Analyzing the Land Cover Change
and Degradation in Sundarbans Mangrove
Forest Using Machine Learning and Remote
Sensing Technique
Ashikur Rahman Khan, Anika Khan, Shehzin Masud, and Rashedur M. Rahman(B)
Department of Electrical and Computer Engineering,
North South University, Dhaka, Bangladesh
{ashikur.khan01,anika.khan01,[Link],
[Link]}@[Link]
Abstract. The purpose of our research work is to understand the efficiency and
advantage of applying machine learning technique on remote sensing data col-
lected from one of the largest mangrove forests in the world, named Sundarbans.
Our study area was Sundarbans mangrove forest, and we have detected land cover
changes in this area. The images we have used were collected from Landsat 8 OLI,
ETM+, TM data. After pre-processing the images, we classified them applying
the Maximum Likelihood classifier. We got overall accuracy of 80%, 75%, and
77.1% and kappa efficiency 0.80, 0.62, and 0.69 for the years 2001, 2011, 2021
respectively. To determine the overall accuracy and kappa efficiency, we have used
confusion matrix. In the last 20 years, Sundarbans mangrove forest has declined by
0.2% due to human settlements, deforestation, natural calamity, increasing water
salinity etc.
Keywords: LULC · Classification · Sundarbans · Remote sensing · Landsat
image · ArcGIS
1 Introduction
Sundarbans is a collection of naturally made facultative halophytes, widely known as
the mangrove forest. These forests are found in hundred and twenty countries all over
the world and the maximum amount of South Asian beautiful greenwood is associated
in Bangladesh and India. The contiguous forest covers a total area of 10,000 km2 with
6000 km2 lying in Bangladesh and the rest 4000 km2 in India [1, 2] Located in the
Ganges-Brahmaputra delta, over 300 flora and 1760 fauna species are found in this World
Heritage site. It also attracts a huge number of tourists [3, 4]. An intricate network of
river channels covers the forest along with small islands and mudflats formed through
the deposition of sediments from the river system. Sundarbans, the largest mangrove
forest in the world, serves as a carbon sink, provides shelters from massive damage of
intense storms, flood destruction, cyclone barrier, atmospheric imbalance, and a good
source of income for about 2.5 million people.
© Springer Nature Switzerland AG 2021
I. Rojas et al. (Eds.): IWANN 2021, LNCS 12862, pp. 429–438, 2021.
[Link]
430 A. R. Khan et al.
The versatile forest has been decreasing for ages, due to overexploitation and defor-
estation by humans to create agricultural land, shrimp farming, and natural calamities
[5]. While the condition in Bangladesh is already dire, China has also lost 2/3rd of
its mangrove forest from Guangdong province [7]. Cyclones have the general capacity
of destroying ecosystems, and habitation; and therefore, Sundarbans has been facing
greater risk due to being on the front, and experiencing higher wind speeds for its geo-
graphical location which leads it being hit by cyclones every 7 to 12 years while smaller
disturbances occur at even lesser time intervals [3]. This frequency is problematic since
it takes around 10–15 years to reach pre-disturbance levels. Sidr was the most robust
cyclone to hit Sundarbans in 2007, and next in 2009, Indian part of Sundarbans was
hit by Aila. Depletion in mangroves has become a global concern, thus mapping and
monitoring the mangrove has become a vital step for the counties carrying these forests.
Conservation of Sundarbans is crucial since it is already an endangered ecosystem.
Climate change inevitably has severe repercussions on it [3]. It has been primarily
predicted that the coastal area around the Bay of Bengal will get immersed because
of the rise of sea level in the future. So, to better understand what to expect and what
not to, it is wise to study the past decades of historical changes in this tropical land.
Mangrove mapping, change detection, and biomass estimation, the conservation of the
forest demand accurate research results [4].
By remote sensing (RS) an object can be monitored and data regarding it can be
noted without having to make any physical contacts or field visits. It comes into play
due to the inability to carry out sufficient field investigation in Sundarbans, and the
complexity of the area combined with its remoteness, and prevalence of floods make
it somewhat inaccessible [6]. RS and satellite imagery has made huge advancements
through the decades starting from introduction of sensors, and continues to be the prime
source for studies as computational costs became more affordable, and the need for
accurate information keeps rising. Therefore, inaccessible parts of the Sundarbans can
also be monitored with reliability. Moreover, Landsat datasets can be used to retrieve
information for long period of time. Several different aspects such as mangrove extent,
canopy closure, LULC, species et cetera have been studied, and the changes in high
resolutions can be noticed before, during, and after an event [20].
This study aims to analyze the land use/land cover change (LULC) in total area of
Sundarbans in the last 20 years (2001 to 2021), using satellite imagery and remote sensing
data. Three cloud-free multi-temporal Landsat Images of 2001, 2011, and 2021 were
classified using Maximum Likelihood classification with the help of ArcGIS software
which monitored the changes in Mangrove Area, Water Body, Mudflat, Sandy Bench,
and Bare Land.
2 Related Works
In paper [4], the authors carried out a study on the Sundarbans to southeast coastal belt
of Noakhali Chittagong-Cox’s Bazar. 28 Landsat imagery from the years 1976, 1989,
2000 and 2015 were used. Landsat MSS, TM, ETM+, and L8 OLI classified images
were found to be 80%, 80%, 87%, and 97% accurate, respectively. Also, perceived
changes of Mangrove in Bangladesh were owed to the deforestation, farms built for
shrimp & salt production, coastal erosion and sedimentation. A decision-tree learning
Analyzing the Land Cover Change and Degradation in Sundarbans 431
method for integrating multi-temporal Landsat TM data and ancillary GIS data to identify
mangroves in the Pearl River Estuary has been used in this paper [7]. They used C5.0
for UNIX and its Windows counterpart See5 to construct the training set, decision tree
training and decision tree rules in the paper. The result shows that from 1988 to 1995
there has been a significant decrease but from 1995 to 2002 there has been an increase
in the area, which lead to 45.8 ha reduction. In this paper [11], the authors carried out a
study on Khulna and Satkhira region. They collected Landsat images from 1980, 1989,
2002 and 2009. Sensors from Landsat, Multi Spectral Scanner (MSS), Thematic Mapper
(TM) and Enhanced Thematic Plus (ETM+) were used. The findings of their research
show a decrease from 49% area in 1989 to 29% 2009 in fallow land and 31% to 29%
from 1989 to 2009 in mangrove, and increase from 5% in 1989 to 19% in 2009 in
homestead and 14% in 1989 to 24% in 2009 in water bodies. As Maximum Likelihood
Classification is a prominent classifier in LULC, it was used in paper [10, 13, 14, 17, 19]
to secure accuracy between 77.0% to 85.71%. Whereas, unsupervised and supervised
classification was used in paper [18] got 78.32% and 81.62% accuracy respectively.
The most dominant species reported in [14] is Avicennia sp. In [9], the authors made
the Indian part of Sundarbans Biosphere Reserve as their subject of study where they
presented a negative correlation between LST and NDVI. The authors in [8] conducted
a study on 30 years of Landsat data using CCDC. Results showed that despite the overall
stability of the mangrove forest, there was a 25% regressive change owing partially to
injurious changes caused by cyclone Sidr and to die- back of Heritiera fomes.
The authors in [12] used cloud-free Landsat 5 TM and Landsat 8 OLI images to
study the land cover change for 1998, 2008 and, 2018. ENVI software version 5.3
was used for Image enhancement and classification. The result shows that 80% of the
human habitat area has been increased from 1998 to 2018; also 116% of aquaculture has
been increased between 1998 to 2008, which are 654 km2 and 231 km2 , respectively.
MLP-Markov Chain-based modeling is used to show the future prediction of land cover
change. Land use/land cover change between 1975 and 2006 on south-western part of
Indian Sundarbans was presented in this paper [15]. The approximate loss is 0.42% in
mangrove coverage. The three (1975, 1990 and 2006) multitemporal cloud-free satellite
image (Landsat) was downloaded from United States Geological Service Glovis. The
authors of this [16] paper, used freely available Landsat TM of 1989, TM of 2014, and L8
OLI imaginary of 2019 to generate land cover change in the eastern Sundarbans using the
maximum likelihood algorithm (MaxLike). Accuracy is reported as 80%, 82.85% and
84.28% respectively. The paper [16] comes out with a predictable output that says 0.44%
vegetation cover has been decreased in the past 30 years (1989–2019). As many studies
were carried out in this particular region, we have tried to focus on current circumstances
only. We have selected last 20 years of data for this study to analyze the changes that
has been occurred using the Maximum Likelihood Classifier.
3 Methodology
3.1 Data Acquisition
Our region of interest (ROI) is Sundarbans Mangrove Forest that spread between
Bangladesh and India. Its total area consists of 9,200 sq km2 , and a significant part
of the forest fall in Bangladesh territory (Fig. 1).
432 A. R. Khan et al.
Fig. 1. Location of study area
Our study area lies between 21° 54’–22° 51’ latitude and 88° 44’–89° 89’ longitude.
Heritiera fomes, Excoecaria agallocha, Ceriops roxburghian and Sonneratia apetala are
some common trees in this area [4]. The height of mangrove varies from region to region,
and on average, it stands 0.9 m and 2.1 m above sea level [3]. Average temperature
ranges from 12–25° and average rainfall is 700 mm per year [3]. May to October are
rainy seasons, and 80% of rainfall occur at this time. Several factors are reasonable for
the growth and development of mangrove forest, such as saline water, sedimentation,
ocean tidal etc.
3.2 Image Collection
We have collected twenty-eight level-1 terrain (LIT) Landsat image from United States
Geological Survey (USGS) website. This image fell into WGS_1984_UTM_Zone_46N
coordinate system. We collected 2021, 2011, 2001 images gradually from Landsat 8
OLI, ETM+7, Landsat 5 TM. Table 1 shows detailed information about the multispectral
imaginary. Usually, the dry seasons are the perfect time to collect the satellite image.
So, our collected images vary from January to March. The cloud cover of the images we
have used in image Pre-processing are less than 5%.
Table 1. Satellite data
Serial Number Path row Acquisition Time Landsat Spacecraft Cloud Cover Pixel Size UTM zone
1 P137 R045 21/01/2001 Landsat 5 TM 0 30 46N
2 P138 R045 5/3/2011 Landsat 7 ETM 0 30 46N
3 P138 R045 4/2/2021 Landsat 8 OLI 0.03 30 46N
Analyzing the Land Cover Change and Degradation in Sundarbans 433
3.3 Image Pre-processing
Figure 2 depicts our working policy. At first, we have combined different band image
like near-infrared, red, green, blue, and made a composite image. This different band
combination helped us to identify different object later on. Then we used ENVI 5.3v
Software to perform radiometric calibration pre-processing. It improved the Landsat
data by making it sharper and more visible. This process is done by converting the DN
(Digital Number) value of the data to the top of atmosphere reflectance. Fast line-of-
sight atmospheric analysis of hypercubes also known as FLAASH is an atmospheric
correction that corrects visible wavelengths to near infrared regions up to 3 µm. So,
we have applied FLAASH in our Landsat images. Then we performed QUAC (Quick
Atmospheric Correction). It is also another atmospheric correction that works between
VNIR-SWIR range. Finally, we applied the Dark Object Subtraction (DOS) in our pre-
processing section to complete it. As the study area is large, so it covers two Landsat
scenes. Thus, we had to mosaic the two Landsat images and convert them into one image.
After that, we have masked the image according to our area of interest. Finally, after
masking, our data is ready for classification. As our data had significantly less cloud
cover, so we did not need to perform any cloud masking (Fig. 2).
Multispectral Image
Composite Band
Radiometric Calibration
FLAASH
QUAC
DOS
Mosaic Image
Masking
Classify Image
Fig. 2. Workflow of image pre-processing and classification
434 A. R. Khan et al.
3.4 Image Classification
We used ArcGIS 10.8 software for image selection and classification. Training samples
were collected very carefully from the multispectral image. We have created five different
classes to identify different objects. They are Mangrove, Water Body, Mudflat, Sandy
Bench, and Bare Land. Training data were collected in the form of polygons. As the study
area lies in the coastal region, the water color is different between the open ocean and
coastal regions due to sedimentation. So, we had to take data from different locations. On
average, we took 50 polygons from each class. However, we have taken more polygons
for mangrove and water classes. We used supervised classification for detecting our
land cover. Maximum Likelihood Classification (MLC) is a very promising algorithm
for classification that we have used by the assist of ArcGIS software. This algorithm
is based on two principles. One is, each class sample cells are being distributed in
multidimensional space, and another one uses the Bayes theorem for decision making.
The classifier classified the images of 2001, 2011, and 2021 years (Figs. 3, 4, 5).
Fig. 3. Image classification of 2001. Fig. 4. Image classification of 2011.
Analyzing the Land Cover Change and Degradation in Sundarbans 435
Fig. 5. Image classification of 2021.
3.5 Accuracy Assessment
We have taken 35 points for our validation. It was randomly taken for each different
classes. As ground survey was not possible, so we used Google Earth Pro software to
choose the validation data. For accuracy assessment, we used confusion matrix. With
using it we have found the user accuracy, producer accuracy, overall accuracy and kappa
coefficient by the following formulas (Table 2).
Total Number of Correctly Classified Pixel (Diagonal)
Overall accuracy = × 100
Total number of Reference Pixels
Total Number of Correctly Classified Pixel in each class
Producer accuracy = × 100%
Total number of Reference Pixels in each class (the column total)
Total Number of Correctly Classified Pixel in each class
User accuracy = × 100%
in each class (The row total)
Total number of Reference Pixels
(TS × TCS) − (ColumnTotal × RowTotal)
Kappa Efficiency (t) =
T S 2 − (ColumnTotal − RowTotal)
Table 2. Accuracy assessment table consist of user accuracy, producer accuracy, overall accuracy
and kappa efficiency.
Mangrove Water Body Mudflat Bare Land Sandy bench Accuracy
Landsat Time Sample Size PA(%) UA(%) PA(%) UA(%) PA(%) UA(%) PA(%) UA(%) PA(%) UA(%) OA(%) Kappa
TM 2001 35 84.60% 100% 90% 90% 85.70% 60% 50% 66.70% 0% 0% 80% 0.72
ETM 2011 75 67% 97% 81% 100% 83% 50% 100% 60% 100% 10% 75% 0.62
L8 OLI 2021 35 84.60% 91.74 80% 88.80% 57.10% 57.10% 75% 100% 100% 25% 77.10% 0.69
436 A. R. Khan et al.
4 Result and Discussion
To calculate our area, we have converted the classified raster image into vector data.
Then we calculated the area of the three different years. The findings are shown below
(Table 3):
Table 3. Landcover area of each class in our study.
2001 2001 2011 2011 2021 2021
Landcover Class
sq km % sq km % sq km %
Mangrove 5684.7 62.28% 5818.5 63.25% 5671.8 61.65%
Water Body 2407.5 26% 2425.8 26.37% 2626.4 28.55%
Bare Land 178.2 1.95% 127.6 1.39% 145.2 1.58%
Mudflat 704.5 7.72% 611.3 6.64% 558.1 6.07%
Sandy Bench 153.2 1.68% 216.5 2.35% 198.3 2.16%
Total area 9128.1 100.00% 9199.7 100.00% 9199.8 100.00%
We can see that the area of mangrove has increased by 134 sq km2 from 2001 to 2011
but again it deceased by 147 sq km2 from 2011 to 2021. If we compare it from 2001
to 2021, the result says, total mangrove forest has been deceased by 13 sq km2 or 0.2%
in last 20 years. If we increase the time period the deforestation will surely increase in
future (Fig. 6).
Fig. 6. Mangrove area changes in last 20 years
Global Mangrove Watch [21] is an online platform that provides remote sensing
data for mangroves. By using ALOS Palsar and Landsat data it creates a baseline map
Analyzing the Land Cover Change and Degradation in Sundarbans 437
of mangrove in all over the world. From there, we collected a vector dataset of 2010 to
compare with our classification data. Then we clipped the vector dataset to our study
area and calculated the total area. The Dataset shows that in our study area 5,608 sq
km2 mangrove forest exists in 2010. In our 2011 year classification, we have found that
the mangrove area is around 5818.5 sq km2 . So, our findings are close to the Dataset of
Global Mangrove Watch.
We come to know that, in the last 20 years, mangrove forest landcover has decreased
by 0.2%. It might not seem a massive loss at first glance but if the time scale increases,
then we will gradually observe immense deforestation. Last few years, our government
took many reasonable steps to control the loss. There are several reasons for forest
declines such as natural calamity, increase of water salinity, human disturbance, forest
fire etc. No massive cyclone like Sidr has hit mangrove forest recently, so the decline
rate is still low. Government should impose heavy restriction on entering the reserved
area of the forest. Besides, mangrove plantation program can also play a significant role
to flatten the decline rate of deforestation. We should be more vigilant to protect these
halophytes.
5 Limitations and Future Work
The shortcomings of our analysis are that, the Image color band of Open Sea and Coastal
area are very different due to sedimentation. As a result, ArcGIS 10.8 software could not
segregate them properly and the change detection for Mudflat came out very low. Also,
as our study does not include any information related to field visits, the actual situation
might affect our predicted accuracy. Our plan for further work is to build a model by
using previous year’s data that can predict land cover changes in Sundarbans.
6 Conclusion
In our study, we tried to focus on the changes that occurred in the Sundarbans mangrove
forest from 2001 to present days. We used high multispectral satellite imagery to conduct
the research. Remote sensing is an emerging field of research due to its strong advantage.
In our study area, some regions were hazardous to collect data and sample, but we have
easily got it by using remote sensing data. From our results, we can see a decline in
the mangrove forest. Various factors work for this reduction but we must try our best to
stop this deforestation in order to save this beautiful ecosystem. Many researchers have
found that mangrove forests are gradually decreasing. So, it is high time we should pay
attention to them.
Reference
1. Shimu, S.A., Aktar, M., Afjal, M.I., Nitu, A.M., Uddin, M.P., Mamun, M.A.: NDVI based
change detection in Sundarban mangrove forest using remote sensing data. In: 4th Inter-
national Conference on Electrical Information and Communication Technology (EICT)
(2019)
438 A. R. Khan et al.
2. Rahman, M.M., Ullah, M.R., Lan, M., Sumantyo, J.S., Kuze, H., Tateishi, R.: Comparison
of Landsat image classification methods for detecting mangrove forests in Sundarbans. Int.
J. Remote Sens. 34(4), 1041–1056 (2013)
3. Mandal, M.S.H., Hosaka, T.: Assessing cyclone disturbances (1988–2016) in the Sundarbans
mangrove forests using Landsat and Google Earth Engine. Nat. Hazards 102(1), 133–150
(2020). [Link]
4. Islam, M.M., Borgqvist, H., Kumar, L.: Monitoring mangrove forest landcover changes in
the coastline of Bangladesh from 1976 to 2015. Geocarto Int. 34(13), 1458–1476 (2019)
5. Mondal, S.H., Debnath, P.: Spatial and temporal changes of Sundarbans reserve forest in
Bangladesh. Environ. Nat. Res. J. 15(1), 51–61 (2017)
6. Rahman, M.M., Lagomasino, D., Lee, S., Fatoyinbo, T.: Improved assessment of man-
grove forests in Sundarbans east wildlife sanctuary using WorldView 2 and TanDEM-X high
resolution imagery. Remote Sens. Ecol. Conserv. 5(2), 136–149 (2019)
7. Kai Liu, X.L., Shi, X., Wang, S.: Monitoring mangrove forest changes using remote sensing
and GIS data with decision-tree learning. Wetlands 28(2), 336–346 (2008)
8. Awty-Carroll, K., Bunting, P., Hardy, A., Bell, G.: Using continuous change detection and
classification of Landsat data to investigate long-term mangrove dynamics in the Sundarbans
region. Remote Sens. 11(23), 2833 (2019)
9. Thakur, S., et al.: Assessment of changes in land use, land cover, and land surface temperature
in the mangrove forest of Sundarbans. Environ. Dev. Sustain. 23(2), 1917–1943 (2020)
10. Rahman, M., Begum, S.: Land cover change analysis around the Sundarbans mangrove forest
of Bangladesh using remote sensing and GIS application. JSF 9(1–2), 95–107 (2013)
11. Sardar, P., Samadder, S.R.: Understanding the dynamics of landscape of greater Sundarban
area using multi-layer perceptron Markov chain and landscape statistics approach. Ecol. Ind.
121, 106914 (2021)
12. Datta, D., Deb, S.: Analysis of coastal land use/land cover changes in the Indian Sunderbans
using remotely sensed data. Geo-spatial Inf. Sci. 15(4), 241–250 (2012)
13. Kumar, M., Mondal, I., Pham, Q.B.: Monitoring forest landcover changes in the Eastern
Sundarban of Bangladesh from 1989 to 2019. Acta Geophys. 69, 561–577 (2021). [Link]
org/10.1007/s11600-021-00551-3
14. Giri, S., et al.: A study on abundance and distribution of mangrove species in Indian Sundarban
using remote sensing technique. J. Coast. Conserv. 18(4), 359–367 (2014). [Link]
1007/s11852-014-0322-3
15. Kar, N.S., Bandyopadhyay, S.: Tropical storm Aila in Gosaba block of Indian Sundarban:
remote sensing based assessment of impact and recovery. Geogr. Rev. India 77(1), 40–54
(2015)
16. Ghosh, M.K., Kumar, L., Roy, C.: Mapping long-term changes in mangrove species
composition and distribution in the Sundarbans. Forests 7(12), 305 (2016)
17. Debnath, A.: Land use and land cover change detection of Gosaba Island of the Indian
Sundarban Region by using multitemporal satellite image. Int. J. Hum. Soc. Sci. 7(1), 209–217
(2018)
18. Salam, M.A., Ross, L.G., Beveridge, C.M.C.: The use of GIS and remote sensing techniques
to classify the Sundarbans mangrove vegetation. J. Agrofor. Environ. 1(1), 7–15 (2007)
19. Ramteke, I.K., et al.: Land Use/Land Cover Change Dynamics in Coastal Ecosystem of
Sundarban Delta, West Bengal-A Case Study of Bali Island (2017)
20. Prusty, B.A.K., Chandra, R., Azeez, P.A. (eds.): Wetland Science. Springer, New Delhi (2017).
[Link]
21. Global Watch. [Link]
