remote sensing

Article
High-Resolution NDVI from Planet’s Constellation of
Earth Observing Nano-Satellites: A New Data Source
for Precision Agriculture
Rasmus Houborg * and Matthew F. McCabe
King Abdullah University of Science and Technology (KAUST), Water Desalination and Reuse Center (WDRC),
Biological and Environmental Science & Engineering (BESE), Thuwal 23955, Saudi Arabia;
[email protected]
* Correspondence: [email protected]; Tel.: +966-544-700-262

Academic Editors: Mutlu Ozdogan, Ioannis Gitas, Clement Atzberger and Prasad S. Thenkabail
Received: 5 July 2016; Accepted: 12 September 2016; Published: 19 September 2016

Abstract: Planet Labs (“Planet”) operate the largest fleet of active nano-satellites in orbit, offering
an unprecedented monitoring capacity of daily and global RGB image capture at 3–5 m resolution.
However, limitations in spectral resolution and lack of accurate radiometric sensor calibration impact
the utility of this rich information source. In this study, Planet’s RGB imagery was translated into
a Normalized Difference Vegetation Index (NDVI): a common metric for vegetation growth and
condition. Our framework employs a data mining approach to build a set of rule-based regression
models that relate RGB data to atmospherically corrected Landsat-8 NDVI. The approach was
evaluated over a desert agricultural landscape in Saudi Arabia where the use of near-coincident
(within five days) Planet and Landsat-8 acquisitions in the training of the regression models resulted
in NDVI predictabilities with an r2 of approximately 0.97 and a Mean Absolute Deviation (MAD)
on the order of 0.014 (~9%). The MAD increased to 0.021 (~14%) when the Landsat NDVI training
image was further away (i.e., 11–16 days) from the corrected Planet image. In these cases, the use of
MODIS observations to inform on the change in NDVI occurring between overpasses was shown to
significantly improve prediction accuracies. MAD levels ranged from 0.002 to 0.011 (3.9% to 9.1%)
for the best performing 80% of the data. The technique is generic and extendable to any region of
interest, increasing the utility of Planet’s dense time-series of RGB imagery.

Keywords: planet labs; Landsat; data mining; NDVI; precision agriculture; RGB; nano-satellites

1. Introduction
One of the constraints on the effective management and assessment of the terrestrial system at
sub-field scales (e.g., 1–10’s of meters) lies in the challenge of achieving both high spatial and high
temporal retrievals [1]. Satellite observations have made considerable progress in advancing such
a capacity, but are inevitably constrained by optics, bandwidth, or orbital configurations that limit
optimal spatio-temporal characteristics. One approach that has recently been proposed to overcome
such limitations is the use of so-called constellations of nano-satellites i.e., a large number of small,
compact sensor units (~10 kg) that are typically cheap, replaceable, and take advantage of economies
of scale [2]. In recent years, the capacity to survey the entire Earth at very high spatial resolution
and high frequency (daily) has approached operational reality, with a number of commercial efforts
exploring such an observation strategy.
One such company, Planet Labs (www.planet.com), operates the largest constellation of satellite
systems in orbit, with nano-satellite “Doves” collecting RGB (i.e., red, green, blue) imagery at 3–5 m
resolution on a daily scale, based on a full constellation of 150–200 satellites. Through such imagery,

Remote Sens. 2016, 8, 768; doi:10.3390/rs8090768 www.mdpi.com/journal/remotesensing

Remote Sens. 2016, 8, 768 2 of 19

the opportunities for change detection and surface characterization are enormous. From an agricultural
monitoring perspective for instance, timely and repeatable information on within-field variability in
growing conditions has specific utility in precision agriculture [3,4] as a means to optimize production
efficiencies via more sustainable and spatially explicit management practices [5,6]. Potential application
of such data span diverse aspects of earth observation and provide numerous opportunities for
advancing science and operational outcomes. Such an observational strategy represents a revolution
in earth observation.
Traditional single-source satellite missions are costly and physically constrained in terms of spatial
resolution and frequency of image capture. For instance, Landsat-8 is the continuation of an impressive
40-year mission heritage that has demonstrated great capacity for land surface characterization and
monitoring at high spatial resolution [7]. The advances that have been made with Landsat observations
result partly from the high quality sensor designs, high signal to noise ratios, and impressive absolute
radiometric accuracies and reflectance error specifications [8,9]. However, the 16-day revisit cycle
constitutes a major limitation for applications requiring time-critical information on surface conditions.
Improvements in spatio-temporal frequency will be realized by the pair of Sentinel-2 satellites,
which will deliver 10–60 m resolution multi-spectral data with a five-day revisit capability [10].
In addition, Landsat-8 and Sentinel-2 synergies offer opportunities to further enhance the temporal
resolution [11]. Nevertheless, the simultaneous requirements of very high spatial resolution (<10 m)
and near-daily frequency are currently only met through targeted acquisition via programmable
commercial multi-sensor systems such as WorldView and RapidEye [12]: and even then on a spatially
limited basis.
While Planet’s Doves will provide very high spatial resolution imagery at an unprecedented
frequency over the terrestrial surfaces of the Earth, the deliberately inexpensive sensor designs and
commercial off the shelf components do not match the signal-to-noise characteristics, radiometric
performance, cross-sensor consistency, and spectral enhancements of satellite imagers that comprise
more traditional space agency based operational missions. In addition, the lack of at-sensor radiance
conversions and atmospheric correction of the RGB imagery will affect the interpretation and time and
space consistency of any time-series data. Bands in the visible domain are particularly sensitive to the
atmospheric correction process [13,14] and failure to account for contaminations from time-varying
atmospheric influences may significantly impair the utility of the RGB imagery for reliably inferring
actual changes in surface cover conditions.
Broadband imaging in just the RGB domain presents a potential limitation for vegetation
monitoring and categorization. Such applications typically require information in the near-infrared
(NIR) to represent the reflectance plateau of green vegetation and is the spectral region most responsive
to changes in vegetation density [15]. The Normalized Difference Vegetation Index (NDVI), developed
based on pioneering work by Rouse et al. [16] and Tucker [17] using bands in the NIR and red domain,
i.e., NDVI = (NIR − red)/(NIR + red), has shown to be an effective and widely used indicator of
spatio-temporal changes in vegetation growth and distribution [18–20], vegetation stress [21–23],
and vegetation productivity [24–26]. However, indices based entirely on the visible range of the
spectrum have also demonstrated utility for assessing vegetation growth and development [27].
The Green-Red Vegetation Index serves as a visible analog to NDVI and has been shown to correlate
strongly with that index [17,28], while the Visible Atmospherically Resistant Index (VARI) has
demonstrated utility for estimating green vegetation fractions [27,29]. However, the robustness
of these indices for assessing vegetation density remains particularly challenged by confounding
factors introduced by the atmospheric medium [30] and pigment absorption [31], and significantly
reduced sensitivity to changes in vegetation density with increasing leaf area [32]. Important advances
in monitoring of vegetation status and retrieval of plant biochemical, structural, and physiological
quantities have been realized with the development of a variety of broadband and narrowband
vegetation indices that exploit specific regions and features of the electromagnetic spectrum [33–35].
Remote Sens. 2016, 8, 768 3 of 19

However, a standard index such as the NDVI continues to serve as a very simple and useful satellite
observable metric in the remote sensing of vegetation response [36].
A key objective of the current study is to exploit the high radiometric quality and near-infrared
imaging capability of Landsat-8 to maximize the utility of Planet’s RGB time series imagery for
agricultural applications. The capacity of the RGB to reproduce insights provided by traditional
Landsat NDVI dynamics was investigated using a data mining approach. An accurate translation of
the RGB imagery into maps of Landsat consistent NDVI is seen as a particularly useful step given
the recognition of NDVI as a key metric for the state and functioning of vegetation. To do this,
associations between the NDVI and RGB data were established on a scene-specific basis using
atmospherically corrected Landsat NDVI, in an attempt to minimize cross-sensor inconsistencies
(i.e., resulting from many individual Dove cameras) and time and space varying contaminations from
atmospheric influences. Coarser resolution (i.e., 250 m) NDVI from the Moderate Resolution Imaging
Spectroradiometer (MODIS) were also integrated into the downscaling framework to account for
surface variations in NDVI occurring between the Planet and Landsat scene acquisitions. An important
objective was to develop an operational framework, building upon an automated processing stream
and generic techniques that are capable of translating dense time-series of raw RGB imagery into
Landsat consistent and atmospherically corrected estimates of NDVI at 3–5 m resolution.

2. Data and Methodology

2.1. Study Region

The Tawdeehiya agricultural facility (24.174◦ N, 48.015◦ E) located southeast of Riyadh in
Saudi Arabia, was used as a test site for the analysis undertaken herein. Tawdeehiya is a commercial
farm occupying an area of approximately 120 km2 (12,000 ha) comprised of 47 fields employing center
pivot irrigation systems and planted with a rotation of alfalfa, corn, Rhodes grass, carrots, and other
vegetables and forage crops. The irrigated pivots (~800 m diameter) are situated within desert terrain
approximately 380 m a.s.l. and are exposed to harsh environmental conditions that include very limited
rainfall (~100 mm·year−1 ), extreme temperatures, and dust events [37]. The data used for this study
was collected across the period from 8 October 2015 to 1 January 2016 (day of year, DOY, 281 and 1,
respectively). The pivots were primarily planted with Alfalfa, corn, and Rhodes grass during the study
period. Cumulative rainfall was 27 mm and the average daytime high varied from ~35 ◦ C in October
to ~20 ◦ C in December.

2.2. Satellite Data

Clear-sky imagery from Planet’s Doves (Section 2.2.1), Landsat-8 (Section 2.2.2), and MODIS
(Section 2.2.3) were acquired within a 90-day window over the Tawdeehiya farm as detailed below.

2.2.1. Planet
A total of four scenes were acquired from Planet’s Dove satellites deployed into the International
Space Station Orbit (i.e., a low Earth orbit with an altitude of ~420 km). The Doves are characterized
as 3U CubeSats (i.e., 10 × 10 × 30 cm) and weigh around 4 kg. The scenes are captured in broad red
(610–700 nm), green (500–590 nm), and blue (420–530 nm) (RGB) wavelength bands with variable
view and sun illumination geometries and a ground sampling distance of approximately 3 m (Table 1).
The scenes are provided orthorectified in the Univeral Transverse Mercator (UTM) projection with
a geo-location accuracy approximated as 20 m relative to Landsat-8. While Planet’s images have
undergone a series of sensor and basic radiometric calibrations involving corrections for sensor
distortions and adjustments for sun angle variations, corrections for atmospheric effects or conversions
into top of atmosphere radiances have not been implemented. The analytic datasets are distributed as
digital counts in a 16-bit GeoTiff file format.
Remote Sens. 2016, 8, 768 4 of 19

Table 1. Acquisition details for the Planet (in italics) and Landsat-8 satellite images. DOY is the day of
year, θsz is the sun zenith angle, θvz is the view zenith angle, and GSD is the ground sampling distance.

Satellite Dates DOY Time θsz (◦ ) θvz (◦ ) GSD (m)

Landsat-8 8 October 2015 281 10:11 am 36.3 7.5 30
Landsat-8 15 October 2015 288 10:18 am 37.6 4.8 30
Planet 19 October 2015 292 10:27 am 30.1 3.2 2.6
Landsat-8 24 October 2015 297 10:11 am 40.9 7.4 30
Landsat-8 31 October 2015 304 10:17 am 42.3 4.9 30
Planet 14 November 2015 318 08:31 am 45.0 6.3 3.0
Landsat-8 16 November 2015 320 10:17 am 46.6 4.9 30
Planet 18 December 2015 352 09:36 am 47.7 2.4 3.0
Landsat-8 18 December 2015 352 10:17 am 52.1 4.9 30
Landsat-8 27 December 2015 361 10:11 am 53.1 7.4 30
Planet 1 January 2016 001 01:08 pm 67.1 8.4 3.0

2.2.2. Landsat-8
Multi-spectral data at 30 m resolution were acquired from the Ocean Land Imager (OLI) onboard
Landsat-8 for eight dates in 2015 (Table 1) via the Landsat archive on the Google cloud storage.
While the repeat cycle of Landsat-8 is 16 days, the location of the farm near the swath edges facilitates
acquisitions from two different swaths, with an optimal repeat cycle of eight days. The collected
scenes have been resampled using cubic convolution and processed to Level 1T by the United States
Geological Survey (USGS) to ensure systematic radiometric and geometric accuracy.
The positional accuracy has been reported as approximately 12 m [38]. Landsat red (640–670 nm),
green (530–590 nm), and blue (450–510 nm) bands are considerably narrower than the Planet
counterparts and have acquisition times that are fixed at approximately 10:15 am (local time). As the
farm is located in proximity to the swath edge, the exact view angle was calculated as a function of
satellite altitude, scene orientation, and distance to the center meridian [39] (Table 1).
The automated framework adopted for translating the at-sensor radiances into bidirectional
surface reflectances is detailed in Houborg et al. [39] and Houborg and McCabe [37]. In brief,
the 6SV atmospheric correction code [40,41] was used with inputs of acquisition specific view and
sun illumination angles, surface elevation from the GTOPO30 digital elevation model, total columnar
ozone from the Ozone Monitoring Instrument (OMI), total precipitable water and atmospheric profiles
from the Atmospheric Infrared Sounder (AIRS), and aerosol optical thickness at 550 nm (AOT550 )
from Terra MODIS deep blue algorithm retrievals [42]. The aerosol model was parameterized using
AOT550 -dependent aerosol radiative characteristics based on observations at the nearby Solar Village
AERONET site.
Adjacency effects can be substantial over the farm due to large reflectance contrasts between
dark vegetated pivots and surrounding bright desert surfaces. In fact, correction for adjacency effects
was shown to increase the NDVI by up to 50% over vegetated fields in a recent investigation [37].
This phenomenon was considered in accordance with Houborg and McCabe [37] using a scene and
AOT550 -specific optimization of the horizontal range of the adjacency effect.

2.2.3. MODIS
The 250 m resolution Terra MODIS surface reflectance product (MOD09Q1, collection 6) was
acquired for the corresponding Landsat eight-day collections over the duration of the study period
(i.e., total of 11 images). The MOD09Q1 product composites the best (i.e., low view angle, absence of
clouds, lowest aerosol loading) near-infrared (NIR) and red spectral observations collected during
an eight-day period, with corrections for atmospheric influences applied. Specific DOY records for
individual pixel observations collected during the eight-day retrieval period are not included in the
MOD09Q1. Instead, these were extracted from the 500 m surface reflectance product (MOD09A1)
and resampled to match the resolution of the MOD09Q1. While daily NDVI could be calculated
Remote Sens. 2016, 8, 768 5 of 18
Remote Sens. 2016, 8, 768 5 of 19

and resampled to match the resolution of the MOD09Q1. While daily NDVI could be calculated
directly from
directly from the
the MODIS
MODIS daily
daily surface
surface reflectance
reflectance product
product (MOD09GQ),
(MOD09GQ), existence
existence ofof clouds
clouds and
and
cloud shadows, high view angles, and high aerosol loadings all affect the utility of the daily products
cloud shadows, high view angles, and high aerosol loadings all affect the utility of the daily products
for accurately
for accurately describing
describing day-to-day
day-to-day dynamics
dynamics inin NDVI.
NDVI. A more complete
A more complete coverage
coverage ofof best
best quality
quality
observations represents a key advantage of using the eight-day products in the present
observations represents a key advantage of using the eight-day products in the present context. context.

2.3. Method Description

A multi-scale framework was developed with an aim to seamlessly translate broadband Planet
imageryinto
RGB imagery into Landsat-8
Landsat-8 consistent
consistent estimates
estimates of atmospherically
of atmospherically correctedcorrected NDVI.1 describes
NDVI. Figure Figure 1
describes this process. MODIS based NDVI estimates at 250 m resolution
this process. MODIS based NDVI estimates at 250 m resolution are integrated into this approach are integrated into thisto
approachfor
account topotential
account for potential
surface surface
variations in variations
NDVI thatinmight
NDVI that between
occur might occur between
the Planet andthe Planet
Landsat
and Landsat
scene scene This
acquisitions. acquisitions. This supplemental
supplemental MODIS processing MODIS processing
stream (Section stream
2.3.2) (Section 2.3.2) isifonly
is only invoked the
invoked if the Planet and Landsat acquisitions are more than four days apart.
Planet and Landsat acquisitions are more than four days apart. Otherwise, the Landsat-only processing Otherwise, the
Landsat-only
stream (Sectionprocessing
2.3.1) takesstream (Section
precedence. The2.3.1) takes
scheme precedence.
employs a dataThe scheme
mining employs
technique (seea Section
data mining
2.3.3)
technique
to (seeofSection
build a set 2.3.3)
rule-based to build a set
multi-variate of rule-based
regression multi-variate
models relating theregression
Planet RGB models
data torelating
Landsat-8 the
Planet RGB data to Landsat-8 consistent NDVI. The generated regression models
consistent NDVI. The generated regression models are specific to each Planet scene, thereby avoiding are specific to each
Planet scene,
potential thereby avoiding
uncertainties induced by potential uncertainties
variations induced
in the calibration andby variationscharacteristics
acquisition in the calibration and
(i.e., view
acquisition characteristics (i.e., view and illumination geometry) of each Dove satellite
and illumination geometry) of each Dove satellite sensor (Table 1). The end result is a high-resolution sensor (Table
1). The end result
Planet-based NDVI is amap
high-resolution
at 3 m ground Planet-based
samplingNDVI map at
distance. The3m ground
full sampling
processing distance.
stream has beenThe
full processing
established in anstream has been
automated andestablished
unsupervised in an automated
manner and unsupervised
to facilitate straightforward manner to facilitate
execution over
straightforward
any selected regionexecution over(see
of interest anyFigure
selected
1). region of interest (see Figure 1).

Figure 1. Overview of the multi-sensor processing framework used for converting the Planet (P) RGB
information into
information intoLandsat-8
Landsat-8 (L) consistent
(L) consistent estimates
estimates of atmospherically
of atmospherically correctedcorrected
NDVI. TheNDVI. The
integration
integration
of MODIS 250 of data
MODIS 250 data
is needed is needed
to correct to correct
for changes for conditions
in cover changes inoccurring
cover conditions occurring
between the Planet
between
and the Planet
Landsat and Landsatwhen
scene acquisitions scenethey
acquisitions
are more when they
than ~4 areapart.
days moreSpecific
than ~4 details
days apart. Specific
are provided
details
in are2.3.
Section provided in Section 2.3.
Remote Sens. 2016, 8, 768 6 of 19

2.3.1. Landsat-Only Processing Stream

When a Landsat-8 acquisition occurs within a few days of the Planet acquisition, changes in
surface conditions between the two satellite acquisitions are expected to be relatively minor. In this
case, the near-coincident Landsat scene is resized to the domain of the Planet imagery that has first
been resampled (using pixel aggregation) to the Landsat 30 m resolution, to ensure a scale consistent
comparison and regression analysis. Geolocation mismatches between the Planet and Landsat imagery
are minimized by applying appropriate pixel shifts (in x and y directions) based on optimizing the
goodness of fit between scatters of Planet and Landsat red reflectances [43,44]. Atmospheric correction
of the Landsat data is performed on the resized imagery as described in Section 2.2.2. Subsequently,
the atmospherically corrected NDVI is calculated and then directed to the Cubist regression modeling
as the target variable, the process of which is described in Section 2.3.3.

2.3.2. MODIS Processing Stream

When the acquisition time between the Planet and nearest valid Landsat-8 overpass is greater than
a few days (set here to four days), changes in surface conditions between overpasses may influence
the fidelity of the developed regressions. To account for this, the framework ingests higher frequency
MODIS 250 m resolution imagery to predict a Landsat scale NDVI at the Planet acquisition DOY
(PDOY ), that is more representative of the actual surface cover conditions for that specific day. For this
purpose, all Landsat and composited eight-day MODIS scenes within ±45 days of the Planet satellite
overpass are acquired. However, for near real-time processing the window may be extended further
backwards in time. The Landsat scenes are then resized, co-registered, and atmospherically corrected
as described in the single scene processing of Section 2.3.1. MODIS NDVI maps are produced from
NIR and red reflectance images, and subsequently resized to the common Landsat/Planet domain,
re-projected from Sinusoidal to UTM projection, and resampled (using a nearest neighbor approach)
to 30 m resolution. Precise co-registration of the resized MODIS and Landsat imagery is ensured by
identifying appropriate pixel shifts based on the maximum goodness of fit between a near-coincident
Landsat and MODIS NDVI scene for different 30 m pixel shifts. The resized MODIS eight-day NDVI
time-series dataset is then interpolated to the daily scale as follows. Each pixel observation within
a given eight-day period is associated with a specific DOY (i.e., corresponding to the best acquisition
conditions). Associated values of NDVI and DOY are collected for each of the eight-day products
over the scene domain, amounting to a total of 11 pairs (i.e., NDVI–DOY) for each pixel over a 90-day
window. The NDVI values are then interpolated to the daily scale by assuming a linear progression
between the observation DOYs. This process produces needed information on day-specific changes in
surface conditions and also serves to fill any existing gaps in the MODIS eight-day product (e.g., due to
clouds or low retrieval quality over the eight-day interval).
In order to reliably predict the change occurring at the Landsat scale, each Landsat pixel must
be associated with a MODIS 250 m pixel. Due to the mismatch in scale, the nearest 250 m pixel may
not be the best representation of the temporal vegetation dynamics occurring at the given Landsat
pixel location. Instead, the optimum MODIS pixel is identified based on the correlation between
time-series of day-coincident MODIS (interpolated) and Landsat NDVI within the neighborhood
(i.e., ±375 m, which corresponds to a 3 × 3 MODIS pixel box) of the given Landsat pixel. The MODIS
pixel associated with the highest correlation coefficient is then assigned as the best representation of
surface cover changes for the given Landsat pixel.
As a final step, Landsat scale NDVI at PDOY can be predicted by multiplying the Landsat NDVI
image closest to PDOY (i.e., LDOY ) by the relative change in MODIS NDVI (MNDVI) observed between
LDOY and PDOY (Figure 1). This can be formulated as

MNDVIPDOY [ xo, yo ] − MNDVILDOY [ xo, yo ]

 
NDVIPDOY [ x, y] = NDVILDOY [ x, y] · 1 + (1)
MNDVILDOY [ xo, yo ]
Remote Sens. 2016, 8, 768 7 of 19

where x and y represent the pixel location in the 2D image array and xo and yo represent the
corresponding pixel location of the associated MODIS pixel. The resulting Landsat scale NDVI
(NDVIPDOY ) is then used as the target variable in the Cubist regression modeling described below.

2.3.3. Cubist Regression Modeling

Cubist (RuleQuest; www.rulequest.com) represents a data mining technique for building
a collection of rule-based predictive models on the basis of a training dataset of target and input
explanatory variables. Each of the resulting rules represents a multi-variate linear regression model
that is used to calculate the predicted value whenever the conditions of the rule are met. The algorithm
is based on the model tree approach described in Quinlan [45,46]. Cubist is computationally fast and
specifically designed to efficiently handle very large datasets. As a result, it has been widely used in
remote sensing applications [47–49].
For the analysis undertaken here, Landsat NDVI (at or close to PDOY ) represents the target variable
and Planet RGB data represent the input explanatory variables. Cubist is applied to RGB data that
is resampled (using pixel aggregation) to 30 m resolution in order to be consistent with the scale of
the Landsat NDVI (Figure 1). The blue band does not provide relevant information in the context
of NDVI so is excluded from the list of explanatory attributes. The Green-Red Vegetation Index,
GRVI = (G − R)/(G + R) [17], is added as an input to the model training along with the squared and
cubed form of R, G, and GRVI. The GRVI is expected to act as a useful proxy for NDVI, as it records
changes in vegetation density via variations in the surface reflectance contrast between the green
reflectance peak and red reflectance minimum characteristic of dense canopies [28].
Cubist is operated in an iterative manner in order to learn the most effective regression
relationships based on pixels that present the most consistent patterns between the Planet and Landsat
imagery. For instance, pixels that have undergone change in surface cover between LDOY and PDOY ,
or that are being poorly predicted based on the MODIS information, should be excluded from the
model training. Initially, the target and explanatory data are extracted from all pixels within the
domain of the Planet scene, and Cubist is applied to the training dataset with the maximum number of
allowed rules (nrul) set to 2. If the resulting NDVI predictions diverge from the target Landsat-based
NDVI by more than 40% (i.e., in terms of the relative mean absolute deviation), those pixels are
removed from the next Cubist run. The nrul is gradually increased (to a maximum of 10) concurrent
with a decrease in the divergence threshold (to a minimum of 25%). The iterative process stops if
the overall relative mean absolute deviation falls below 10% or the maximum number of iterations is
reached (n = 10). Once completed, the developed rule-based regression models are applied to the full
scene 3 m resolution RGB imagery to enable computation of Landsat consistent NDVI estimates at
PDOY (Figure 1). Note that a new suite of rule-based regression models will be established for each
Planet scene.

2.4. Evaluation of Approach

The fidelity of the approach in translating Planet RGB imagery into Landsat consistent estimates
of NDVI is evaluated via scale consistent (i.e., 30 m) pixel-wise comparisons of the predicted (Planet)
and actual (Landsat) NDVI for near-coincident acquisitions. In the Landsat-only mode (Section 2.3.1),
the Landsat imagery at LDOY is used for both model training and validation of the NDVI predictions
on PDOY . The validation is done on a full scene basis and thus includes pixels not involved in the
training of the regression models. The traditional separation of the training database into independent
training and validation groups [37] is not relevant in this case, as the developed rule-based models
are specific to the given Planet scene and will not be applied to other data (i.e., another Planet scene).
As such, the resulting validation statistics represent the accuracy level expected when the Planet scene
is associated with a near-coincident Landsat-8 acquisition.
The approach is also tested using a Landsat acquisition that is temporally displaced (i.e., 11–16 days)
from the PDOY as the training base and using MODIS to inform on changes in NDVI occurring over
Remote Sens. 2016, 8, 768 8 of 19

the timespan between acquisitions. In this case, the imagery used for model training is independent of
the actual Landsat imagery near PDOY used for model validation.
Standard statistical metrics are used to assess the prediction performance. These include the
coefficient of determination (r2 ), the Mean Absolute Deviation (MAD), the relative MAD (calculated
by dividing the MAD with the mean of the observed Landsat NDVI values), and the relative Mean
Bias Deviation (MBD; positive if the Planet predictions overestimate the Landsat based observations).
In addition, density scatter plots and histograms of normalized probability density functions for
Landsat and Planet based NDVI are used to visualize the utility of the approach in reproducing
magnitudes and spatial features in NDVI. In the statistical evaluations, the worst performing 1% of
data points were removed from the analyses to limit the impact of extreme outliers at pivot boundaries
resulting from co-registration, scale, and pixel resampling issues.

3. Results

3.1. Case of Near-Coincident Acquisitions

In the ideal scenario, a clear-sky Landsat-8 image would be available within a few days of the
Planet acquisition. In this case, changes in surface cover conditions (i.e., NDVI) between acquisitions
will typically be minor and the training of the rule-based regression models will be based on observed
NDVI spatial dynamics. Figure 2a shows true color (i.e., RGB) images over the farm from Landsat-8 and
Planet acquired two days apart. As can be seen, the spatial features are generally similar, with dense
vegetation pivots appearing dark (due to radiation absorption by chlorophyll pigments) and the
surrounding highly reflective desert soil appearing bright. Some of the sparsely vegetated pivots also
appear in darker colors, which is a result of the active irrigation. In fact, the position of the center
pivot irrigation arm and the clockwise rotation of the center pivot can be observed, with darker shades
trailing and brighter shades leading the irrigation arm (Figure 2a insert).
Figure 2b compares the Landsat-8 derived NDVI on DOY 320 with the Planet NDVI on DOY 318.
The Planet NDVI was calculated via Cubist, using Landsat-8 NDVI from DOY 320 as the target
variable and the Planet R, G, and GRVI information as the explanatory variables (see Section 2.3.3).
The application of Cubist to the training dataset (total of 139,200 30 m resolution pixels; ~95% of
all pixels) resulted in 10 rules, each associated with a unique multi-variate linear regression model.
As an example, rule 8 is established based on 3535 pixels, has a target NDVI varying between 0.327
and 0.943 and is expressed by:

NDVIPlanet = 2.6827423 − 0.015379 · R + 0.013228 · G − 33.48 · GRVI + 88.5 · GRVI2 (2)

In converting RGB imagery, rule 8 is used whenever R ≤ 1011 and G > 997 and GRVI ≤ 0.09592
apply. The R, G, and RGVI attributes were used in all of the rules, whereas R2 and RGVI2 were
used in 40% and 60% of the rules, respectively. Evaluation of the converted Planet NDVI map
against the training data resulted in a coefficient of determination of 0.99 for this particular example.
Visually, the spatial features and magnitudes of the Landsat NDVI are accurately reproduced by
the Planet predicted NDVI (Figure 2). As a counter-example, the NDVI of two pivots towards the
northwestern edge of the farm (indicated by white arrows) differ significantly between the two
images. These discrepancies are attributed to partial harvesting of the fields after the Planet acquisition,
which can also be seen in the true color maps. Another distinguishing feature is the relatively high
NDVI recorded in the Planet image over the paved road surfaces (Figure 2). The red and green band
reflectances are relatively low and mostly the same over these surfaces, which may be mistaken for
vegetation development at low to intermediate stages. In these cases, NIR capability is needed to
accurately reproduce the true NDVI.
The benefits of the enhanced resolution (3 m) are clearly evident in the level of spatial detail
discernable from the zoomed NDVI images of Figure 2b. While a high degree of within-field variability
resulting from irrigation and soil variability can be seen in both the Landsat and Planet imagery,
Remote Sens. 2016, 8, 768 9 of 19

Remote
the Sens. 2016,
fine-scale 8, 768 of center-pivot irrigation wheel trenches on the NDVI signal is only evident
impact 9 ofin
18

the Planet imagery (rightmost zoom in Figure 2b). In addition, variations in NDVI dynamics at the
dynamics at the Landsat sub-pixel scale occur in several places, which provide important insight
Landsat sub-pixel scale occur in several places, which provide important insight into crop conditions
into crop conditions for precision management purposes.
for precision management purposes.
The density scatter plot of NDVI for all pixels (30 m) over the full domain of the farm
The density scatter plot of NDVI for all pixels (30 m) over the full domain of the farm demonstrates
demonstrates significant utility of the rule-based regression approach in predicting Landsat-8 NDVI
significant utility of the rule-based regression approach in predicting Landsat-8 NDVI from Planet’s
from Planet’s RGB imagery, with a low bias (1.6%) and a relative MAD of 8.5% (Figure 3a). In fact,
RGB imagery, with a low bias (1.6%) and a relative MAD of 8.5% (Figure 3a). In fact, the large
the large majority (i.e., 80%) of the data points are characterized by an r2 of 0.998 and a MAD of only
majority (i.e., 80%) of the data points are characterized by an r2 of 0.998 and a MAD of only 0.005
0.005 (3.9%) (Table 2). Comparative normalized frequency distributions of actual (Landsat on DOY
(3.9%) (Table 2). Comparative normalized frequency distributions of actual (Landsat on DOY 320) and
320) and predicted (Planet on DOY 318) NDVI also demonstrate a close correspondence between the
predicted (Planet on DOY 318) NDVI also demonstrate a close correspondence between the estimates
estimates both at the lower and higher extremes of the NDVI range (Figure 3b).
both at the lower and higher extremes of the NDVI range (Figure 3b).

Figure 2. (a) True color RGB imagery from Landsat-8 and Planet acquisitions two days apart.
Figure 2. (a) True color RGB imagery from Landsat-8 and Planet acquisitions two days apart. The
The insert in the Planet image is included to demonstrate the impact of pivot irrigation on the RGB
insert in the Planet image is included to demonstrate the impact of pivot irrigation on the RGB signal.
signal; (b) Intercomparison of Landsat-8 NDVI and Planet NDVI computed from RGB imagery using
(b) Intercomparison of Landsat-8 NDVI and Planet NDVI computed from RGB imagery using Cubist
Cubist regressions generated on the basis of Landsat-8 training data from DOY 320. The two white
regressions generated on the basis of Landsat-8 training data from DOY 320. The two white arrows
arrows highlight fields that have undergone significant change in cover conditions between the two
highlight fields that have undergone significant change in cover conditions between the two
acquisitions. Zoomed images are included to demonstrate the enhanced spatial detail achievable with
acquisitions. Zoomed images are included to demonstrate the enhanced spatial detail achievable
Planet data.
with Planet data.
 Remote Sens. 2016, 8, 768 10 of 18
Remote Sens. 2016, 8, 768 10 of 19

Figure 3. (a) Density scatter plot intercomparing Landsat-8 and Planet NDVI. The Planet NDVI at
Figure 3. (a) Density scatter plot intercomparing Landsat-8 and Planet NDVI. The Planet NDVI at
DOY 318 was estimated from multi-variate regression models using Landsat NDVI from DOY 320 as
DOY 318 was estimated from multi-variate regression models using Landsat NDVI from DOY 320 as
training data; (b) Normalized frequency distributions of the Landsat-8 and Planet NDVI using DOY 320
training data. (b) Normalized frequency distributions of the Landsat-8 and Planet NDVI using DOY
as the Landsat training base. Note that the Planet data have been resampled to 30 m resolution to
320 as the Landsat training base. Note that the Planet data have been resampled to 30 m resolution to
facilitate a scale consistent intercomparison and statistic evaluation.
facilitate a scale consistent intercomparison and statistic evaluation.

Table 2. Statistical metrics of the agreement in NDVI between four Planet acquisition days (PDOY ) and
3.2. Case When Satellite Acquisitions are Further Apart
near-coincident (i.e., left columns) Landsat-8 acquisitions (LDOY ). Landsat NDVI close to (0–5 days) or
Whenaway
further the (11–16
Planetdays)
andfrom
Landsat-8
PDOY wasacquisitions areCubist
used to train the further apart,models.
regression changes in latter
In the surface
case, cover
MODIS NDVI
conditions can bewas used to inform
significant, on will
which any change
impactoccurring between
the utility of theLLandsat-only
DOY and PDOY . processing
The statisticsstream
are in
based on Planet
reproducing data resampled
the NDVI dynamics to on30Pm,
DOYand
. Toare reported
examine for the
such full set ofNDVI
a scenario, data points (i.e.,
at PDOY (DOY99%)318)
in was
addition
predicted to a reduced
using LandsatsetNDVI
incorporating
training80%
dataoffrom
the data.
DOY 304. As seen in Figure 4a, the majority of the
pivots on the eastern half of the farm experience a significant green up between DOY 304 and 320.
This is reflectedPDOY :
in the 292
Planet NDVI 318
predictions (Figure 4c, top plot),352
which underestimate 001
the actual
NDVI magnitudes, LDOY : particularly
288 281 * those
over 320areas 304 *
(Figure 352 bottom
4a, 336 *plot). 361
Improved 352 NDVI
*
r2
predictability is achieved when the0.843
0.907 day-specific
0.977MODIS 0.957NDVI 0.967
imagery 0.918
(Figure 4b) is used 0.950
0.956 to inform
upon MAD 0.032 0.043 0.014 0.019 0.014 0.024 0.016 0.022
99% the
incl.Landsat-scale changes in NDVI occurring between DOY 304 and PDOY. In that case, the
MAD (%) 18.5 25.1 8.5 11.6 9.5 16.4 9.8 13.6
NDVI magnitudes are more accurately reproduced and the spatial features in the predicted NDVI
MBD (%) 8.5 9.4 1.6 5.1 1.3 10.9 1.1 −5.4
(Figure 4c, bottom plot) are generally consistent with the observed patterns in the Landsat NDVI on
2 0.993
DOY 320 (Figurer4a, bottom plot). 0.978 0.998 0.997 0.995 0.991 0.991 0.987
MAD 0.010 0.013 0.005 0.005 0.002 0.008 0.006 0.011
80% incl.
MAD (%) 7.3 11.1 3.9 4.5 4.6 7.7 4.9 9.1
Table 2. Statistical
MBD (%)metrics 1.4of the agreement
3.1 in NDVI between
0.2 −0.2 four0.5 Planet acquisition
5.4 days
1.0 (PDOY−
) and
7.6
near-coincident (i.e., left columns) Landsat-8 acquisitions (LDOY). Landsat NDVI close to (0–5 days) or
* Predictions with the MODIS processing stream invoked (see Section 2.3.2 and Figure 1).
further away (11–16 days) from PDOY was used to train the Cubist regression models. In the latter
case, MODIS NDVI was used to inform on any change occurring between LDOY and PDOY. The
statistics
3.2. Case Whenare basedAcquisitions
Satellite on Planet dataAreresampled to 30 m, and are reported for the full set of data points
Further Apart
(i.e., 99%) in addition to a reduced set incorporating 80% of the data.
When the Planet and Landsat-8 acquisitions are further apart, changes in surface cover conditions
PDOY:
can be significant, which will292
impact the utility of 318
the Landsat-only 352 processing stream in 001
reproducing
LDOY: 288 281 * 320 304 * 352 336 * 361 352 *
the NDVI dynamics on PDOY . To examine such a scenario, NDVI at PDOY (DOY 318) was predicted
r2 0.907 0.843 0.977 0.957 0.967 0.918 0.956 0.950
99% incl.

using Landsat NDVI training

MAD 0.032
data0.043
from DOY 304. As0.019
0.014
seen in Figure
0.014
4a, the majority0.016
0.024
of the pivots
0.022
on the
eastern halfMAD
of the farm
(%) experience
18.5 a
25.1significant
8.5 green up
11.6 between
9.5 DOY 304
16.4and 320. This
9.8 is reflected
13.6 in
the Planet NDVI
MBD (%] predictions
8.5 (Figure
9.4 4c, top1.6plot), which
5.1 underestimate
1.3 the actual 1.1
10.9 NDVI magnitudes,
−5.4
particularly over
r2 those areas
0.993 (Figure
0.9784a, bottom
0.998 plot).0.997
Improved NDVI predictability
0.995 0.991 is achieved
0.991 0.987 when
80% incl.

the day-specific
MAD MODIS0.010
NDVI imagery
0.013 (Figure
0.005 4b) is 0.005
used to inform
0.002 upon the Landsat-scale
0.008 0.006 changes
0.011 in
MAD (%) 7.3 11.1 3.9 4.5 4.6 7.7
NDVI occurring between DOY 304 and PDOY . In that case, the NDVI magnitudes are more accurately 4.9 9.1
reproducedMBD and(%) 1.4
the spatial 3.1 in the0.2predicted
features −0.2NDVI (Figure
0.5 4c,5.4bottom 1.0
plot) are −7.6
generally
* Predictions with the MODIS processing stream invoked (see Section 2.3.2
consistent with the observed patterns in the Landsat NDVI on DOY 320 (Figure 4a, bottom plot). and Figure 1).
Remote Sens. 2016, 8, 768 11 of 19
Remote Sens. 2016, 8, 768 11 of 18

Figure4.4.(a)
Figure (a)Landsat-8
Landsat-8NDVI
NDVIon onDOY
DOY304 304(top)
(top)and
and320
320(bottom);
(bottom).(b)
(b)Day-specific
Day-specific(i.e.,
(i.e.,interpolated)
interpolated)
MODISNDVI
MODIS NDVIononDOY DOY 304304 (top)
(top) andand
318318 (bottom).
(bottom); (c) Planet
(c) Planet NDVINDVI on DOY
on DOY 318 predicted
318 predicted on theon the
basis
basis
of of Landsat-8
Landsat-8 NDVI
NDVI from DOYfrom DOY
304. The304. Theplot
bottom bottom plot demonstrates
demonstrates the utilitythe utilityhigher
of using of using higher
frequency
frequency
MODIS dataMODIS datathe
to inform to Cubist-based
inform the Cubist-based regression
regression models whenmodels when near-coincident
near-coincident Landsat andLandsat
Planet
and Planet acquisitions
acquisitions do not exist. do not exist.

While the trends in NDVI recorded in the MODIS imagery are consistent with the Landsat
While the trends in NDVI recorded in the MODIS imagery are consistent with the Landsat
observations, peak MODIS NDVI values over the vegetated pivots (~0.8) differ significantly from the
observations, peak MODIS NDVI values over the vegetated pivots (~0.8) differ significantly from
corresponding Landsat values (~0.95) (Figure 4a,b). These high Landsat NDVI values agree with
the corresponding Landsat values (~0.95) (Figure 4a,b). These high Landsat NDVI values agree with
field collected spectroradiometer measurements over the farm (not shown), and are the result of
field collected spectroradiometer measurements over the farm (not shown), and are the result of
well-watered dense canopies and generally dark (i.e., wet) soil backgrounds. The impact of
well-watered dense canopies and generally dark (i.e., wet) soil backgrounds. The impact of adjacency
adjacency effects can be substantial over vegetated pivots surrounded by bright desert soils [39], and
effects can be substantial over vegetated pivots surrounded by bright desert soils [37], and algorithm
algorithm differences in the treatment of this complex phenomenon provide a plausible reason for
differences in the treatment of this complex phenomenon provide a plausible reason for the observed
the observed discrepancy between the MODIS and Landsat-based NDVI estimates. Regardless,
discrepancy between the MODIS and Landsat-based NDVI estimates. Regardless, these issues are
these issues are expected to have a relatively minor influence on the results, as the framework
expected to have a relatively minor influence on the results, as the framework utilizes the relative
utilizes the relative change in MODIS NDVI (Equation (1)).
change in MODIS NDVI (Equation (1)).
The density scatter plots of observed Landsat NDVI (DOY 320) versus the Landsat-only (DOY
The density scatter plots of observed Landsat NDVI (DOY 320) versus the Landsat-only (DOY 304)
304) and MODIS informed NDVI predictions on PDOY (318) (Figure 5a) demonstrate the benefit of
and MODIS informed NDVI predictions on PDOY (318) (Figure 5a) demonstrate the benefit of
integrating MODIS information, with the r2 increasing from 0.92 to 0.96 and the relative MAD
integrating MODIS information, with the r2 increasing from 0.92 to 0.96 and the relative MAD
decreasing from 14.9% to 11.6%. The normalized frequency distribution for the Landsat-only
decreasing from 14.9% to 11.6%. The normalized frequency distribution for the Landsat-only
predictions indicates a displacement of the frequency peak at the high end of the NDVI range
predictions indicates a displacement of the frequency peak at the high end of the NDVI range relative to
relative to the observed NDVI (Figure 5b), consistent with the underestimation of NDVI over the
the observed NDVI (Figure 5b), consistent with the underestimation of NDVI over the dense vegetation
dense vegetation pivots (Figure 4c, top plot). Fortunately, this issue is largely resolved when both
pivots (Figure 4c, top plot). Fortunately, this issue is largely resolved when both Landsat and MODIS
Landsat and MODIS NDVI are employed in the prediction process (Figure 5b).
NDVI are employed in the prediction process (Figure 5b).

Figure 5. (a) Density scatter plots intercomparing Landsat-8 and Planet NDVI. In the leftmost plot,
Planet NDVI on DOY 318 was estimated from Cubist regression models using Landsat NDVI from
integrating MODIS information, with the r2 increasing from 0.92 to 0.96 and the relative MAD
decreasing from 14.9% to 11.6%. The normalized frequency distribution for the Landsat-only
predictions indicates a displacement of the frequency peak at the high end of the NDVI range
relative to the observed NDVI (Figure 5b), consistent with the underestimation of NDVI over the
dense vegetation pivots (Figure 4c, top plot). Fortunately, this issue is largely resolved when12both
Remote Sens. 2016, 8, 768 of 19
Landsat and MODIS NDVI are employed in the prediction process (Figure 5b).

Figure 5. (a) Density scatter plots intercomparing

intercomparing Landsat-8
Landsat-8 and
and Planet
Planet NDVI.
NDVI. In
In the leftmost
leftmost plot,
plot,
Planet NDVI on DOY 318 was estimated from Cubist regression models using Landsat NDVI from
DOY 304 as training data. In the rightmost plot, MODIS NDVI from DOY 304 and 318 were used
to inform on the change in cover conditions occurring between the Landsat and Planet overpasses;
(b) Normalized frequency distributions of the Landsat-8 and Planet NDVI using DOY 304 as the
Landsat training base. Note that the Planet data have been resampled to 30 m resolution to facilitate
a scale consistent intercomparison and statistic evaluation.

3.3. Planet NDVI Time-Series

Figure 6 displays spatio-temporal features in (a) MODIS, (b) Landsat-8 and (c,d) Planet NDVI
over the eastern part of the farm during the period of study, and showcases the two procedures
(Landsat only or combination of Landsat and MODIS) for generating time-series of Planet NDVI. In all
cases, each Cubist simulation is independent, with resultant rules specific to the given Planet scene.
In Figure 6c, a near-coincident Landsat NDVI image was used to train the rule-based NDVI predictive
models, whereas a Landsat scene further away from PDOY together with MODIS NDVI imagery from
LDOY and PDOY were used to produce the Planet NDVI maps in Figure 6d.
Rapid changes in NDVI are evident in the Landsat-8 based time-sequence of maps (Figure 6b).
The majority of the fields (predominantly alfalfa) are harvested between DOY 281 and 288, followed
by a period of rapid green up. The mature alfalfa fields are harvested again between DOY 336 and 352.
These same tendencies are evident in the coarser scale MODIS NDVI images (Figure 6a). Planet NDVI
predictions on DOY 292 (Figure 6c,d) differ significantly from the Landsat NDVI on DOY 288 over
some of the pivots, as illustrated by a relative MAD of 18.5% (c) and 25.1% (d) (Table 2). The relatively
poor performance is believed to be the result of rapid changes in NDVI occurring between DOY 288
and 292 (Figure 6a,b), which partly invalidates the use of the Landsat NDVI on DOY 288 as the basis
for evaluating NDVI predictions on DOY 292.
As previously demonstrated, the predicted Planet NDVI on DOY 318 closely resembles the
observed NDVI on DOY 320 for both cases (Figure 6c,d). Planet NDVI predictions on DOY 352, based
on day-coincident Landsat NDVI training data, accurately reproduce the observed NDVI (Figure 6c)
with an r2 of 0.967, a MAD of 0.014 and a relative MAD of 9.5% (Table 2). Planet NDVI predictions
based on Landsat NDVI from DOY 336 and MODIS NDVI from DOY 336 and 352 are characterized by
a larger relative MAD (16.5%) (Table 2), which is mainly attributable to a positive bias (i.e., NDVI is
overestimated by 10.9%) over low-to-intermediate NDVI pixels (Figure 6d and Table 2). On DOY 001
(2016), the two approaches provide very similar NDVI predictions over the displayed section of the
farm that are in good correspondence with the observed NDVI on DOY 361 (Figure 6b–d and Table 2).
Remote Sens. 2016, 8, 768 13 of 19

Based on these results (excluding DOY 292), Planet NDVI time-series data can be predicted with
an overall MAD of 0.014 (9.2%) when a near-coincident (i.e., 0–5 days) Landsat NDVI image is used to
train the regression models (Table 3). Robust predictabilities are observed for all of the dominant cover
types (i.e., MAD varying from 7.8% to 11.2%) (Table 3), and the majority (i.e., 80%) of the data points are
described by a relative MAD less than 5% (Table 2). The level of overall agreement decreases to a MAD
of 0.021 (13.7%) when the Landsat NDVI training image is further away (i.e., 11–16 days) from PDOY
and the MODIS processing stream is invoked (Table 3). In these cases, the Planet NDVI predictions
tend to overestimate the Landsat observations over alfalfa (MBD = 8.5%) and corn (MBD = 14.5%),
which reflect uncertainties associated with describing the actual rate of change in cover specific NDVI
between the Landsat training image and PDOY . Regardless, the majority of the MODIS informed
predictions reproduce the observed NDVI with an r2 ≈ 0.99 and a relative MAD ranging from 4.5% to
Remote(Table
9.1% Sens. 2016,
2). 8, 768 13 of 18

Figure 6. Time-series of (a) 250 m resolution MODIS; (b) 30 m resolution Landsat-8; and (c,d) 3 m
Figure 6. Time-series of (a) 250 m resolution MODIS, (b) 30 m resolution Landsat-8, and (c,d) 3 m
resolution Planet NDVI over a subset of the farm. In (c), NDVI was predicted on the basis of a
resolution Planet NDVI over a subset of the farm. In (c), NDVI was predicted on the basis of a
near-coincident (i.e., within 0–5 days) Landsat acquisition that is indicated by a red boundary box (b).
near-coincident (i.e., within 0–5 days) Landsat acquisition that is indicated by a red boundary box
In (d), the Landsat-8 NDVI image used as training base is within 11–16 days of the Planet acquisition,
(b). In (d), the Landsat-8 NDVI image used as training base is within 11–16 days of the Planet
and MODIS NDVI is used to inform on the change occurring over these timespans.
acquisition, and MODIS NDVI is used to inform on the change occurring over these timespans.

Table 3. Overall (i.e., Planet DOY 318, 352 and 001) NDVI prediction results on a cover specific basis.
Results based on the Landsat-only (L8) and MODIS informed (M) processing streams are reported in
separate columns (i.e., L8 and L8 + M). The statistics are based on Planet data resampled to 30 m, and
are reported for the full set of data points (i.e., 99%). The number of pixel samples (n), and average
NDVI and standard deviation of each cover type are also given.

All covers Alfalfa Grass Corn Bare/Sparse

L8 L8 + M L8 L8 + M L8 L8 + M L8 L8 + M L8 L8 + M
r2 0.970 0.941 0.953 0.941 0.935 0.874 0.959 0.892 0.750 0.636
MAD 0.014 0.021 0.043 0.07 0.048 0.066 0.033 0.066 0.009 0.013
MAD [%] 9.2 13.7 9.5 15.5 7.8 10.7 11.2 22.0 8.7 12.1
MBD [%] 1.3 3.7 1.4 8.5 −2.5 −1.0 3.1 14.5 1.4 0.58
Remote Sens. 2016, 8, 768 14 of 19

Table 3. Overall (i.e., Planet DOY 318, 352 and 001) NDVI prediction results on a cover specific basis.
Results based on the Landsat-only (L8) and MODIS informed (M) processing streams are reported
in separate columns (i.e., L8 and L8 + M). The statistics are based on Planet data resampled to 30 m,
and are reported for the full set of data points (i.e., 99%). The number of pixel samples (n), and average
NDVI and standard deviation of each cover type are also given.

All Covers Alfalfa Grass Corn Bare/Sparse

L8 L8 + M L8 L8 + M L8 L8 + M L8 L8 + M L8 L8 + M
r2 0.970 0.941 0.953 0.941 0.935 0.874 0.959 0.892 0.750 0.636
MAD 0.014 0.021 0.043 0.07 0.048 0.066 0.033 0.066 0.009 0.013
MAD (%) 9.2 13.7 9.5 15.5 7.8 10.7 11.2 22.0 8.7 12.1
MBD (%) 1.3 3.7 1.4 8.5 −2.5 −1.0 3.1 14.5 1.4 0.58
NDVI 0.16 ± 0.17 0.45 ± 0.27 0.62 ± 0.23 0.30 ± 0.24 0.10 ± 0.04
n 401,902 38,516 11,148 9660 342,578

4. Discussion
Planet’s flock of Dove satellites has been collecting high spatial resolution imagery in broad red,
green, and blue spectral bands since March 2014. Daily image capture on a global scale is anticipated
with a full constellation of Dove satellites in the near future. While unique in terms of spatio-temporal
resolution, the limited spectral resolution and lack of atmospherically corrected or radiometrically
calibrated radiance data may pose a limitation for reliably characterizing the changing conditions
of land surface features generally, and vegetation canopies specifically. To address such potential
restrictions, Planet’s RGB imagery were translated into Landsat-8 consistent and atmospherically
corrected NDVI.
Whenever a NIR retrieval capability is lacking, the translation of the visible (i.e., RGB) imagery
into estimates of NDVI is not straightforward. To overcome the RGB data constraint, a data mining
approach based on establishing rule-based predictive regression models was shown to be effective
at predicting Landsat observed NDVI from information contained in the red and green bands of
the Planet imagery. NDVI predictability was best, i.e., characterized by a MAD of 0.014 (9.2%),
when a Landsat acquisition within a few days of the Planet acquisition was available to train the
rule-based models. The MAD increased to 0.021 (13.7%) when the Landsat NDVI training image was
further away (i.e., 11–16 days) from the Planet acquisition and MODIS NDVI images were used to
inform on the change in NDVI occurring between the overpasses.
The Planet NDVI predictions were aggregated to the Landsat 30 m resolution in order to facilitate
scale consistent pixel-wise comparisons. While efforts were made to ensure precise co-registration
of the imagery (via the pixel shifting approach described in Section 2.3.1), differences between the
Planet pixel aggregation and the Landsat-8 cubic convolution resampling technique are likely to
affect the pixel comparability, particularly at field boundaries. As a result, the prediction accuracy
was also assessed on the best performing majority (i.e., 80%) of the data, which reduced the overall
MAD to ~0.006 (~4.5%) and ~0.008 (~7%) (Table 2) for the Landsat-only and MODIS informed cases,
respectively. The retrieval accuracy levels may be impacted by the number of days between the Planet
acquisition and the Landsat NDVI image used to validate the Planet NDVI predictions. For example,
rapid changes in surface cover conditions between the Landsat and Planet overpasses on DOY 288
and 292, respectively, raised some questions on the validity of the Landsat NDVI image on DOY 288
for evaluating the Planet predictions on DOY 292. Overall, reported prediction accuracies were highest
for Planet imagery associated with a Landsat NDVI validation image acquired within two days of the
Planet imagery.
The high correlations (r2 > 0.9) reported between the Planet explanatory data (red, green, GRVI)
and the Landsat target NDVI, might seem surprising given the lack of NIR information. Dense green
vegetation canopies are characterized by a surface reflectance peak in the green wavelength domain
and a reflectance minimum in the red domain (absorption maximum by chlorophyll pigments),
Remote Sens. 2016, 8, 768 15 of 19

whereas the typical background soil reflectance signal is characterized by a steady increase over the
visible domain [15,50]. As a result, the reflectance contrast between green and red reflectances will
directly relate to changes in vegetation density. This makes a visible normalized difference index
such as the GRVI a particularly good proxy for NDVI, as has been observed in a number of previous
studies [17,28,29]. Confusion between the RGB and NDVI signals is most likely over non-vegetated
non-soil surfaces (e.g., roads, buildings, water), where the distinctive spectral signatures can be
mistaken for vegetation development without the presence of NIR information. Clearly, the link is not
universal and will depend on time-specific and spatially varying surface characteristics. The adopted
data mining approach is particularly useful at inferring these links on a repeatable scene-specific
basis via the definition of multiple rule-based regressions. The rule-based models will not be applied
beyond the given Planet scene, thereby avoiding traditional issues associated with the transferability
of regression-based predictions [51]. As such, changes in environmental or surface characteristics are
expected to have little effect on the algorithm performance, as the Cubist-based regressions will adapt
to any given condition.
The standard 16-day revisit cycle of Landsat-8 in addition to cloud cover issues reduces the
chance of near-coincident Landsat and Planet acquisitions over many regions. In those situations,
effective predictions of NDVI at the Planet acquisition DOY will rely on the ability to properly
describe variations in NDVI at the Landsat scale over the acquisition timespan. Day-specific (linearly
interpolated) MODIS NDVI images (250 m) computed from eight-day products of NIR and red
reflectances were found to be useful for this purpose. The 250 m MODIS resolution may not always
be sufficient for distinguishing the change in NDVI of individual land cover types in heterogeneous
landscapes. However, uncertainties associated with the mixed pixel issue were reduced by searching
for the “purest” (i.e., least mixed) MODIS pixel within the neighborhood of the given Landsat pixel
based on pixel-wise time-series correlations of coincident MODIS and Landsat NDVI images collected
within 45 days of the Planet acquisition (Section 2.3.2). Uncertainties are naturally associated with
describing the change in NDVI occurring between the DOYs of the optimal pixel observations in the
eight-day periods, which may vary from 1 to 16 days. A linear rate of change in NDVI between DOYs
(as adopted here) may not always be a good approximation for extended timespans. Supplementing
the data mining approach with eight-day Aqua MODIS products (i.e., MYD09Q1) or daily surface
reflectance products (i.e., MOD/MYD09GQ) during optimal acquisition conditions is likely to help
in establishing a higher frequency account of observed NDVI dynamics at the 250 m scale. Overall,
the approach relies on high quality retrievals for ensuring time and space consistent time-series records
of NDVI, and reliably describing relative variations in land cover specific NDVI occurring over the
acquisition timespan.
Reducing the time interval between the Planet acquisition DOY and the NDVI training image is
key to achieving good NDVI predictabilities. Unfortunately, extended timespans between overpasses
are likely to be an issue in regions with more frequent cloud coverage. While the framework has
been demonstrated using Landsat-8, it can be easily extended to high spatial resolution multi-spectral
sensors such as RapidEye (5 m) and Sentinel-2 (10 m) using their respective red and near-infrared
bands for providing NDVI target inputs to the Cubist regression modeling. Synergistic use of
multi-sensor NDVI data streams, atmospherically corrected and normalized to a common standard,
would significantly increase the chance of usable clear-sky NDVI training images close to the Planet
scene acquisition.
The developed approach uses generic techniques, adaptable scene-specific model regressions,
and automated processing streams to ensure transferability to any region of interest. The approach
involves the acquisition and atmospheric correction of the Landsat NDVI imagery, which is
accomplished in a fully automated manner using extendable atmospheric correction techniques [37,39].
Careful removal of atmospheric influences in the target NDVI images is particularly pertinent over
desert agricultural systems due to high aerosol loadings, a heterogeneous surface reflectance field,
and bright desert soils [37], and represent a critical step in the construction of high quality NDVI
Remote Sens. 2016, 8, 768 16 of 19

time-series from Planet RGB imagery. Contamination from space and time varying atmospheric effects,
inherent variations in the spectral and radiometric characteristics of individual Planet RGB sensors,
and differences in the orbit and acquisition characteristics of individual Doves can easily confound the
time-series signal relating to the actual change in surface cover conditions. A key objective of ongoing
work is to evaluate the robustness of the framework over other regions with contrasting environmental
and surface conditions.
The framework developed here exploits the high absolute radiometric accuracy and reflectance
error specification of the Landsat-8 push-broom sensor [8,9] to correct the Planet time-series imagery
into estimates of NDVI, with a comparable level of accuracy and consistency over time and space
domains. This is expected to significantly enhance the utility of Planet’s dense time-series of imagery
for near real-time monitoring of the state and condition of vegetation canopies. Capability will be
further enhanced with the addition of near-infrared sensors on-board the Dove platforms in the
near future. While this will enable Planet NDVI estimation directly from the imagery, the data
mining approach remains useful in producing Landsat consistent and atmospherically corrected NDVI
estimates at the scale and time of the Planet acquisition. Obviously, the approach outlined here may
also be applied directly to the Planet spectral band (RGB) data in order to retrieve Landsat consistent
surface reflectance data in associated visible wavelength bands, which could further advance the utility
of Planet’s imagery in agricultural applications.

5. Conclusions
The expanding flock of Planet’s Dove satellites surveys the globe with frequent RGB imagery
capture at high spatial resolution (~3 m). A daily frequency of global coverage is foreseen in the near
future with a full constellation of Dove satellites. In order to enhance the use of this unique capacity
for earth observation, the Planet RGB imagery was translated into Landsat-consistent estimates of
atmospherically corrected NDVI via a set of scene-specific rule-based predictive models. The adopted
data mining technique was shown to be effective at reproducing observed features and magnitudes
in Landsat NDVI from information contained in the red and green bands of the Planet imagery.
Predictabilities were optimal when a near-coincident Landsat NDVI image was available to train the
regression models: but accuracy levels remained high with larger time intervals between acquisitions
and when MODIS 250 m data were used to inform on the change in NDVI occurring between the
Landsat and Planet acquisitions.
The developed framework builds upon an automated processing stream and generic techniques,
which makes it extendable, scalable, and transferable to other regions of interest. The approach offers
significant potential for translating dense time-series of Planet RGB imagery, captured by a multitude
of Dove satellites during variable acquisition conditions, into consistent and atmospherically corrected
estimates of an important metric of vegetation density and health. Apart from being a useful metric
of vegetation health globally, this newly developed earth observation resource serves as a high
resolution tool for precision agriculture with the potential to significantly enhance the ability to
identify within-field variability in crop growing conditions on a timely and repeatable basis, which is
likely to help farmers optimize agricultural production and manage resources more sustainably.

Acknowledgments: Research reported in this publication was supported by the King Abdullah University of
Science and Technology (KAUST). We acknowledge Planet’s Ambassadors program for providing access to their
imagery archive as well as the outreach efforts of Planet’s Nuno Vilaça and Joseph Mascaro. We greatly appreciate
the logistical, equipment, and scientific support offered to our team by Jack King, Alan King, and employees of
the Tawdeehiya Farm in Al Kharj, Saudi Arabia, without whom this research would not have been possible.
Author Contributions: R.H. and M.F.M. conceived the project. R.H. developed the algorithm and designed
the framework, processed and analyzed the data, and prepared the figure material. R.H. wrote the paper with
assistance from M.F.M.
Conflicts of Interest: The authors declare no conflict of interest. The funding sponsors had no role in the design
of the study; in the collection, analyses, or interpretation of data; in the writing of the manuscript, and in the
decision to publish the results.
Remote Sens. 2016, 8, 768 17 of 19

References
1. Berni, J.; Zarco-Tejada, P.J.; Suarez, L.; Fereres, E. Thermal and Narrowband multispectral remote sensing for
vegetation monitoring from an unmanned aerial vehicle. IEEE Trans. Geosci. Remote Sens. 2009, 47, 722–738.
[CrossRef]
2. Esper, J.; Panetta, P.V.; Ryschkewitsch, M.; Wiscombe, W.; Neeck, S. NASA-GSFC Nano-satellite technology
for Earth Science missions. Acta Astronaut. 2000, 46, 287–296. [CrossRef]
3. Moran, M.S.; Inoue, Y.; Barnes, E.M. Opportunities and limitations for image-based remote sensing in
precision crop management. Remote Sens. Environ. 1997, 61, 319–346. [CrossRef]
4. Zhang, N.; Wang, M.; Wang, N. Precision agriculture—A worldwide overview. Comput. Electron. Agric. 2002,
36, 113–132. [CrossRef]
5. Robert, P.C. Precision agriculture: A challenge for crop nutrition management. Plant Soil 2002, 247, 143–149.
[CrossRef]
6. Gebbers, R.; Adamchuk, V.I. Precision agriculture and food security. Science 2010, 327, 828–831. [CrossRef]
[PubMed]
7. Roy, D.P.; Wulder, M.A.; Loveland, T.R.; Woodcock, C.E.; Allen, R.G.; Anderson, M.C.; Helder, D.; Irons, J.R.;
Johnson, D.M.; Kennedy, R.; et al. Landsat-8: Science and product vision for terrestrial global change research.
Remote Sens. Environ. 2014, 145, 154–172. [CrossRef]
8. Irons, J.R.; Dwyer, J.L.; Barsi, J.A. The next Landsat satellite: The Landsat data continuity mission.
Remote Sens. Environ. 2012, 122, 11–21. [CrossRef]
9. Czapla-Myers, J.; McCorkel, J.; Anderson, N.; Thome, K.; Biggar, S.; Helder, D.; Aaron, D.; Leigh, L.;
Mishra, N. The ground-based absolute radiometric calibration of Landsat 8 OLI. Remote Sens. 2015, 7, 600–626.
[CrossRef]
10. Drusch, M.; Del Bello, U.; Carlier, S.; Colin, O.; Fernandez, V.; Gascon, F.; Hoersch, B.; Isola, C.; Laberinti, P.;
Martimort, P.; et al. Sentinel-2: ESA’s optical high-resolution mission for GMES operational services.
Remote Sens. Environ. 2012, 120, 25–36. [CrossRef]
11. Yan, L.; Roy, D.P.; Zhang, H.; Li, J.; Huang, H. An Automated Approach for sub-pixel registration of Landsat-8
Operational Land Imager (OLI) and Sentinel-2 Multi Spectral Instrument (MSI) imagery. Remote Sens.
2016, 8, 520. [CrossRef]
12. Houborg, R.; Fisher, J.B.; Skidmore, A.K. Advances in remote sensing of vegetation function and traits.
Int. J. Appl. Earth Obs. Geoinf. 2015, 3, 1–6. [CrossRef]
13. Kaufman, Y.J.; Tanre, D. Strategy for direct and indirect methods for correcting the aerosol effect on remote
sensing: from AVHRR to EOS-MODIS. Remote Sens. Environ. 1996, 55, 65–79. [CrossRef]
14. Vermote, E.F.; El Saleous, N.Z.; Justice, C.O. Atmospheric correction of MODIS data in the visible to middle
infrared: First results. Remote Sens. Environ. 2002, 83, 97–111. [CrossRef]
15. Asner, G.P. Biophysical and Biochemical sources of variability in canopy reflectance. Remote Sens. Environ.
1998, 64, 234–253. [CrossRef]
16. Rouse, J.; Haas, R.; Schell, J.; Deering, D. Monitoring vegetation systems in the Great Plains with ERTS.
In Proceedings of the Third ERTS-1 Symposium NASA SP-351, Washington, DC, USA, 10–14 December 1973;
pp. 309–317.
17. Tucker, C.J. Red and photographic Infrared linear combinations for monitoring vegetation. Remote Sens. Environ.
1979, 8, 127–150. [CrossRef]
18. Tucker, C.J.; Townshend, J.R.G.; Goff, T.E. African land-cover classification using satellite data. Science 1985,
227, 369–375. [CrossRef] [PubMed]
19. Myneni, R.B.; Keeling, C.D.; Tucker, C.J.; Asra, G.; Nemani, R.R. Increased plant growth in the northern high
latitudes from 1981 to 1991. Nature 1997, 386, 698–702. [CrossRef]
20. Fensholt, R.; Rasmussen, K.; Nielsen, T.T.; Mbow, C. Evaluation of earth observation based long term
vegetation trends—Ntercomparing NDVI time series trend analysis consistency of Sahel from AVHRR
GIMMS, Terra MODIS and SPOT VGT data. Remote Sens. Environ. 2009, 113, 1886–1898. [CrossRef]
21. Olsen, J.L.; Stisen, S.; Proud, S.R.; Fensholt, R. Evaluating EO-based canopy water stress from seasonally
detrended NDVI and SIWSI with modeled evapotranspiration in the Senegal River Basin. Remote Sens. Environ.
2015, 159, 57–69. [CrossRef]
Remote Sens. 2016, 8, 768 18 of 19

22. Peters, A.J.; Walter-Shea, E.A.; Ji, L.; Viña, A.; Hayes, M.; Svoboda, M.D. Drought monitoring with
NDVI-Based standardized vegetation index. Photogramm. Eng. Remote Sens. 2002, 68, 71–75.
23. Karnieli, A.; Agam, N.; Pinker, R.T.; Anderson, M.; Imhoff, M.L.; Gutman, G.G.; Panov, N.; Goldberg, A.
Use of NDVI and land surface temperature for drought assessment: Merits and limitations. J. Clim. 2010,
23, 618–633. [CrossRef]
24. Running, S.W.; Nemani, R.R.; Heinsch, F.A.; Zhao, M.; Reeves, M.; Hashimoto, H. A continuous
satellite-derived measure of global terrestrial primary production. Bioscience 2004, 54, 547–560. [CrossRef]
25. Gitelson, A.A.; Peng, Y.; Huemmrich, K.F. Relationship between fraction of radiation absorbed by
photosynthesizing maize and soybean canopies and NDVI from remotely sensed data taken at close range
and from MODIS 250m resolution data. Remote Sens. Environ. 2014, 147, 108–120. [CrossRef]
26. Ruimy, A.; Saugier, B.; Dedieu, G. Methodology for the estimation of terrestrial net primary production from
remotely sensed data. J. Geophys. Res. 1994, 99, 5263–5283. [CrossRef]
27. Gitelson, A.A.; Kaufman, Y.J.; Stark, R.; Rundquist, D. Novel algorithms for remote estimation of vegetation
fraction. Remote Sens. Environ. 2002, 80, 76–87. [CrossRef]
28. Motohka, T.; Nasahara, K.N.; Oguma, H.; Tsuchida, S. Applicability of Green-Red Vegetation Index for
remote sensing of vegetation phenology. Remote Sens. 2010, 2, 2369–2387. [CrossRef]
29. Schneider, P.; Roberts, D.A.; Kyriakidis, P.C. A VARI-based relative greenness from MODIS data for
computing the Fire Potential Index. Remote Sens. Environ. 2008, 112, 1151–1167. [CrossRef]
30. Vermote, E.F.; Kotchenova, S. Atmospheric correction for the monitoring of land surfaces. J. Geophys.
Res. Atmos. 2008, 113, 1–12. [CrossRef]
31. Blackburn, G.A. Hyperspectral remote sensing of plant pigments. J. Exp. Bot. 2007, 58, 855–867. [CrossRef]
[PubMed]
32. Houborg, R.; Soegaard, H.; Boegh, E. Combining vegetation index and model inversion methods for
the extraction of key vegetation biophysical parameters using Terra and Aqua MODIS reflectance data.
Remote Sens. Environ. 2007, 106, 39–58. [CrossRef]
33. Thenkabail, P.S.; Lyon, J.G.; Huete, A. Hyperspectral Remote Sensing of Vegetation; CRC Press: Boca Raton,
FL, USA, 2011.
34. Haboudane, D. Hyperspectral vegetation indices and novel algorithms for predicting green LAI of crop
canopies: Modeling and validation in the context of precision agriculture. Remote Sens. Environ. 2004,
90, 337–352. [CrossRef]
35. Broge, N.; Leblanc, E. Comparing prediction power and stability of broadband and hyperspectral vegetation
indices for estimation of green leaf area index and canopy chlorophyll density. Remote Sens. Environ. 2000,
76, 156–172. [CrossRef]
36. Candiago, S.; Remondino, F.; De Giglio, M.; Dubbini, M.; Gattelli, M. Evaluating Multispectral images and
vegetation indices for precision farming applications from UAV images. Remote Sens. 2015, 7, 4026–4047.
[CrossRef]
37. Houborg, R.; McCabe, M.F. Adapting a regularized canopy reflectance model (REGFLEC) for the retrieval
challenges of dryland agricultural systems. Remote Sens. Environ. 2016, 186, 105–120. [CrossRef]
38. Landsat 8. Available online: http://landsat.usgs.gov/landsat8.php (accessed on 14 September 2016).
39. Houborg, R.; McCabe, M.; Cescatti, A.; Gao, F.; Schull, M.; Gitelson, A. Joint leaf chlorophyll content
and leaf area index retrieval from Landsat data using a regularized model inversion system (REGFLEC).
Remote Sens. Environ. 2015, 159, 203–221. [CrossRef]
40. Kotchenova, S.Y.; Vermote, E.F.; Matarrese, R.; Klemm, F.J. Validation of a vector version of the 6S radiative
transfer code for atmospheric correction of satellite data. Part I: Path radiance. Appl. Opt. 2006, 45, 6762–6774.
[CrossRef] [PubMed]
41. Vermote, E.F.; Tanre, D.; Deuze, J.L.; Herman, M.; Morcette, J.J. Second simulation of the satellite signal in
the solar spectrum, 6S: An overview. IEEE Trans. Geosci. Remote Sens. 1997, 35, 675–686. [CrossRef]
42. Hsu, N.C.; Tsay, S.C.; King, M.D.; Herman, J.R. Aerosol properties over bright-reflecting source regions.
IEEE Trans. Geosci. Remote Sens. 2004, 42, 557–569. [CrossRef]
43. Houborg, R.; McCabe, M.F.; Gao, F. A Spatio-Temporal Enhancement Method for medium resolution LAI
(STEM-LAI). Int. J. Appl. Earth Obs. Geoinf. 2016, 47, 15–29. [CrossRef]
44. Wang, P.; Gao, F.; Masek, J.G. Operational data fusion framework for building frequent Landsat-like imagery.
IEEE Trans. Geosci. Remote Sens. 2014, 52, 7353–7365. [CrossRef]
Remote Sens. 2016, 8, 768 19 of 19

45. Quinlan, R. Learning with continuous classes. In Proceedings of the 5th Australian Joint Conference on
Artificial Intelligence, Hobart, Australian, 16–18 November 1992; pp. 343–348.
46. Quinlan, R. Combining instance-based and model-based learning. In Proceedings of the Tenth International
Conference on Machine Learning, Amherst, MA, USA, 27–29 June 1993; pp. 236–243.
47. Moisen, G.G.; Freeman, E.A.; Blackard, J.A.; Frescino, T.S.; Zimmermann, N.E.; Edwards, T.C., Jr. Predicting
tree species presence and basal area in Utah: A comparison of stochastic gradient boosting, generalized
additive models, and tree-based methods. Ecol. Model. 2006, 199, 176–187. [CrossRef]
48. Gleason, C.J.; Im, J. Forest biomass estimation from airborne LiDAR data using machine learning approaches.
Remote Sens. Environ. 2012, 125, 80–91. [CrossRef]
49. Walton, J.T. Subpixel urban land cover estimation: Comparing cubist, random forests, and support vector
regression. Photogramm. Eng. Remote Sens. 2008, 74, 1213–1222. [CrossRef]
50. Price, J.C. On the information content of soil reflectance spectra. Remote Sens. Environ. 1990, 33, 113–121.
[CrossRef]
51. Atzberger, C.; Darvishzadeh, R.; Immitzer, M.; Schlerf, M.; Skidmore, A.; le Maire, G. Comparative analysis
of different retrieval methods for mapping grassland leaf area index using airborne imaging spectroscopy.
Int. J. Appl. Earth Obs. Geoinf. 2015, 43, 19–31. [CrossRef]

© 2016 by the authors; licensee MDPI, Basel, Switzerland. This article is an open access
article distributed under the terms and conditions of the Creative Commons Attribution
(CC-BY) license (http://creativecommons.org/licenses/by/4.0/).