V., Remote Sens. 44, 2176e2183.
Geng, X., Li, Y.H., Twedt, K., Wang, Z.P., Wilson, T., Zhang, D.Y., Zhang, W., Huang, W., Hong, Z.M., Meng, L.K.,
Wu, A.S., 2018. Updates of moderate resolution imaging 2017a. Upscaling of surface soil moisture using a deep
spectroradiometer on-orbit calibration uncertainty learning model with VIIRS RDR. ISPRS Int. J. Geo-Inf. 6, 20.
assessments. J. Appl. Remote Sens. 12, 18. Zhang, H.K.K., Huang, B., Zhang, M., Cao, K., Yu, L., 2015.
Xu, T., Bateni, S., Liang, S., 2015. Estimating turbulent heat A generalization of spatial and temporal fusion methods
fluxes with a weak-constraint data assimilation scheme: for remotely sensed surface parameters. Int. J. Remote
a case study (HiWATER-MUSOEXE). IEEE Geosci. Sens. 36, 4411e4445.
Remote Sens. Lett. 12, 68e72. Zhang, L.J., Dobslaw, H., Stacke, T., Guntner, A., Dill, R.,
Xu, T., Bateni, S.M., Liang, S., Entekhabi, D., Mao, K., 2014. Thomas, M., 2017b. Validation of terrestrial water storage
Estimation of surface turbulent heat fluxes via variational variations as simulated by different global numerical
assimilation of sequences of land surface temperatures models with GRACE satellite observations. Hydrol. Earth
from Geostationary Operational Environmental Syst. Sci. 21, 821e837.
 References 57
Zhang, L.P., Zhang, L.F., Du, B., 2016. Deep learning for Zhou, J., Jia, L., Menenti, M., 2015. Reconstruction of global
remote sensing data a technical tutorial on the state of MODIS NDVI time series: performance of harmonic
the art. IEEE Geosci. Remote Sens. Mag. 4, 22e40. ANalysis of time series (HANTS). Remote Sens. Environ.
Zhang, X., Liang, S., Wang, K., Li, L., Gui, S., 2010. Analysis 163, 217e228.
of global land surface shortwave broadband Albedo from Zhou, Y., Wang, D., Liang, S., Yu, Y., He, T., 2016. Assess-
multiple data sources. IEEE J. Special Top. Appl. Earth ment of the Suomi NPP VIIRS land surface albedo data
Obs. Remote Sens. 3, 296e305. using station measurements and high-resolution albedo
Zhang, X., Liang, S., Zhou, G., Wu, H., Zhao, X., 2014a. maps. Remote Sens. 8, 137.
Generating Global LAnd Surface Satellite incident short- Zhu, X.L., Chen, J., Gao, F., Chen, X.H., Masek, J.G., 2010. An
wave radiation and photosynthetically active radiation enhanced spatial and temporal adaptive reflectance
products from multiple satellite data. Remote Sens. Envi- fusion model for complex heterogeneous regions. Remote
ron. 152, 318e332. Sens. Environ. 114, 2610e2623.
Zhang, X.Y., Friedl, M.A., Schaaf, C.B., Strahler, A.H., Zhu, X.X., Tuia, D., Mou, L., Xia, G.S., Zhang, L., Xu, F.,
Hodges, J.C.F., Gao, F., Reed, B.C., Huete, A., 2003. Moni- Fraundorfer, F., 2017a. Deep learning in remote sensing:
toring vegetation phenology using MODIS. Remote Sens. a comprehensive review and list of resources. IEEE Geo-
Environ. 84, 471e475. sci. Remote Sens. Mag. 5, 8e36.
Zhang, Y., He, T., Liang, S., Wang, D., Yu, Y., 2018. Estima- Zhu, Y.H., Liu, K., Liu, L., Myint, S.W., Wang, S.G., Liu, H.X.,
tion of all-sky instantaneous surface incident shortwave He, Z., 2017b. Exploring the potential of WorldView-2
radiation from Moderate Resolution Imaging Spectroradi- red-edge band-based vegetation indices for estimation
ometer data using optimization method. Remote Sens. of mangrove leaf area index with machine learning
Environ. 209, 468e479. algorithms. Remote Sens. 9, 20.
Zhang, Y., Liang, S., Sun, G., 2014b. Mapping forest biomass Zhu, Z., Piao, S., Myneni, R.B., Huang, M., Zeng, Z.,
with GLAS and MODIS data over Northeast China. IEEE Canadell, J.G., Ciais, P., Sitch, S., Friedlingstein, P.,
J. Special Top. Appl. Earth Obs. Remote Sens. 7, Arneth, A., Cao, C., Cheng, L., Kato, E., Koven, C.,
140e152. Li, Y., Lian, X., Liu, Y., Liu, R., Mao, J., Pan, Y., Peng, S.,
Zhang, Y., Qu, Y., Wang, J., Liang, S., 2012. Estimating leaf Penuelas, J., Poulter, B., Pugh, T.A.M., Stocker, B.D.,
area index from MODIS and surface meteorological Viovy, N., Wang, X., Wang, Y., Xiao, Z., Yang, H.,
data using a dynamic Bayesian network. Remote Sens. Zaehle, S., Zeng, N., 2016. Greening of the earth and its
Environ. 127, 30e43. drivers. Nat. Clim. Chang. 6, 791e795.