Water Across Synthetic Aperture Radar Data (WASARD): SAR Water Body Classification

for the Open Data Cube

1
Zachary Kreiser, 2Brian Killough, 1Syed R Rizvi

1
Analytical Mechanics Associates, Hampton, VA
2
NASA Langley Research Center, Hampton, VA

ABSTRACT labeling all pixels for which this band’s value is below a
certain threshold as containing water. The thresholding
The detection of inland water bodies from Synthetic method works since water tends to return a much lower
Aperture Radar (SAR) data provides a great advantage over backscatter value to the satellite than land [1]. However, this
water detection with optical data, since SAR imaging is not method can be flawed since estimating the proper threshold
impeded by cloud cover. Traditional methods of detecting is often imprecise, complicated, and labor intensive for the
water from SAR data involves using thresholding methods end user. Thresholding also tends to use data from only one
that can be labor intensive and imprecise. This paper SAR polarization, when a combination of polarizations can
describes Water Across Synthetic Aperture Radar Data provide insight into whether water is present. [2]
(WASARD): a method of water detection from SAR data In order to alleviate these problems, this paper presents
which automates and simplifies the thresholding process an application for the Open Data Cube to detect water from
using machine learning on training data created from SAR data using support vector machine (SVM)
Geoscience Australia’s WOFS algorithm. Of the machine classification.
learning models tested, the Linear Support Vector Machine
was determined to be optimal, with the option of training 2. PLATFORM
using solely the VH polarization or a combination of the VH
and VV polarizations. WASARD was able to identify water WASARD is an application for the Open Data Cube, a
in the target area with a correlation of 97% with WOFS. mechanism which provides a simple yet efficient means of
ingesting, storing, and retrieving remote sensing data. Data
Index Terms— Sentinel-1, Open Data Cube, can be ingested and made analysis ready according to
Earth Observations, Machine Learning, Water Detection whatever specifications the researcher chooses, and easily
resampled to artificially alter a scene’s resolution. Currently
1. INTRODUCTION WASARD supports water detection on scenes from ESA’s
Sentinel-1 and JAXA’s ALOS. When testing WASARD,
Water classification is an important function of Earth Sentinel-1 was most commonly used due to its relatively
imaging satellites, as accurate remote classification of land high spatial resolution and its rapid 6 day revisit cycle [5].
and water can assist in land use analysis, flood prediction, With minor alterations to the application's code, however, it
climate change research, as well as a variety of agricultural could support data from other satellites.
applications [2]. The ability to identify bodies of water
remotely via satellite is immensely cheaper than contracting 3. METHODOLOGY
surveys of the areas in question, meaning that an application
that can accurately use satellite data towards this function Using supervised classification, WASARD compares SAR
can make valuable information available to nations which data to a dataset pre-classified by WOFS in order to train an
would not be able to afford it otherwise. SVM classifier. This classifier is then used to detect water in
Highly reliable applications for the remote detection of other SAR scenes outside the training set. Accuracy was
water currently exist for use with optical satellite data such measured according to the following metrics:
as that provided by LANDSAT. One such application,  Precision: a measure of what percentage of the
Geoscience Australia’s Water Observations from Space points WASARD labels as water are truly water
(WOFS) has already been ported for use with the Open Data  Recall: a measure of what percentage of the total
Cube [6]. However, water detection using optical data from water cover WASARD was able to identify.
Landsat is constrained by its relatively long revisit cycle of  F1 Score: a harmonic average of the precision and
16 days [5], and water detection using any optical data is recall scores
constrained in that it lacks the ability to make accurate Both precision and recall are calculated at the end of the
classifications through cloud cover [2]. The alternative training phase, when the trained classifier is compared to a
solution which solves these problems is water detection testing dataset. Because the WOFS algorithm’s
using SAR data, which images the Earth using cloud- classifications are used as the truth values when training a
penetrating microwaves. WASARD classifier, when precision and recall are
Because of its advantages over optical data, much mentioned in this paper, they are always with respect to the
research has been done into water detection using SAR data. values produced by WOFS on a similar scene of Landsat
Traditionally, this has been done using the thresholding data, which themselves have a classification accuracy of
method, which involves picking a polarization band and 97% [6].

978-1-5386-7150-4/18/$31.00 ©2018 IEEE 437 IGARSS 2018

Visual representations of water identified by WASARD in
this paper were produced using the function wasard_plot(),
which is included in WASARD.

3.1 Algorithm Selection

Figure 2: Accuracy of classifiers trained using different
polarization bands. Precision and Recall were measured
The machine learning model used by WASARD is the
with respect to the values produced by WOFS.
Linear Support Vector Machine (SVM). This model uses a
supervised learning algorithm to develop a classifier,
Figure 2 demonstrates that using both the VV and VH
meaning it creates a vector which can be multiplied by the
bands trades slightly lower recall for significantly greater
vector formed by the relevant data bands to determine
precision when compared with the VH band alone, and that
whether a pixel in a SAR scene contains water. This
using the VV band alone is inferior in both metrics.
classifier is trained by comparing data points from selected
WASARD therefore defaults to using both the VV and VH
bands in a SAR scene to their respective labels, which in this
bands, and includes the option to use solely the VH band.
case are “water” or “not water” as given by the WOFS
The VV polarization’s lower precision compared to the VH
algorithm. The SVM was selected over the Random Forest
polarization is in contrast to results from previous research
model, which outperformed the SVM in training speed, but
and may merit further analysis [4].
had a greater classification time and lower accuracy, and the
Multilayer Perceptron Artificial Neural Network, which had
3.2 Training a Classifier
a slightly higher average accuracy than the SVM, but much
greater training and classification times.
The steps in training a classifier with WASARD are
1. Selecting two scenes (one SAR, one optical) with
the same spatial extents, and acquired close to
each other in time, with a preference that the
scenes are taken on the same day.
2. Using the WOFS algorithm to produce an array of
the detected water in the scene of optical data, to
be used as the labels during supervised learning
3. Data points from the selected bands from the SAR
acquisition are bundled together into an array with
the corresponding labels gathered from WOFS. A
random sample with an equal number of points
labeled “Water” and “Not Water” is selected to be
partitioned into a training and a testing dataset
4. Using Scikit-Learn’s LinearSVC object, the
training dataset is used to produce a classifier,
which is then tested against the testing dataset to
determine its precision and recall
Figure 1: Visual representation of the SVM Classifier. The result is a wasard_classifier object, which has the
Each white point represents a pixel in a SAR scene. following attributes:
1. f1, recall, and precision: 3 metrics used to
In Figure 1, the diagonal line separating pixels determine the classifier’s accuracy
determined to be water from those determined not to be 2. Coefficient: Vector which the SVM uses to make
water represents the actual classification vector produced by its predictions. The classifier detects water when
the SVM. It is worth noting that once the model has been the dot product of the coefficient and the vector
trained, classification of pixels is done in a similar manner formed by the SAR bands is positive
as in the thresholding method. This is especially true if only 3. Save(): allows a user to save a classifier to the disk
one band was used to train the model. in order to use it without retraining
4. wasard_classify(): Classifies an entire xarray of
3.1 Feature Selection SAR data using the SVM classifier
All of the above steps are performed automatically
Sentinel-1 collects data from two bands: the when the user creates a wasard_classifier object.
Vertical/Vertical polarization (VV) and the
Vertical/Horizontal polarization (VH). When 100 SVM 3.3 Classifying a Dataset
classifiers were created for each polarization individually,
and for the combination of the two, the following results Once the classifier has been created, it can be used to detect
were achieved: water in an xarray of SAR data using wasard_classify(). By
taking the dot product of the classifier’s coefficients and the
vector formed by the selected bands of SAR data, an array
of predictions is constructed. A classifier can effectively be
used on the same spatial extents as the ones where it was
trained, or on any area with a similar landscape. While

438
testing WASARD, it was observed that a classifier trained
on one lake in Vietnam detected water accurately across the
entire nation.

3.4 Noise Reduction

One major drawback of SAR data is that the intensity of the

returned signals from a SAR satellite’s radar pulses will vary
in frequency from pixel to pixel due to the waves falling out
of phase after hitting the Earth’s surface. This results in an
image which has pixels that vary in intensity across a
homogenous area, referred to as speckle [1]. This speckle
noise can potentially cause the classifier to mislabel points, Figure 4: WASARD classification (left) vs WOFS
and so reducing speckle is necessary to ensuring WASARD classification (right). Pixels identified as water by each
makes the most accurate classifications possible. Since algorithm are denoted in blue.
speckle shows up as points normally too small to be separate
bodies of water, WASARD reduces noise by scanning over Figure 4 demonstrates WASARD’s utility in identifying
the classified dataset with a moving window and removing water no matter the conditions. The Sentinel scene on the left
isolated pixels labeled as water. and the Landsat scene on the right have the same latitudinal
and longitudinal extents and were acquired on the same day,
5. RESULTS AND DISCUSSION but the region was covered by clouds at the time of the
Landsat acquisition. The WOFS data here is unusable since
WASARD was chiefly tested on bodies of water in Southern the clouds prevent accurate classification, but WASARD is
Vietnam, which is appropriate in large part due to the still able to perform. Only one clear Landsat scene is needed
region’s tropical climate and jungle covered landscape, to train a WASARD classifier, making WASARD
which contribute to the area being frequently covered by invaluable in areas frequently covered by clouds.
heavy clouds. Since this weather makes water detection with On the areas studied, WASARD was able to correctly
optical data less effective, it is an ideal region on which to classify points with an accuracy which is comparable or
use WASARD. superior to existing methods for SAR data [1]. WASARD’s
classifications had a precision of .960 and recall of .980 on
5.1 Accuracy of Classifications average, with an overall correlation with WOFS of 97%.
These numbers are in comparison to the labels produced by
WOFS on a Landsat scene with the same spatial extents and
as close temporally as possible. It is possible, therefore, that
these results are slightly lower than WASARD’s true
accuracy due to changes in water cover between when the
Landsat scenes and SAR scenes were acquired. Overall, it
would seem that WASARD offers a viable alternative to
classification with WOFS for when weather conditions
preclude WOFS’s use.

5.2 Applications
Figure 3: Water Detection by WASARD (Raw SAR image Figure 5 shows a composite of WASARD’s classifications
on left, WASARD’s predictions denoted in white on the in Buon Tua Sarh in Dak Nong, Vietnam. This composite
right hand side). was built from 17 scenes spanning 22 months. It is clear that
there is a large amount of variation in the presence of water
The comparison in Figure 3 demonstrates WASARD’s in this region, as there are large portions colored blue where
water detection on a reservoir in Southern Vietnam. This water was found 80-100% of the time, as well as significant
classifier has a precision score of .963, and recall of .983. portions colored yellow or orange where water was found
The Support Vector Classifier is represented by the 20-60% of the time. Closer research reveals that this
following equation: variation is due to the body of water being a reservoir whose
water level is controlled by a dam at its northernmost point.
SVC = (VH Coefficient) (VV Band Value) + (VH Coefficient) While the example of identifying a body of water with a dam
(VH Band Value) + Bias Constant, where SVC < 0 is water is somewhat trivial, it is a valid demonstration of how time
and SVC > 0 is non-water. series data might be used to analyze flooding patterns in a
For Figure 3: SVC = -45.899 * VH - 1.271 * VV + 1.007 region. WASARD makes running time series analysis on
SAR scenes easy with the included function
wasard_time_plot().

439
 from 100 classifiers trained back to back on the same area.
Standard deviations of .008 for precision and .004 for recall
were found. Therefore, trial and observation of multiple
classifiers is recommended to find the optimal one, which
can then be saved and reused. Included in WASARD is a
function get_best_classifier() which automates this process,
training a given number of classifiers and returning them to
the user in a list sortable by precision, accuracy, or f1 score.

6. CONCLUSION

WASARD is an effective application for use with the Open

Data Cube that addresses the problem of threshold
estimation, one of the greatest obstacles to detecting water
using SAR data. WASARD solves this problem by using the
WOFS algorithm to produce a dataset which can then be
used to train an effective classifier for SAR data, with no
intermediate steps required by the user. Identifying and
removing false positives caused by speckle further improves
Figure 5: Time series representation of water identified by WASARD’s predictions to the point that it can match or
WASARD from February 2015 to December 2016. 17 defeat existing methods in both its simplicity and accuracy.
scenes were analyzed to create this image. Demonstrations shown in this paper were all performed on
data from ESA’s Sentinel. Preliminary tests done on JAXA’s
5.3 User Friendliness ALOS were promising but constrained by lack of available
data. Despite the fact that it is limited by its decreased
In addition to its accuracy and its functionality despite cloud performance in desert landscapes, WASARD offers an
cover, WASARD’s value also stems from the ease with effective and easy to use SAR water detection application
which it can be used. Every step of generating an effective for the Open Data Cube in tropical environments such as
classifier is automated. WASARD automatically handles all South Vietnam.
of the following:
 Selecting a clear Landsat scene and a 7. REFERENCES
corresponding SAR scene on which to train the
SVM 1. S. Hong et al., “Water Area Extraction Using
 Adjusting the resolution of the SAR data to fit that RADARSAT SAT Imagery Combined with
of the Landsat data Landsat Imagery and Terrain Information,”
 Creating a training dataset, and fitting the Sensors, vol. 15(3), pp. 6652-6657, 2015.
classifier to the data. 2. A. Behnamian et al., “Semi-Automated Surface
All the user need do is feed in two datasets with Water Detection with Synthetic Aperture Radar
identical spatial extents, one of SAR data and one from Data: A Wetland Case Study,” Remote Sensing. 23
Landsat. Additionally, since WASARD operates in November 2017.
conjunction with the Open Data Cube, it allows the user to 3. R. Rahman, P. Thakur, “Detecting, mapping, and
analyze patterns of water cover without the use of a GIS. analyzing of flood water propagation using
This user-friendliness contrasts most current methods of synthetic aperture radar (SAR) satellite data and
detecting water from SAR, which involve the user creating GIS: A case study from the Kendrapara District of
histograms to determine thresholds [2] and using a GIS to Orissa State of India,” The Egyptian Journal of
run analysis on water cover patterns [3]. Remote Sensing and Space Science. 24 October
2017.
5.4 Tradeoffs 4. M.A. Clement et al., “Multi-temporal synthetic
aperture radar flood mapping using change
Despite its benefits, WASARD is constrained in a few areas. detection,” Journal of Flood Risk Management. 27
WASARD struggles to return precise measurements in flat APR 2017.
areas without vegetation such as deserts due to the way 5. L. Zeng et al., “Analyzing changes of the Poyang
smooth, flat terrain reflects microwaves in a similar fashion Lake water area using Sentinel-1 synthetic
to water. This same constraint applies to any smooth, flat aperture radar imagery,” International Journal of
surface such as roofs, as well as terrain such as mountains Remote Sensing, vol. 38(23), 13 August 2017.
that may cause the wave to be reflected away from the 6. N. Mueller et al., “Water observations from space:
satellite’s sensor. Incorporating a digital elevation model Mapping surface water from 25 years of Landsat
may be able to alleviate this problem [1]. imagery across Australia,” Remote Sensing of
Additionally, due to the random selection of points to be Environment, vol. 174, pp. 341-352, 1 March
used for training data, WASARD will create a slightly 2016.
different classifier each time it is trained on a given area.
This was tested by calculating the precision and recall scores

440