Content uploaded by Dr. Tasneem Ahmed
Author content
All content in this area was uploaded by Dr. Tasneem Ahmed on Sep 30, 2022
Content may be subject to copyright.
XXX-X-XXXX-XXXX-X/XX/$XX.00 ©20XX IEEE
RANDOM FOREST BASED FLOOD MONITORING USING SENTINEL-1 IMAGES: A
CASE STUDY OF FLOOD PRONE REGIONS OF NORTH-EAST INDIA
Mohammed Siddique1,2
Advanced Computing Research Lab,
Department of Computer Application
1Integral University, Lucknow, India
studentp@student.iul.ac.in
Faculty of Information Technology,
2Majan University College, Muscat, Oman
muhammed.siddique@majancollege.edu.om
Tasneem Ahmed
Advanced Computing Research Lab,
Department of Computer Application
Integral University, Lucknow, India
tasneemfca@iulac.in
Mohd Shahid Husain
Department of Information Technology
CAS-Ibri University of Technology and
Applied Sciences, Muscat, Oman
mshahid.ibr@cas.edu.om
Abstract — The use of SAR satellite images will be very helpful
in flood monitoring as the acquisition of synthetic aperture
radar (SAR) images is possible day-night in all weather
conditions and are very sensitive to water bodies and the
changes in their behaviour. The usage of SAR (like Sentinal-1)
images is an added advantage in handling the rescue operations
and damage assessments based on images acquired before
flood, flood at peak, and after flood effects. This paper
discusses flood mapping and results in two different case
studies. This is covered in phase-1 with RGB composite images
of the cities of Gorakhpur and Ayodhya and phase-2 to analyse
the flood situation using an accuracy assessment of Basti city
based on the supervised classification method on SAR data. In
this paper, random forest classification (RF) technique has
been used to identify the flood prone areas by using Sentinel-1
satellite images and interpreted the changes detected for rescue
operations. Sentinel-1 images are classified as Crisis image and
Archive image, and further analysed to identify the flood prone
areas (water bodies due to flood), permanent water bodies,
urban (Built-up area), and vegetation.
Keywords – Sentinel 1, Supervised Classification, Image
Processing, Flood mapping, SAR images.
I. INTRODUCTION
The recent floods in the state of Uttar Pradesh made
news with several flood-related incidents. August 2020, saw
as many as 666 villages in 17 districts hit by floods in Uttar
Pradesh, India which includes the cities of Gorakhpur,
Ayodhya, and Basti. Hence, it is vital to implement an image
classification technique and identify the Flood Prone areas
by using Sentinel-1 satellite images to further derive
required precautionary actions [1]. The images acquired
after the floods are compared with the images during the
flood. These images will be processed in order to distinguish
the flooded areas and compare them with the permanent
water bodies at normal times. As followed in any disaster
mapping applications, the image acquired during flood
events is referred to as crisis image, and the images acquired
after the floods are referred to as archive image.
The Synthetic Aperture Radar (SAR) data for the cities
of Gorakhpur and Ayodhya from the state of Uttar Pradesh,
India were collected from the Copernicus repository. As
stated by Lee, the processed SAR image can be used to
measure the forest loss rate and the processed data help
calculate the loss rate using MATLAB [2]. This could be
achieved by Pre-processing, post-processing, color
manipulation, and cropping the image in the editing
programs. With high-spatial resolution Sentinel-1 SAR data,
an operational and automated approach was followed for the
systematic identification of surface water bodies in near
real-time (NRT). For the incessant monitoring of surface-
water bodies, a WebGIS platform was developed in which
all produced maps are made available along with the new
maps being uploaded in NRT [3].
The time-series based statistics driven normalization of
sentinel intensity observations built a Bayesian probability
function. This would decrease the intensity mainly due to
the specular reflection from the signal. Similarly, it will
increase the intensity due to double bounce cases. The prior
and likelihood probability of floods were further computed.
Upon comparison with this result with an optimal uniform
threshold approach in a prior and during SAR intensity pair
analysis, the results from the flood detections in terms of a
cut-off probability of 0.5 displaying improved performance.
The outcome of the mapping result revealed that the Sentinel
intensity increase is linked with a high percentage of the
urban flood which is a double-bounce effect [4].
The flooded regions due to flash floods could be detected
effectively by using the Interferometric Synthetic Aperture
Radar (InSAR) tools for flood modeling and also for the
surface deformation modeling. The SARPROZ tool in the
MATLAB programming language provides better results.
During Ramsar floods, the comparison of flood images
based on intensity before and after floods depicted that an
area of 0.205 km2 was affected that was based on spatial
resolution of SAR SLC IW images. The satellite sensors
play a vital role in providing information to monitor the
current phenomena and future predictions [5]. The flood
image classification based on pixels and objects was carried
out by following the SAR time-series approach. This would
ensure measuring the impact on the flood extent derivation.
There was an implementation of K-Means clustering
algorithm on this data. The hierarchical thresholding
approach in the classification process-chain based on image
pixels or segments investigated and identified the
significance of polarization and time-series features for the
temporary flooded vegetation (TFV) derivation
contributing to the results. Through augmenting the
temporary open water (TOW) with TFV areas, the time-
series approach allows the detection of the entire flood
extent [6]. The image classification techniques on SAR data
on areas covered by clouds provide clear distinctions
between flooded and high-moisture areas. It helps the
stakeholders in decision making for rescue and relief
operations during the disaster and damage assessment [7].
4883978-1-6654-2792-0/22/$31.00 ©2022 IEEE IGARSS 2022
IGARSS 2022 - 2022 IEEE International Geoscience and Remote Sensing Symposium | 978-1-6654-2792-0/22/$31.00 ©2022 IEEE | DOI: 10.1109/IGARSS46834.2022.9884483
Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY ROORKEE. Downloaded on September 29,2022 at 19:07:04 UTC from IEEE Xplore. Restrictions apply.
This paper is characterized with Section I covering
literature on similar work and Section II has two case studies
covered in two different segments. With the first, the RGB
composite images of crisis images with clear segregation of
water and soil categories are created. The second phase is
the implementation of the supervised image classification
techniques on the SAR data images. In Section IV, the
results are discussed with the detailed interpretation of
images and the accuracy of each class measured for multiple
flood prone regions. In the end, Section V deliberates the
conclusion covering the outcomes of these case studies and
future work recommended.
II. STUDY AREA AND SATELLITE DATA USED
A. Study Area: Gorakhpur, Ayodhya, and Basti
The unprecedented floods in different parts of the world are
one of the most common natural calamities that affect the
loss of life and property severely every year. Many parts of
India and specifically UP-East is vulnerable to floods.
Fig. 1. Extents of SAR data study area of North-East India [9]
Therefore, the cities of Gorakhpur (26.7606° N, 83.3732°
E), Ayodhya (26.7922° N, 82.1998° E) and Basti (26.8140°
N, 82.7630° E) from these regions have been considered as
study areas.
Fig. 2. Photo credits: People wade through waterlogged street in UP [14].
B. Satellite Data Used
The SAR data from Copernicus Open Access Hub
provides geographical coordinates with flood mapping from
space which can help in analysing the situation on the
ground more accurately [8]. The SAR sensors have their
own source of illumination and hence, the images captured
by them are not affected by the atmospheric conditions. It
also provides much stronger data to differentiate between
water bodies and land areas. The level-1 data mostly in
Ground Range Detected (GRD) products has focused SAR
data with well-defined resolution. They are in Full
resolution, High resolution, and Medium resolution format.
In this paper, Sentinel 1A images are taken which has VV
and VH bands in IW mode with 10 m spatial resolution
during the floods (i.e. Crisis images) and after floods (i.e.
Archive image) from the year 2020. The particulars of these
images along with the Acquisition-ID and dates for the
selected cities are provided in the below Table 1.
Table 1: Sentinel 1 (SAR) Data (2020) Acquisition details
Acquisition-ID
Date
Image
City
S1A_IW_GRDH_1SDV_20200821T123008_
20200821T123033_034005_03F250_FA1D
21-
Aug
Crisis
Gorakhpur
S1A_IW_GRDH_1SDV_20201219T123008_
20201219T123033_035755_042F37_2718
19-Dec
Archive
S1A_IW_GRDH_1SDV_20200802T123825_
20200802T123850_033728_03E8C0_481F
2-Aug
Crisis
Ayodhya
S1A_IW_GRDH_1SDV_20200910T002821_
20200910T002846_034289_03FC44_EB16
10-Sep
Archive
S1A_IW_GRDH_1SDV_20200805T002754_
20200805T002819_033764_03E9E5_716C
5-Aug
Crisis
Basti
S1A_IW_GRDH_1SDV_20201219T123008_
20201219T123033_035755_042F37_2718
19-Dec
Archive
III. METHODOLOGY
The SAR images of Gorakhpur and Ayodhya from the
state of Uttar Pradesh, India were collected from Copernicus
repository. Subsequently, image processing techniques are
implemented and the changes are detected in terms of water,
land, vegetation and the bare soil areas. Further the data
generated using flood mapping is analysed to confirm the
accuracy of identifying the flooded areas.
A. Case Study1: Gorakhpur and Ayodhya Cities
The images acquired are from 21st August, 2020 and 19th
December 2020 which are during the floods and after the
floods respectively [14]. The Fig. 3 and Fig. 4 shown below
is an archive which after the floods and crisis image which
is during the floods. The images are over Gorakhpur city
where the river Rapti flows and Sarayu river in Ayodhya that
can be seen under the bands with amplitude details. The
abstract metadata of both the products are same in terms of
location proximity and the acquisition characteristics.
Fig. 3. SAR images during Archive and Crisis: Gorakhpur [8]
Fig. 4. SAR images during Archive and Crisis: Ayodhya [8]
The crisis image shows the flooded areas are patches of
a low backscattering return. The dark areas are due to
specular reflection over the smooth water surfaces. Hence,
the signal gets reflected away from the sensor. Whereas the
surrounding land is much rough and the urban areas have
rugged terrain with lots of distortion due to the layover and
foreshortening effects. A subset of areas where the flooding
is likely to have taken place is extracted which is around the
river. To analyse the inner urban areas, multi looking is
applied on the subsets with value 3 for range looks though,
it might affect the overall resolution. Further, a calibration
is applied which is essential to compare two images with VV
polarization [10]. The Sigma naught is retained as default
which in a ground range is the ratio instant to receive
4884
Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY ROORKEE. Downloaded on September 29,2022 at 19:07:04 UTC from IEEE Xplore. Restrictions apply.
backscatter/unit area. It was observed in the color
manipulation process that many of the pixels had low
backscatter values and hence, there is a switch from digital
numbers to the physical quantity which is Sigma naught
backscatter. Hence, the pixels are converted from a linear
scale to decibel in a non-linear logarithm scale to have a
clear visualization. The histogram generated is easier to
manipulate for achieving a clear image that distinguished
between water and other areas in the city and is further
refined with the contrast stretch on respective histograms.
Fig 5: Histograms for Crisis and Archive images (Gorakhpur)
Fig 6: Histograms for Crisis and Archive images (Ayodhya)
To overlay images over the map together, they are
stacked into output extends using the product Geolocation
to ensure they are well registered. The bands of both crisis
and archive images are combined into a new image. These
images were simply combined to create an RGB composite
so as to differentiate between the flooded areas and the
permanent water bodies. The archive image is selected as a
red band for the red channel to highlight the flooded areas
as a high radar response. The crisis image is assigned with
green and blue bands and the flooded areas will have a low
backscatter return. Hence, flooded areas appear in red with
high response in the red channel and low response in the
green and blue channels. The RGB composite images of
Gorakhpur and Ayodhya are shown in Fig. 7.
Fig 7: RGB Composite-Gorakhpur (Left) and Ayodhya (Right)
After that, RGB composite images are overlayed on Google
Earth for more accurate flood analysis as shown in Fig. 8.
Over the permanent water bodies especially in the river
stream, there would be uniform dark return with low
backscatter return for both images before/after floods and
during the floods with low response. From Fig. 8, it is
observed that red color covers a large portion of the cities’
built-up areas.
Fig 8: RGB composite images overlayed at Google Earth to highlight the
Flooded areas of Gorakhpur (Left) and Ayodhya (Right).
B. Case Study 2: Flood mapping with Supervised
Classifier-Basti
The images acquired are from 05th August, 2020 and 12th
December, 2020 that are during the floods and after the
floods respectively. The classification of land cover is
conducted using Sentinel-1 on the SNAP application tool
[11]. The processes start by stacking an input file and
creating a subset on the study area. The scene start pixel
coordinates of X and Y are 11115 and 13338 and scene end
coordinates are 17784 and 16759 respectively.
(a) VV image of Archive
(b) VV image of Crisis
Fig 9: Subset Area of Basti
The next stage creates a calibration to Radar for
generating radiometric calibration with spackle filter. As
evident from the images shown in Fig. 9, there are patches
of a low backscattering return with the dark areas over the
smooth water surfaces. The surrounding soil land is much
rough with lots of distortion due to the layover and
foreshortening effects. The terrain correction is applied to
these images to correct the distortion. Further, the images
were combined to create an RGB composite in order to
distinguish between the flooded areas and the permanent
water bodies. The archive and crisis images based on the
RGB Composite of Dual Pol Ratio VV+VH and in decibel
were reviewed.
To categorize different land covers, vectors are created
that include Water, Urban, Vegetation, and Bare soil. The
appropriate segments are identified manually. The Random
Forest Classifier type of supervised classification was
implemented on the vector-based image with all Water,
Urban, Vegetation, and Bare soil segments. The cross-
validation outcome generated provided four types of classes
for the respective segments [12]. The comparative analysis
of each of these classes is carried out in terms of accuracy,
precision, correlation, and error rates, and the outcome of
the distribution of water class is recorded. It is evident that
the flooded areas constitute over 16% that spread over urban
populated areas and vegetation fields [13]. The changes in
water level are detected based on water images from
different periods of time.
4885
Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY ROORKEE. Downloaded on September 29,2022 at 19:07:04 UTC from IEEE Xplore. Restrictions apply.
The confidence level is found to be greater than or equal
to 0.5. By using the testing dataset of total samples of 9024,
the correct predictions are found to be 61.79%.
Fig. 10. Classified Crisis image - Basti
Fig. 10. Distribution of class – cross validation.
The assessment outcome from crisis classifier data
shows that the water class TruePositives values were
702.000 and FalsePositives were 28.0000. In the archive
classifier, the water class TruePositives values were 18.000
and FalsePositives were 0.0000. This depicts that the
flooded area with water bodies stretches beyond the areas of
permanent water bodies.
IV. RESULTS AND DISCUSSION
The SAR data processing for the flood prone cities such as
Gorakhpur, Ayodhya, and Basti was processed by using the
techniques such as Batch processing, Binarisation, Color
Manipulation, Band-Maths expressions, and created water
images. The conceptual design for carrying out the research
work directs the set of tasks that were applied in order to
achieve the desired outcome. The testing feature importance
score retrieved from the classification technique ensures that
each feature is perturbed three times and the correct
prediction percentage is averaged. The prediction function
(f(x)) for the RF is generated using:
Where, Cfull is the average of the complete data set and K is
the number of features. In case study 2, the distribution of
each of the classes in pixels is analysed and the percentage
of flooded water areas is identified as shown in Table 2. The
correct predictions were 61.85% which confirms that the
dataset matches the schema. The detailed review further will
support extended research on implementing the change
detection techniques and refine the process of monitoring
the flood prone areas.
Table 2. Distribution of Classes.
SAR
IMAGE
WATER
AREA
URBAN
AREA
VEGETATION
BARE
SOIL
ARCHIVE
6.4%
47.0%
44.9%
1.6%
CRISIS
16.8%
27.7%
27.7%
27.7%
V. CONCLUSION
The study based on the RF classification approach is
very eminent for Sentine1 imagery. The results from RGB
composite and the RF classification confirm that the region
with water class shows more concentration of flooded areas.
Yet, the wrong predictions were at 38.11% which can be
addressed by processing the image using unsupervised
classification and further assess the results. The future work
will be to continue to draw outcomes based on algorithms
comparing images taken during different periods of time. A
WebGIS platform setup would support the incessant
monitoring of surface-water bodies. Further, the
implementation of change detection techniques will refine
the outcomes and a time series analysis support in
identifying the flood pattern by using Sentinel-1 images.
REFERENCES
[1] Ahmed, T., 2021. Monitoring and Mapping of Flash Flood of
Patna City Using Sentinel-1 Images: A Case of India’s Most
Flood Prone State. Academia Letters.
[2] Lee, Jung-Hoon & Sri Sumantyo, Josaphat & Waqar, Mirza
& Kim, Jae-Hyun. 2020. Analysis of forest loss by Sentinel1
SARtimeseries,182.184.10.1109/ICTC49870.2020.9289607.
[3] Ovakoglou, G., Cherif, I., Alexandridis, T., Pantazi, X.,
Tamouridou, A., Moshou, D., Tseni, X., Raptis, I.,
Kalaitzopoulou, S. and Mourelatos, S., 2021. Automatic
detection of surface-water bodies from Sentinel-1 images for
effective mosquito larvae control. Journal of Applied Remote
Sensing, 15(01).
[4] Lin, Yun, Bhardwaj and Hill, 2019. Urban Flood Detection
with Sentinel-1 Multi-Temporal Synthetic Aperture Radar
(SAR) Observations in a Bayesian Framework: A Case Study
for Hurricane Matthew. Remote Sensing, 11(15), p.1778.
[5] Jamali, A. and Abdul Rahman, A., 2019. Flood Mapping
Using Synthetic Aperture Radar: A Case Study of Ramsar
Flash Flood. ISPRS - International Archives of the
Photogrammetry, Remote Sensing and Spatial Information
Sciences, XLII-4/W16, pp.291-295.
[6] Tsyganskaya, V., Martinis, S. and Marzahn, P., 2019. Flood
Monitoring in Vegetated Areas Using Multitemporal
Sentinel-1 Data: Impact of Time Series Features. Water,
11(9), p.1938.
[7] Borah, S., Sivasankar, T., Ramya, M. and Raju, P. 2018.
Flood inundation mapping and monitoring in Kaziranga
National Park, Assam using Sentinel-1 SAR data.
Environmental Monitoring and Assessment, 190(9).
[8] Scihub.copernicus.eu. 2022. Open Access Hub. [online]
Available at: <https://scihub.copernicus.eu/> [Accessed 16
January 2022].
[9] Team, M., 2022. India: Maps. [online] Maphill.com.
Available at: <http://www.maphill.com/india> [Accessed 7
May 2022].
[10] C.M. Bhatt & G.S. Rao, 2016, Ganga floods of 2010 in Uttar
Pradesh, north India: a perspective analysis using satellite
remote sensing data, Geomatics, Natural Hazards and Risk,
7:2, 747-763.
[11] Abazaj, F. and Hasko, G., 2020. Use of Sentinel-1 Data in
Flood Mapping in the Buna River Area. International Journal
of Environment and Climate Change, pp.147-156.
[12] Rahaman, S. and Shermin, N., 2021. Identifying Built-up
Area Expansion and Comparing Two Conventional Built-up
Area Extraction Method from LANDSAT Imagery: A Case
Study on Khulna City. Academia Letters.
[13] Brivio, P. A., Colombo, R., Maggi, M., & Tomasoni, R. 2002.
Integration of remote sensing and GIS for accurate mapping
of flooded areas. International Journal of Remote Sensing,
23(3), 429–441.
[14] NDTV.com. 2022. Over 700 villages in 15 UP districts
affected by floods. [online] Available at:
<https://www.ndtv.com/india-news/over-700-villages-in-15-
uttar-pradesh-districts-affected-by-floods-2280943>
[Accessed 11 January 2022].
0.000
0.500
1.000
1.500
0246810
Distribution of classes - Cross Validation
class 1.0: Water class 2.0: Urban
class 3.0: Vegetation class 4.0: Bare Soil
4886
Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY ROORKEE. Downloaded on September 29,2022 at 19:07:04 UTC from IEEE Xplore. Restrictions apply.