论文信息

Pallavi M / Shilpa N / P. Shanmugasundaram 等 4 人

Environmental Earth Sciences

论文解读

AI 辅助整理
方法
NDWI/MNDWI结合六种阈值法与U-Net-ResNet101深度学习,Gr
发现
Otsu与Li阈值最可靠,NDWI优于MNDWI,U-Net-ResNet101精度达99.0%且泛
机会点
可将该可解释深度学习水体提取框架迁移至农田水体和灌溉水体监测,结合多时相遥感做季节动态分析。

论文摘要

摘要 准确提取和监测城市水体对于可持续水资源管理和环境规划至关重要。本研究提出了一种稳健的地球观测框架,该框架集成了归一化差异水体指数(NDWI)、改进的归一化差异水体指数(MNDWI)、自适应阈值技术以及深度学习模型,利用Sentinel-2影像实现水体的自动制图与季节分析。研究评估了六种阈值方法,即Otsu、Li、Triangle、Minimum、Mean和Yen,并在QGIS中对提取的水体进行了处理。采用全卷积神经网络(FCNN)、U-Net、U-Net-ResNet50和U-Net-ResNet101进行了基于瓦片的分割。实验结果表明,Otsu和Li阈值法提供了最可靠的水体 delineation,而NDWI在分割精度上始终优于MNDWI。尽管FCNN和U-Net在班加罗尔湖泊上取得了优异性能,但其在印度地理多样性湖泊上的泛化能力有限。所提出的U-Net-ResNet101模型表现出更优的稳健性,在多个区域实现了99.0%的训练和测试准确率,以及近乎完美的精确率、召回率和F1分数。Grad-CAM可视化和全面的消融研究验证了所提出架构的有效性。季节预测与历史水体动态高度吻合,展示了该框架在快速城市化地区可扩展的城市水体监测、环境信息学和决策支持方面的潜力。

全文

Environmental Earth Sciences

ORIGINAL ARTICLE
Environmental Earth Sciences (2026) 85:471
https://doi.org/10.1007/s12665-026-13208-5
An explainable earth informatics framework for automated urban
water body mapping and seasonal monitoring using multi-temporal
sentinel-2 imagery
Pallavi M1 · Shilpa N2 · P . Shanmugasundaram3  · Srinivasarao Thota4
Received: 20 July 2026 / Accepted: 6 November 2026
© The Author(s) 2026
P. Shanmugasundaram
psserode@mtu.edu.et
Pallavi M
dr.pallavi_cs@sirmvit.edu
Shilpa N
shilpa.n@presidencyuniversity.in
Srinivasarao Thota
t_srinivasarao@av.amrita.edu
1 Department of Computer Science, Sir MVIT, Bengaluru,
Karnataka 562157, India
2 Department of Mathematics, Presidency University,
Ittgalpura, Rajanukunte, Bengaluru, Karnataka 560119, India
3 Department of Mathematics, College of Natural &
Computational Sciences, MizanTepi University, 21, Tepi
Bushira, Tepi, Ethiopia
4 Department of Mathematics, Amrita School of Engineering,
Amrita Vishwa Vidyapeetham, Amaravati,
Andhra Pradesh 522503, India
Abstract
Accurate extraction and monitoring of urban water bodies are essential for sustainable water resource management and
environmental planning. This study presents a robust Earth observation framework that integrates the Normalized Differ -
ence Water Index (NDWI), Modified Normalized Difference Water Index (MNDWI), adaptive thresholding techniques,
and deep learning models for automated water body mapping and seasonal analysis using Sentinel-2 imagery. Six thresh -
olding methods, namely Otsu, Li, Triangle, Minimum, Mean, and Yen, were evaluated, and the extracted water bodies
were processed in QGIS. Tile-based segmentation was performed using Fully Convolutional Neural Networks (FCNN),
U-Net, U-Net-ResNet50, and U-Net-ResNet101. Experimental results demonstrate that Otsu and Li thresholding provide
the most reliable water delineation, while NDWI consistently outperforms MNDWI in segmentation accuracy. Although
FCNN and U-Net achieved excellent performance for lakes in Bengaluru, their generalization across geographically
diverse Indian lakes was limited. The proposed U-Net-ResNet101 model exhibited superior robustness, achieving 99.0%
training and test accuracy with near-perfect precision, recall, and F1-score across multiple regions. Grad-CAM visualiza -
tions and comprehensive ablation studies validated the effectiveness of the proposed architecture. Seasonal predictions
closely matched historical water dynamics, demonstrating the framework’s potential for scalable urban water monitoring,
environmental informatics, and decision support in rapidly urbanizing regions.
Keywords Sentinel-2 imagery · Urban water body mapping · Deep learning · Remote sensing · Earth informatics
Introduction
Water covers nearly three-quarters of Earth’s surface, yet
only a minor fraction of this is directly usable and of good
quality for human consumption and ecosystems. In rap -
idly urbanising regions, increasing pressures from indus -
trial growth, land-use change and hydrological disruption
threaten water quality and availability. In particular, small
urban surface water bodies — ponds, lakes, tanks and other
depressions found in cities — whether natural or artificial,
play disproportionately important roles in the urban envi -
ronment. They support groundwater recharge, regulate sur-
face runoff, provide habitat for aquatic and semi-aquatic
species, act as local climate moderators (e.g., reducing
urban heat island effects) and serve recreation and cultural
functions. However, these urban water bodies face serious
threats: pollution (from industrial, domestic and storm-
water sources), encroachment, siltation, habitat degradation

1 3
471 Page 2 of 26
Environmental Earth Sciences (2026) 85:471
Table 1 Summary of representative studies on automated water body
and remote sensing image extraction
Reference Approach Key Findings Research
Gap/Scope
Wieland et al.
(2024)
CNN-based
S1S2-Water
dataset using
Sentinel-1 and
Sentinel-2
High segmen-
tation accuracy,
with IoU
values of 0.845
and 0.965,
respectively.
Improve
atmospheric-
condition han-
dling, flood
samples, and
land-cover
diversity.
Wang et al.
(2020)
Thresholding and
MSLWENet with
ResNet-101
98.53% accu-
racy; effective
for small lakes
and varying
classes.
Expand
datasets and
improve mor-
phological
processing.
Dong et al.
(2023)
TSLWEM with
MNDWI
Up to 99.93%
accuracy and
effective detec-
tion of small
tributaries.
Validate
across diverse
regions and
higher-resolu-
tion imagery.
Sathish (2023) Global, determin-
istic, and Otsu
thresholding with
Sentinel-2
Otsu and itera-
tive threshold-
ing showed
effective
surface-water
classification.
Integrate
additional
indices and
improve
MNDWI-
based
discrimination.
Li et al.
(2021c)
U-Net with atten-
tion and pyramid
modules using
SAR data
Approximately
95% accuracy
and robust
performance
across different
conditions.
Improve
extraction
in com-
plex urban
environments.
Liu et al.
(2023)
Ten water indices
with Otsu thresh-
olding using
Landsat
NWI and EWI
provided stable
water extrac-
tion over time.
Combine
indices with
machine
learning for
improved
accuracy.
Li et al.
(2021a)
Six machine-
learning algo-
rithms and three
thresholding
methods
Otsu performed
effectively with
low computa-
tional cost.
Improve
transferability
in complex
environments.
Lilay and
Taye (2023)
SVM, RF, and
LinkNet using
Sentinel-2
Deep learn-
ing provided
improved
land-cover
classification.
Use larger
datasets, addi-
tional spectral
bands, and
SAR data.
Pan et al.
(2020)
U-Net compared
with RF and
OBIA
U-Net achieved
improved
urban seg-
mentation
performance.
Improve
detailed
mapping
of complex
urban areas.
Mohanty et al.
(2020)
U-Net and Mask
R-CNN for
satellite image
segmentation
High detection
performance,
with AP up to
0.937.
Improve high-
resolution and
multispec-
tral image
processing.
and disease-transmission risks in high-density zones. Effec-
tive identification, mapping, classification and temporal
monitoring of these water bodies is therefore a critical com-
ponent of sustainable urban water management (Luo et al.
2023).
Traditionally, monitoring water bodies relied on labour-
intensive field campaigns: manual surveying, in-situ sam -
pling and laboratory analysis of water quality, along with
ground mapping of water-body extent and condition. While
such methods are indispensable, they are time-consuming,
costly and spatially limited — especially in dynamic urban
settings and for small, distributed water bodies. In recent
years, remote sensing and satellite imagery have become
powerful alternatives (or complements) to field methods,
enabling large-scale, frequent, and objective assessments of
surface water bodies and state changes (Sun and Li 2024).
Advances in satellite sensors, image processing, and com -
putational methods have made monitoring of urban hydro -
logical features far more efficient and accurate (Cai et al.
2024; Yan et al. 2023).
In remote sensing, there has been a shift towards index-
based methods for detecting and mapping water bodies.
These methods utilize the specific spectral properties of
water, which include the strong absorption of water in both
the near-infrared (NIR) and shortwave infrared (SWIR)
spectral bands, and the apparent differences of reflectance
between the visible spectral bands and these other surfaces,
to be able to separate water from vegetation, soil, built sur -
faces, shadows, clouds and other background surfaces (Li
et al. 2021b; Kang et al. 2021; Li et al. 2022). Among the
most widely used are the Normalised Difference Water
Index (NDWI) and the Modified NDWI (MNDWI). NDWI
typically uses green and NIR bands to highlight water fea -
tures, while MNDWI replaces NIR by a MIR (or SWIR)
band to better suppress built-up or soil background in urban
environments (and to more effectively exclude shadows)
(Liu et al. 2023.) These indices provide a computationally
efficient first step in delineating water bodies and are often
integrated into classification workflows for urban water
monitoring.
Beyond indices, contemporary studies emphasise the
integration of higher spatial and temporal resolution imag -
ery (e.g., from Sentinel-2, Landsat 8 and commercial satel -
lites) with machine learning and deep-learning algorithms to
improve accuracy (Pallavi et al. 2022a, b), handle small or
complex water bodies, reduce misclassification (e.g., shad -
ows mistaken as water), and enable change-detection over
time (Sun and Li 2024; Pallavi et al. 2022c). For instance,
very recent research has developed CNN-based semantic
segmentation models tailored for high-resolution urban

1 3
Page 3 of 26 471
Environmental Earth Sciences (2026) 85:471
Fig. 1 Visualization of the dataset description before and after augmentation

mapping in the urban area of Bangalore (India), leverag -
ing advanced spectral-analysis methods and recent method-
ological advances in remote sensing. The use of indices such
as NDWI and MNDWI (Chen et al. 2020) enhances clas -
sification accuracy by exploiting the strong absorption and
reflectance differences of water, while reducing background
noise from built surfaces, shadows and clouds (Francis et al.
2019). The NDWI, by leveraging NIR reflectance sensitiv -
ity, improves water delineation in mixed land-cover envi -
ronments; the MNDWI, by incorporating MIR (or SWIR)
bands, further suppresses non-water signals (especially
built-up and soil backgrounds) and is particularly effective
in dense urban settings. Together with ancillary process -
ing steps, these methods aim to provide precise, efficient
and scalable water-body mapping, thereby supporting
improved conservation, monitoring and management ini -
tiatives in rapidly urbanising regions. Recent studies have
underscored the critical role of accurate remote-sensing-
based monitoring of urban water resources and land-cover
dynamics in the context of rapid urbanization and climate
change. Sharma et al. demonstrated that LULC transitions
in Indian cities, particularly the expansion of built-up areas
at the expense of agricultural land and water bodies, lead to
measurable declines in ecosystem service values (Sharma
et al. 2024). Complementary multi-sensor approaches that
imagery, producing water-body extraction accuracies above
95% and effectively distinguishing small ponds, canals and
narrow streams from urban clutter and shadowing effects.
Also, reviews of the field show a growing trend in fusing
multisensor data, deploying machine learning (Random For-
est, SVM, GBT) and deep models (CNNs, GANs) for moni-
toring both water-body extent and water-quality parameters
such as chlorophyll-a, turbidity and CDOM, even in small
inland waters.
In the context of urban mapping, it is essential to empha-
size that small water bodies are often overlooked by coarser
sensors or coarser-resolution classification schemes, yet
they are ecologically and hydrologically important in urban
catchments. According to Ngamile et al. (2025), high spatial
resolution sensors like Sentinel-2 (10–20 m) with frequent
revisit intervals enable monitoring of small inland sur -
face waters and thus fill a key gap in urban water resource
assessment.
Moreover, recent research pointed out that while many
studies focus on optically active water-quality parameters,
a gap remains in remote sensing of non-optically active
parameters (such as total nitrogen, dissolved oxygen) and
robust atmospheric correction for inland and urban waters.
Given these developments, this work presents a compre-
hensive framework for automated water-body detection and

1 3
471 Page 4 of 26
Environmental Earth Sciences (2026) 85:471
Fig. 2 Methodology Overview for Water Body Mapping and Seasonal Analysis in Bangalore

Fig. 3 (a)U-Net model architecture, (b)FCNN Model Architecture

1 3
Page 5 of 26 471
Environmental Earth Sciences (2026) 85:471
Fig. 4 U-Net-ResNet101 Architecture

integrate optical, DEM and radar data have proven effec -
tive for delineating hydro-structural features controlling
groundwater occurrence in hard-rock terrains (Embaby
et al. 2024). Advances in Sentinel-2 time-series analysis
combined with hybrid deep-learning and cellular-automata
models have further enabled accurate multi-year land-cover
forecasting (Hussein et al. 2025). Beyond mapping, integra-
tive frameworks that couple water and food systems under
climate and socio-economic pressures highlight the need
for resilient management strategies (Thota et al. 2026a),
while graph-based and quantum-enhanced ecological mod -
els offer complementary tools for analysing stability in
aquatic–terrestrial systems (Thota et al. 2026b). Long-term
mathematical modelling studies reinforce the ecological
importance of surface water bodies (Thota et al. 2026c), and
the growing use of explainable deep learning techniques
supports transparent interpretation of environmental predic-
tions (Bikku et al. 2025). Collectively, these works provide
a strong foundation for scalable, explainable Earth infor -
matics frameworks such as the one proposed in this study.
In summary, the key contributions of this research
include:
● A methodological workflow tailored to urban water-
body mapping in an Indian metropolitan context, inte -
grating index-based detection (NDWI, MNDWI) with
modern remote-sensing practices.
● Application to small, distributed urban water bodies
(ponds, tanks, canals) which are often challenging to
detect and monitor.
● Emphasis on accuracy, automation and scalability in the
classification pipeline, aligned with recent advances in
sensor resolution and algorithmic capability.
● Provision of a foundation for ongoing monitoring,
change detection, and urban ecosystem-management
support in the face of rapid urbanisation.

1 3
471 Page 6 of 26
Environmental Earth Sciences (2026) 85:471
Fig. 5 Lakes Around Bangalore and Boundary Indicating Lakes Under BBMP Limits ( a) LULC predicted map using CNN, ( b) Lakes around
Bangalore

Table 2 Computational cost and training time of the proposed UNet-
ResNet101 model
Computational Measure UNet-ResNet101
GPU NVIDIA Tesla T4
Training epochs 100
Steps per epoch 34
Total training steps 3,400
Total training time 2062 s (34.37 min)
Average time per epoch 20.62 s
Approx. GPU-hours 0.573 GPU-hours
Table 3 Number of trainable parameters of the evaluated deep learn -
ing models
Model Number of Trainable Parameters
FCNN 31,031,745
UNet 409,281
UNet-ResNet50 20,985,473
UNet-ResNet101 62,907,585
Table 4 Hardware and software specifications used for the experiments
Specification Details
Platform Google Colab
GPU NVIDIA Tesla T4
GPU Memory 16 GB GDDR6
CPU Intel Xeon processor (Colab-dependent)
Framework TensorFlow/Keras
System RAM Typically 12–13 GB
Programming language Python
The formulas for these indices are as follows:
NDWI = Green − NIRGreen + NIR
NDWI is more effective in distinguishing water bodies from
land but it is affected by vegetation and soil moisture, lead -
ing to false positives.
NDWI( (using bands 8A and 11)) = Green − MIRGreen + MIR
Where Green: Band 8A (Green band) and MIR (Middle
Infrared): Band 11
NDWI( (using bands 8 and 12)) = Green − MIRGreen + MIR
Where Green: Band 8 (Green band) and MIR (Middle Infra-
red): Band 12
MNDWI = Green − MIRGreen + MIR
MNDWI enhances water detection in urban areas by using
MIR to suppress land signals, but it is sensitive to shadows

1 3
Page 7 of 26 471
Environmental Earth Sciences (2026) 85:471
Table 5 Table of hyperparameters
Hyperparameter Value / Description
Epochs 100
Batch Size 8
Optimizer Adam
Loss Function Binary_crossentropy
EarlyStopping Monitor val_loss
EarlyStopping Patience 5
EarlyStopping Mode min
Metrics Plotted Training/Validation
Loss, Training/Vali-
dation Accuracy
Table 6 Hyperparameters used for training the deep learning models
Hyperparameter FCNN UNet UNet-
ResNet50
UNet-
ResNet101
Batch size 8 8 8 8
Learning rate 0.001 0.001 0.001 0.001
Number of epochs 100 100 100 100
Optimizer Adam Adam Adam Adam
Loss function Binary
cross-
entropy
Binary
cross-
entropy
Binary
cross-
entropy
Binary
cross-
entropy
Learning-rate
scheduler
None None None None
Early stopping
criteria
Monitor:
val_loss
Monitor:
val_loss
Monitor:
val_loss
Monitor:
val_loss
Patience:
5 epochs
Patience:
5 epochs
Patience: 5
epochs
Patience: 5
epochs
Mode:
min
Mode:
min
Mode: min Mode: min
Restore
best
weights:
True
Restore
best
weights:
True
Restore best
weights:
True
Restore
best
weights:
True
and may misclassify dark surfaces as water. To improve the
accuracy of the classification of water bodies, many thresh-
olding methods such as Otsu, Triangle, Minimum, Mean,
Li, and Yen were used to determine the best thresholds for
extracting water bodies from NDWI and MNDWI images.
● Otsu’s Method: Selects the optimal threshold by mini -
mizing intra-class variance and maximizing inter-class
variance: It is effective for bimodal histograms but may
perform poorly with unimodal distributions and low-
contrast images.
● Triangle Method: Determines the threshold by forming
a triangle between the histogram peak and its baseline. It
is simple and effective for various images but sensitive
to noise and complex distributions.
● Minimum Method: Selects the threshold correspond -
ing to the minimum difference between the image his -
togram and the background histogram. It is straight -
forward but may be affected by noise and overlapping
intensity distributions. It is simple to implement but
sensitive to outliers and uneven illumination.
● Li’s Method: Determines a threshold by minimizing
the cross-entropy between image classes. It emphasizes
class separation but may be less effective for highly tex-
tured images.
● Yen’s Method: Uses entropy-based thresholding to
maximize class separability. It can be effective for com-
plex intensity distributions but may be sensitive to noise.
In particular, existing approaches predominantly rely on
fixed/manual threshold selection for NDWI/MNDWI
extraction, while the influence of threshold selection under
varying geographical conditions is insufficiently addressed.
Moreover, many existing studies validate their models
within the same or closely related geographical settings,
providing limited evidence of cross-geographical gener -
alizability. The proposed study addresses these limitations
through optimal thresholding of NDWI/MNDWI indices,
evaluation across geographically distinct regions, and
effective utilization of available computational resources.
The latter is particularly relevant for practical deployment
where access to high-end computational infrastructure may
be limited.
Literature review
Urban planning, resource management, and environmen -
tal monitoring all depend on the extraction of water bod -
ies. Effective mapping of surface water bodies is crucial in
areas that are quickly urbanizing in order to solve issues
like urban flooding and the distribution of water resources.
The methods used in automated water body extraction are
examined in this review Table 1, with a focus on the appli -
cation of deep learning techniques in conjunction with the
Normalized Difference Water Index (NDWI) and Modified
NDWI (MNDWI).
Methodology
This dataset visualization (Fig. 1), methodology (Fig. 2),
U-Net Architecture, FCNN Architecture (Fig. 3), U-Net-
ResNet101 Architecture (Fig. 4), LULC Map (Fig. 5)
dataset description in Section “Dataset description" and per-
formance metrics used in Sec. “Performance Metrics” are
all included in this section.
The computational cost, trainable parameters, and hard -
ware and software specifications are presented in Tables 2, 3,
and 4, respectively.

1 3
471 Page 8 of 26
Environmental Earth Sciences (2026) 85:471
Fig. 6 Raster Images: (a) Madiwala 843 band, (b) Ulsoor 843 band, (c) Madiwala 1143 band, and (d) Ulsoor 1143 band

Fig. 7 NDWI image of Madiwala Lake for different thresholding techniques

Fig. 8 NDWI image of Ulsoor Lake for different thresholding techniques

Dataset description
This study investigated a unique method combining deep
learning networks with an encoder-decoder architecture
for the accurate extraction of water bodies from remote
sensing photos. In order to identify possible water bod -
ies inside the Area of Interest (AOI), the methods first
obtained raster images from GoogleEarth Engine and
uploaded them to QGIS to compute the Normalized
Difference Water Index (NDWI) or Modified NDWI
(MNDWI).Numerous band combinations, such as Near-
Infrared (NIR) and Green, bands 8A and 11, and bands
8 and 12, were used to do NDWI calculations.The best
threshold values for binary segmentation were found using
a variety of thresholding approaches, including Otsu, Li,
Triangle, Minimum, Isodata, Mean, and Yen.The relevant
water body areas were computed using these threshold
values, and the results were verified using real data from

1 3
Page 9 of 26 471
Environmental Earth Sciences (2026) 85:471
Fig. 9 MNDWI image of Madiwala Lake for different thresholding techniques

Fig. 10 MNDWI image of Ulsoor Lake for different thresholding techniques

the Karnataka State Remote Sensing Applications Centre
(KSRSAC), Bengaluru-560097.
As input for segmentation models, tiles were produced
from both the original raster (raw) and the mask (binary
image) made using the chosen ideal threshold for the com -
plete city of Bangalore. The methodology incorporated a
two-step process: (1) integration of a state-of-the-art image
classification and segmentation model within the encoder-
decoder framework and (2) the use of deformable convo -
lution to adaptively modify the receptive field, enabling
accurate extraction of water body features. Deep learning
models, specifically Fully Convolutional Neural Networks
(FCNN), U-Net, U-Net_ResNet and U-Net_ResNet, were
trained with these image tiles to improve segmentation
accuracy.
The training images have a dimension of 3930 × 2958
pixels, from which 324 image tiles were generated. Before
augmentation, the dataset consisted of 324 tiles, which was
increased to 1296 tiles after augmentation. The test images
have a dimension of 1614 × 1446 pixels, from which 42
image tiles were generated for testing.
The models were evaluated on reserved test data to
assess their predictive performance and generalizability.
Ultimately, the trained models were used to predict water
masks for raster images, supporting the analysis and map -
ping of water bodies across India
In this paper, we will describe the design of a hybrid
encoder decoder architecture using a pre-trained ResNet101
as the encoder and U-Net-style architectures as a decoder
to enable the creation of segmented images at increased
depths and higher resolutions. In addition, this model will
take advantage of the representational ability of deep resid-
ual networks with their hierarchical feature extraction while
retaining the forward-backward ReLU characteristics of the
U-Net model that allows for spatial representation through-
out training.
Encoder: feature extraction using ResNet101
The encoder is comprised of a ResNet101 architecture that has
been pre-trained on ImageNet and all fully connected classifi-
cation layers have been omitted. The ResNet101 architecture
produces a deep hierarchical pipeline of feature extraction
and feature activation, progressively downsampling spatial
resolution with increasing levels of semantic information con-
tained within feature activations as defined by network stage

1 3
471 Page 10 of 26
Environmental Earth Sciences (2026) 85:471
Table 7 Water Area Analysis for Madiwala Lake
Lake Name Duration Thresholding
Technique
NDWI MNDWI NDWI 8a and 11 NDWI 8 and 12
Threshold
Value
Water
Area
(km2)
Threshold
Value
Water
Area
(km2)
Threshold
Value
Water
Area
(km2)
Threshold
Value
Water
Area
(km2)
Madiwala July-Oct Otsu -0.0962 3.32 0.0390 0.72 0.0961 1.29 0.1357 1.78
Li -0.1158 3.89 0.0199 0.74 0.0848 1.38 0.1191 1.96
Minimum 0.0023 0.67 0.0537 0.71 0.1128 1.18 0.1567 1.57
Triangle -0.1771 4.89 -0.0450 0.85 0.1295 1.08 0.1778 1.40
Isodata -0.0997 3.43 0.0390 0.72 0.0961 1.29 0.1357 1.78
Mean -0.0834 2.85 -0.0327 0.82 0.0901 1.34 0.1169 1.99
Yen -0.1982 5.07 -0.2087 3.07 0.1833 0.25 0.2386 0.46
Nov-Feb Otsu 0.0072 0.69 0.0733 0.71 0.0814 1.14 0.1220 1.56
Li -0.0699 3.25 0.0572 0.72 0.0745 1.19 0.1077 1.73
Minimum -0.3698 4.93 -0.2695 3.08 0.1036 1.02 0.1604 1.23
Triangle -0.1197 4.31 0.2119 0.56 0.1239 0.91 0.1732 1.14
Isodata -0.0591 2.87 0.0733 0.71 0.0796 1.16 0.1220 1.56
Mean -0.0497 2.44 -0.0021 0.78 0.0697 1.23 0.0989 1.87
Yen -0.1822 4.74 -0.2038 3.07 0.1645 0.11 0.2137 0.34
Mar-June Otsu -0.0144 0.69 0.0328 0.70 0.0618 1.13 0.0973 1.62
Li -0.0982 3.02 0.0081 0.71 0.0543 1.19 0.0832 1.81
Minimum -0.3557 4.91 0.0376 0.69 0.0802 1.01 0.1306 1.27
Triangle -0.1777 4.44 -0.0728 0.78 0.1023 0.89 0.1021 1.57
Isodata -0.0144 0.69 0.0328 0.70 0.0618 1.13 0.0973 1.62
Mean -0.0814 2.38 -0.0573 0.76 0.0501 1.23 0.0769 1.90
Yen -0.1652 4.33 -0.2072 2.96 0.2181 0.01 0.2020 0.29
Table 8 Water Area Analysis for Ulsoor Lake
Lake Name Duration Thresholding
Technique
NDWI MNDWI NDWI 8a and 11 NDWI 8 and 12
Threshold
Value
Water
Area
(km2)
Threshold
Value
Water
Area
(km2)
Threshold
Value
Water
Area
(km2)
Threshold
Value
Water
Area
(km2)
Ulsoor July-Oct Otsu -0.0488 0.37 0.0453 0.35 0.1023 0.98 0.1654 1.06
Li -0.0918 0.58 0.0167 0.37 0.0875 1.07 0.1430 1.14
Minimum 0.0105 0.31 0.1162 0.30 0.0541 1.27 0.0948 1.31
Triangle -0.0503 0.37 -0.0344 0.42 0.1817 0.42 0.2184 0.82
Isodata -0.0503 0.37 0.0453 0.35 0.1009 0.98 0.1654 1.06
Mean -0.1058 0.68 -0.0496 0.44 0.1104 0.92 0.1886 0.96
Yen -0.1096 0.71 -0.0653 0.50 0.0866 1.08 0.1601 1.08
Nov-Feb Otsu -0.0183 0.36 0.0867 0.36 0.0915 0.74 0.1505 0.88
Li -0.0709 0.82 0.0596 0.38 0.0773 0.85 0.1289 1.00
Minimum -0.0951 1.06 0.0797 0.37 0.0447 1.13 0.1924 0.66
Triangle 0.0453 0.30 -0.0118 0.44 0.1718 0.30 0.2358 0.43
Isodata -0.0183 0.36 0.0867 0.36 0.0903 0.74 0.1490 0.88
Mean -0.0748 0.86 -0.0125 0.44 0.0824 0.81 0.1536 0.86
Yen -0.0951 1.06 -0.1549 1.77 0.0663 0.95 0.1230 1.03
Mar-June Otsu -0.0479 0.35 0.0594 0.36 0.0958 0.85 0.1562 0.92
Li -0.1319 0.95 0.0174 0.38 0.0771 0.96 0.1298 1.03
Minimum -0.1587 1.12 0.0417 0.37 0.1373 0.62 0.1789 0.82
Triangle 0.0610 0.28 -0.0649 0.43 0.1673 0.48 0.2297 0.62
Isodata -0.0479 0.35 0.0594 0.36 0.0958 0.85 0.1562 0.92
Mean -0.1251 0.90 -0.0625 0.43 0.0941 0.86 0.1653 0.88
Yen -0.1549 1.10 -0.0978 0.52 0.0743 0.98 0.2618 0.23

1 3
Page 11 of 26 471
Environmental Earth Sciences (2026) 85:471
Table 9 The performance metrics of various deep learning models for the Modified Normalized Difference Water Index (MNDWI) and the Nor -
malized Difference Water Index (NDWI)
Index Model Precision Recall F1-score IoU Train Acc Test Acc
MNDWI FCNN 1.00 0.98 0.99 0.9826 0.9978 0.9932
U-Net 1.00 0.98 0.99 0.9800 0.9799 0.9930
U-Net_ResNet50 1.00 0.98 0.99 0.4500 0.9900 0.9807
U-Net_ResNet101 1.00 0.99 1.00 0.6900 0.9978 0.9900
NDWI FCNN 1.00 1.00 1.00 1.0000 0.9980 0.9993
U-Net 1.00 1.00 1.00 1.0000 0.9982 0.9989
U-Net_ResNet50 1.00 1.00 1.00 0.8640 0.9900 0.9900
U-Net_ResNet101 1.00 1.00 1.00 0.8460 0.9900 0.9900
Fig. 11 Performance Evaluation of FCNN Model for the index NDWI: (a) Confusion Matrix, (b) Training Accuracy, (c) Training Loss, (d) Band
1, (e) Band 2, (f) Band 3

selections. Using the selected feature maps of ResNet101 for
skip connections throughout the U-Net decoder (described
below)
● 𝑆𝑆1: conv1_relu(128 × 128) — captures low-level
texture and edge features.
● 𝑆𝑆2: conv2_block3_out(64 × 64) — encodes mid-
level structural patterns.
● 𝑆𝑆3: conv3_block4_out(32 × 32) — contains inter -
mediate semantic features.
● 𝑆𝑆4: conv4_block23_out(16 × 16) — represents
high-level object features.

1 3
471 Page 12 of 26
Environmental Earth Sciences (2026) 85:471
Fig. 12 Performance Evaluation of U-Net Model for the index NDWI: (a) Confusion Matrix, (b) Training Accuracy, (c) Training Loss, (d) Band
1, (e) Band 2, (f) Band 3

The deepest representation, obtained from conv5_
block3_out(8 × 8), is denoted as the bottleneck feature
tensor B, as it encodes the highest-level abstract semantics.
𝑋𝑋 ∈ ℝ256×256×3 𝑓𝑓encoder− −−−−− →{𝑆𝑆1,𝑆𝑆 2,𝑆𝑆 3,𝑆𝑆 4,𝐵𝐵 },
which defines a multiscale feature transformation through
the encoder.
{𝑆𝑆1,𝑆𝑆 2,𝑆𝑆 3,𝑆𝑆 4,𝐵𝐵 } =𝑓𝑓 ResNet101(𝑋𝑋)
Decoder: multi-scale reconstruction
At the decoder stage, high-resolution predictions are cre -
ated by progressively up-sampling the bottleneck repre -
sentation by integrating encoder features through skip
connections. One of the functions of the decoder blocks
is to:
(a) Upsampling The nearest neighbour method will
always up-sample (increase) by double the amount of space
the resolution takes up.
𝑋𝑋up( 𝑖𝑖 , 𝑖𝑖 )=𝑋𝑋(⌊2𝑖𝑖⌋ , ⌊2𝑖𝑖⌋)
(b) Skip Fusion The corresponding encoder feature map
poses a concatenation (concatenation operation) with the
corresponding encoder feature maps (Sk) to create a larger
concatenated tensor.
𝑋𝑋cat = Concat(𝑋𝑋up,𝑆𝑆 𝑘𝑘 )
When these tensors are concatenated, they allow you to
regain the spatial context of the encoder feature maps that
were lost when the encoder feature maps were down-sam -
pled. One will be able to localize precisely as a result of
reconstructing the concatenated tensor.

1 3
Page 13 of 26 471
Environmental Earth Sciences (2026) 85:471
Fig. 13 Performance Evaluation of U-Net-ResNet50 Model for the index NDWI: (a) Confusion Matrix, (b) Training Accuracy, (c) Training Loss,
(d) Band 1, (e) Band 2, (f) Band 3

(c) Feature Refinement (𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐 𝑏𝑏 𝑙𝑙𝑐𝑐𝑐𝑐𝑙𝑙) Once the concat -
enated tensor is created, it passes through a convolutional
refinement operation consisting of two sequences of 3 × 3
convolutions followed by batch normalization and ReLU
activations, improving non-linearity while decreasing inter-
nal covariate shift by maintaining spatial resolution through
same-padding.
𝑋𝑋1 = ReLU(BN(𝑊𝑊1 ∗𝑋𝑋 cat)) ,
𝑋𝑋2 = ReLU(BN(𝑊𝑊2 ∗𝑋𝑋 1)) .
Finally, the decoder will construct feature maps for each
resolution level: 16×16 → 32×32 → 64×64 → 128×128 →
256×256.
3. Output Layer: Pixel-Wise Prediction At the end 1 ×1
convolution produces probability score of each class per
pixel:
𝑍𝑍=𝑊𝑊 out ∗𝑋𝑋 dec, ̂𝑦𝑦(𝑖𝑖, 𝑖𝑖) = 11+𝑒𝑒 −𝑍𝑍𝑍𝑍𝑍𝑍𝑍𝑍𝑍 .
This generates an output map of size 256 x 256 x 2.
4. Overall Network Structure:
The overall architecture obtains the segmented map by
utilizing the encoder to create semantically meaningful
lower resolution features and then using the decoder along
with the combination of hierarchical features to build the
higher resolution structural characteristics. This combina -
tion allows the model to achieve both the benefits of the
high contextual understanding gained from ResNet101 as
well as the high precision in determining the location of
boundaries produced by U-Net, which makes this model
very well suited for medical imaging and remote sensing
segmentation tasks.
̂𝑌𝑌=𝑓𝑓 decoder (𝑓𝑓encoder(𝑋𝑋)) ,

1 3
471 Page 14 of 26
Environmental Earth Sciences (2026) 85:471
Fig. 14 Performance Evaluation of U-Net-ResNet101 Model for the index NDWI: (a) Confusion Matrix, (b) Training Accuracy, (c) Training Loss,
(d) Band 1, (e) Band 2, (f) Band 3

A systematic selection of hyperparameters Table 5 was car-
ried out for the developed deep learning model to optimize
effectively, achieve fast convergence times, and perform
well in generalizing the results. The training process was
conducted using a batch size of 8, which was found to com-
promise between providing stable gradients and taking into
account GPU memory limitations. The maximum number
of epochs for training was set at 100, thereby allowing an
adequate number of iterations to reach convergence, but
preventing excessive computation. The training optimiza -
tion for the model relied on the Adam Optimizer, a well-
established method of adaptively changing the learning
rate, and which has a high level of convergence speed using
default momentum values of 𝛽𝛽1=0.9 and 𝛽𝛽2=0.999. The
existing learning rate was set at 1×1 0 −4; thereby slowly
updating network weights with smooth, accurate adjust -
ments. To minimize overfitting, an early stopping mecha -
nism was also used for monitoring the validation loss, and
if no improvement was observed for five epochs in a row,
training would be terminated. A Model Checkpoint was also
created to track the weights with the lowest validation loss
to ensure that the best model was maintained. Each epoch,
the training loss, validation loss, training accuracy, and vali-
dation accuracy were recorded to enable a detailed exami -
nation of the model’s convergence and performance trends
over time. Collectively, hyperparameters were identified to
provide a stable and effective training process for the mod -
els to generalize well during image segmentation applica -
tions (Table 6).
Performance Metrics

  1. Precision: It measures the accuracy of positive predic -
    tions. It is the ratio of correctly predicted positive obser-
    vations to the total predicted positives.

1 3
Page 15 of 26 471
Environmental Earth Sciences (2026) 85:471
Fig. 15 Performance Evaluation of FCNN Model: (a) Confusion Matrix, ( b) Training Accuracy, (c) Training Loss, ( d) Band 1, ( e) Band 2, ( f)
Band 3

Precision = 𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇 + 𝑇𝑇𝑇𝑇 .
2. Recall: It measures the ability of the model to identify
actual positives. It is the ratio of correctly predicted posi-
tive observations to all the actual positives.
Recall = 𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇 + 𝑇𝑇𝑇𝑇 .
3. F1-score: It is the harmonic mean of Precision and
Recall. It gives a balance between precision and recall,
especially when class distribution is imbalanced.
F1-score =2× Precision × RecallPrecision + Recall.
4. IoU (Intersection over Union:) It measures the over -
lap between the predicted segmentation and the ground
truth. It is the ratio of the area of overlap to the area of
union.
IoU = 𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇 + 𝑇𝑇𝑇𝑇 + 𝑇𝑇𝑇𝑇 .
5. Confusion Matrix: It summarizes the performance
of a classification model. It shows the counts of true
positives (TP), true negatives (TN), false positives
(FP), and false negatives (FN), where TP (True Posi -
tive): Correctly predicted positive observations. FP
(False Positive): Incorrectly predicted as positive. TN
(True Negative): Correctly predicted negative obser -
vations. FN (False Negative): Incorrectly predicted as
negative.
𝐶𝐶𝑐𝑐𝑐𝑐𝑓𝑓𝐶𝐶𝐶𝐶𝑖𝑖𝑐𝑐𝑐𝑐 𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝑖𝑖𝐶𝐶 = [ 𝑇𝑇𝑇𝑇 𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇 𝑇𝑇𝑇𝑇]

1 3
471 Page 16 of 26
Environmental Earth Sciences (2026) 85:471
Fig. 16 Performance Evaluation of U-Net Model: (a) Confusion Matrix, ( b) Training Accuracy, (c) Training Loss, ( d) Band 1, ( e) Band 2, ( f)
Band 3

Results and discussions
The results of this study demonstrate the effectiveness of
the proposed framework for extracting and predicting water
bodies using NDWI and MNDWI indices, combined with
advanced thresholding techniques and deep learning mod -
els. The raster images of Madiwala Lake and Ulsoor Lake
for 843 and 1143 bands are visualized in Fig. 6.
NDWI images of Madiwala Lake and Ulsoor Lake for
different thresholding techniques are visualized in Figs. 7
and 8 respectively. Similarly, MNDWI images of Madiwala
Lake and Ulsoor Lake for different thresholding techniques
are visualized in Figs. 9 and 10 respectively.
Table 7 presents the water area analysis of Madiwala
Lake, showing threshold values and corresponding water
areas for three different sessions of the year, using various
band combinations, including NIR and Green, bands 8A and
11, and bands 8 and 12. Table 8 contains similar analysis
for Ulsoor Lake. Based on the findings, bands 8A and 11,
and 8 and 12 did not yield satisfactory results, leading to
their exclusion from further study. The experiment contin -
ued focusing on NDWI and MNDWI, where Otsu and Li
methods proved most effective.
Table 9 presents the performance comparison of FCNN,
U-Net, U-Net–ResNet50, and U-Net–ResNet101 models
using NDWI and MNDWI indices. For NDWI, all mod -
els achieve perfect precision, recall, and F1-score, indicat -
ing highly accurate water pixel classification. FCNN and
U-Net further attain an IoU of 1.00, reflecting near-perfect
spatial overlap. Although ResNet-based models show
slightly lower IoU values, their overall accuracy remains
high. In regard to MNDWI, we have determined that the
models’ precision/recall metrics show good performance
overall. However, the FCNN and U-Net models pro -
duce better (higher) IoU values than models based on the
ResNet architecture. The relatively low IoU values for the
two U-Net/ResNet50 and U-Net/ResNet101 combinations
may indicate some degree of boundary sensitivity resulting

1 3
Page 17 of 26 471
Environmental Earth Sciences (2026) 85:471
Fig. 17 Performance Evaluation of U-Net-ResNet50 Model: (a) Confusion Matrix, (b) Training Accuracy, (c) Training Loss, (d) Band 1, (e) Band
2, (f) Band 3

from the use of MNDWI as an input source. Overall, the
training and testing accuracy for all models have exceeded
97%, indicating that the models generalize very well. In
general, we observe that NDWI provides more stable seg -
mentation results across all models while MNDWI favors
simpler architecture. Qualitative visualizations presented
in the previous section of this report were consistent with
the results obtained here.
The performance evaluation of the models NDWI-FCNN,
NDWI-U-Net, NDWI-U-Net-ResNet50 and NDWI-U-Net-
ResNet101 are visualized in Figs. 11, 12, 13, and 14.
Furthermore, Figs. 15, 16, 17, and 18 show the per -
formance evaluation of the models MNDWI-FCNN,
MNDWI-U-Net, MNDWI-U-Net-ResNet50, and
MNDWI-U-Net-ResNet101.
Geographical location of selected lakes across India are
listed in Table 10
Implications for water management
The capacity to precisely map and forecast water bodies
has important ramifications for water management in urban
areas dealing with related to water resources and fast urban-
ization. These findings can help local governments keep an
eye on water levels, spot possible conservation zones, and
manage water resources more effectively.
FCNN and U-Net models successfully captured the
seasonal dynamics (pre-monsoon, monsoon, and post-
monsoon) of water bodies in Bangalore urban region,
accurately reflecting increases in water coverage during
the monsoon season and decreases during the pre and post-
monsoon periods Fig. 19. The U-Net and FCNN models
demonstrated strong performance in predicting water bod -
ies within urban regions of Bengaluru. Nevertheless, their
ability to generalize to other parts of India was restricted.

1 3
471 Page 18 of 26
Environmental Earth Sciences (2026) 85:471
Fig. 18 Performance Evaluation of U-Net-ResNet101 Model: (a) Confusion Matrix, ( b) Training Accuracy, (c) Training Loss, ( d) Band 1, ( e)
Band 2, (f) Band 3

Table 10 Geographical location of selected lakes across India
Lake Name State Region
Madiwala Lake Karnataka Southern India
Ulsoor Lake Karnataka Southern India
Tansa–Maha Lake Maharashtra Western India
Nanaksagar Reservoir Uttarakhand Northern India
Apparacheruvu Lake Andhra Pradesh Southern India
Hybrid architectures with deeper feature extractors—
U-Net–ResNet50 and U-Net–ResNet101—were created
to overcome this constraint. Among these, the U-Net–
ResNet101 model demonstrated superior robustness and
generalization performance by consistently and success -
fully predicting water bodies across all assessed geo -
graphic regions of India.The model’s dependability was
further confirmed by the seasonal estimates’ good align -
ment with historical water level data.
The seasonal predictions aligned well with historical data
on water levels, further validating the model’s reliability. For
a few lakes in various parts of India, including Madiwala,
Ulsoor, Tansa–Maha, Nanaksagar, and Apparacheruvu, the
expected lake extents derived from the U-Net, FCNN, U-Net–
ResNet50, and U-Net–ResNet101 models are shown. To assess
model consistency across spectral inputs, predictions are dis-
played for each lake using both NDWI and MNDWI indices.
The NDWI-based lake forecasts for Apparacheruvu (Andhra
Pradesh), Madiwala and Ulsoor (Karnataka), Nanaksagar
(Uttarakhand), and Tansa–Maha (Maharashtra) using four
deep learning models—FCNN, U-Net, U-Net-ResNet50, and
U-NetResNet101—are shown in Figs. 20, 21, 22, 23 and 24.
The subfigures for each lake are forecasts made by FCNN,
U-Net, U-Net–ResNet50, and U-Net–ResNet101, in that order.
Figures 25, 26, 27, 28 and 29 display the similar MNDWI-
based forecasts for the same lakes and models. In total, the
qualitative findings show that deeper encoder backbones—
specifically, U-Net–ResNet101—produce more reliable
lake segmentation performance throughout India’s various
geographical regions.

1 3
Page 19 of 26 471
Environmental Earth Sciences (2026) 85:471
Fig. 19 NDWI Madiwala lake predictions for different seasons: (a) season 1, (b) season 2, and (c) season 3

Fig. 20 NDWI Apparacheruvu: Predictions

Fig. 21 NDWI Madiwala: Predictions

We evaluated the cross-generalizability of the proposed
model across different geographical regions of Karnataka,
Andhra Pradesh, Maharashtra and Uttarakhand as men -
tioned in the article. The training and testing data were
obtained from geographically distinct regions, ensuring that
the model was evaluated on regions that were not used dur-
ing training. This spatially independent evaluation provides
an assessment of the model’s ability to generalize beyond
the specific geographical areas represented in the training
data. The results demonstrate the applicability of the pro -
posed framework across different regions and strengthen
the reliability of the reported performance.
Ablation study
Following five ablation study on the proposed U-Net-Resnet
101 model are detailed as below:

1 3
471 Page 20 of 26
Environmental Earth Sciences (2026) 85:471
Fig. 22 NDWI Ulsoor: Predictions

Fig. 23 NDWI Nanaksagar: Predictions

Fig. 24 NDWI Tansa Maha: Predictions

Fig. 25 MNDWI Apparacheruvu: Predictions

1 3
Page 21 of 26 471
Environmental Earth Sciences (2026) 85:471
Fig. 26 MNDWI Madiwala: Predictions

Fig. 27 MNDWI Ulsoor: Predictions

Fig. 28 MNDWI Nanaksagar: Predictions

Fig. 29 MNDWI Tansa Maha: Predictions

1 3
471 Page 22 of 26
Environmental Earth Sciences (2026) 85:471
Fig. 30 (a) Skip connections full,(b) Skip connections partial, (c) no skip connections

Fig. 31 (a) Deep_decoder, (b) Medium_Decoder and (c) Shallow_Decoder

  1. Skip connections: This model was created with high
    accuracy using the Full-Skip model method. We can
    see that with multi-scaling, it allows for the recovery
    of spatial context and spatial detail. The mid-level fea -
    ture maps provided by the Partial-Skip model were still
    very important, as this led to a small decrease in perfor -
    mance compared to the Full-Skip model. The No-Skip
    model had the largest drop in performance because the
    use of skip connections allows for the reconstruction of
    finer scale details and for the provision of information
    between the downsampling processes efficiently Fig. 30.
  2. Depth of decoder: To see how different levels of com -
    plexity in decoders affect segmentation efficacy, we
    developed three variations of our model using differ -
    ent amounts of complexity in the decoder within the
    ResNet101-U-Net framework. For example, our baseline
    model has four decoders, each decoder corresponds to
    an encoder depth. The medium complexity decoder has
    three decoders, meaning it skips one up-sampling step.
    Finally, our lowest decoding complexity model only has
    two decoders. We found the deeper models consistently
    produce better Dice Score values. This is most likely
    due to their ability to reconstruct multiple scales of the
    original image. While models with fewer depth produce
    similar computational efficiency, they tend to lose some
    level of detail and accuracy when constructing fine struc-
    tures and edges. Therefore, we conclude that decoder
    depth is important in maintaining spatial fidelity to the
    segmentation networks Fig. 31.
  3. Removal of Batch Normalization effect: To evaluate
    the effect of feature normalization on segmentation per -
    formance, we built a modification to the originally pro -
    posed ResNet101-U-Net architecture that removes all
    of the Batch Normalization (BN) layers. BN layers are
    often utilized in practice to normalize the intermediate
    feature maps and avoid the problem of internal covariate
    shift, and to improve the training stability. In this experi-
    ment, we eliminated every BN layer from the decoder’s
    convolutional blocks, which means that both models had
    the same encoder and training parameters. By remov -
    ing BN this way, we were able to determine how much
    of a difference BN makes to the convergence of the
    model, consistency of features, and the model’s ability
    to generalize. The performance difference between the
    ResNet101-U-Net baseline model and the model with -
    out BN demonstrates how important normalization is in
    keeping the features consistent as they propagate through
    the decoder of the model Fig. 32 (a) and (b).

1 3
Page 23 of 26 471
Environmental Earth Sciences (2026) 85:471
Fig. 32 (a) Batch Normalisation accuracy, (b) Batch Normalisation loss (c) U-Net_resnet101 Relu with leakyrelu accuracy, (d) U-Net_resnet101
Relu with leakyrelu loss (e) U-Net_resnet101 upsampling accuracy, (f) U-Net_resnet101 upsampling loss

  1. Replacing ReLU with Leaky ReLU effect: A differ -
    ent version of the ResNet101-U-Net model was created
    to study how activation nonlinearity affects propoga -
    tion of features and convergence. This was achieved by
    swapping the ReLU activations for LeakyReLU ( 𝛼𝛼 =
    0.1) within the architecture. The regression slope for the
    ReLU unit is flat, which can "saturate" when the input is
    negative and the gradient becomes smaller compared to
    positive inputs, causing neurons to become inactive. As
    a result, some LeakyReLU units are able to allow a very
    small negative slope through the activation levels, which
    enables gradients to flow across the full range of activa -
    tion levels, therefore reducing the possibility of inactive
    neurons. The removal of ReLU from the model provides
    a clearer picture of how activation functions impact the
    stability of gradients as they reconstruct feature maps
    in the decoder. Comparison of the performance of these
    types of model will indicate whether improved gradients
    result in increased accuracy with respect to segmentation
    and the sharpness of the boundary of the feature Fig. 32
    (c) and (d).
  2. Upsampling strategy effect: The use of the upsampling
    mechanism on spatial detail reconstruction was assessed
    by substituting all the non-learnable interpolation layers
    from the baseline decoder (bilinear interpolation, which
    enlarges fixed-scale features through fixed-scale meth -
    ods) for transposed convolution layers. Transposed con-
    volution allows for the learning of kernel parameters
    that can automatically learn to adaptively restore spatial
    features and boundary definition rather than limiting
    upsampling to bilinear interpolation, which has a fixed
    enlargement rate. The results from the transposed convo-
    lution upsampling variant were compared to the baseline
    decoder upsampling variant to provide insight into the
    degree of improvement achieved by using transposed
    convolution versus bilinear interpolation for producing
    high-quality segmentations. The results will be used to
    quantify the impacts of decoder design choices (e.g.,
    accuracy, computational cost, artifact generation) on
    segmentation performance using fine detail Fig. 32 (e)
    and (f).
    Grad-CAM visualizations illustrating the attention regions
    of the U-Net ResNet101 model for different lakes are shown
    in Fig. 33.
    In the prediction phase, the models were applied to
    predict water bodies for two major lakes—Madiwala and
    Ulsoor—across three seasonal sessions.

1 3
471 Page 24 of 26
Environmental Earth Sciences (2026) 85:471
Fig. 33 Gradcam Visualisations

Conclusion
In this study, we developed a robust framework for auto -
mated extraction and prediction of urban water bodies using
NDWI/MNDWI indices, thresholding techniques, and deep
learning models. Among the tested thresholding methods,
Otsu and Li provided the most accurate water delineation.
FCNN and U-Net performed well for Bengaluru lakes but
showed reduced generalization across lakes from other parts
of India. To address this, we evaluated U-Net-ResNet50
and U-Net-ResNet101, with the latter achieving signifi -
cantly superior and consistent performance. Grad-CAM
analysis and extensive ablation studies on upsampling,
skip connections, activation functions, decoder configu -
rations, and batch normalization confirmed the proposed
model’s optimal design. The framework’s dependability
was further confirmed by seasonal forecasts that matched
past water dynamics. Overall, the study shows that accurate
and scalable urban water body monitoring is made possible
by combining NDWI/MNDWI with an improved U-Net-
ResNet101 architecture. In order to improve model gener -
alizability, future research will examine bigger datasets and
more geographical areas.
Future scope
This work provides a foundation for future developments
in automated urban water body monitoring. Future research
may explore quantum computing for improved segmenta -
tion and optimization, while incorporating multi-temporal,
higher-resolution imagery and environmental factors such
as soil moisture and land use to enhance regional gener -
alization and modelling of seasonal and long-term water
dynamics.
In future work, a comprehensive statistical significance
analysis will be conducted using multiple independent
experimental runs to assess the reliability and robustness of
the observed performance differences. In addition, the pro -
posed framework will be experimentally evaluated against
recent state-of-the-art segmentation architectures, includ -
ing DeepLabV3+, SegFormer, and Attention U-Net, under
identical experimental conditions. These experiments will
provide a more detailed assessment of segmentation accu -
racy, generalization capability, computational complexity,
and training efficiency. The framework will also be further
developed toward real-time adaptation to support sustain -
able urban water-resource management and flood-risk
monitoring.
Author Contributions PM and SN is involved in the writing and cre -
ation of the proposed model. PS and ST are involved in the valida -
tion and supervision. Finally, all authors read and approved the final
manuscript.
Funding No funding has been received for this research.
Data Availability The Sentinel-2 images used in this study were
obtained using Google Earth Engine.
Declarations
Competing Interests The authors declare no competing interests.
Open Access This article is licensed under a Creative Commons
Attribution 4.0 International License, which permits use, sharing,
adaptation, distribution and reproduction in any medium or format,
as long as you give appropriate credit to the original author(s) and the
source, provide a link to the Creative Commons licence, and indicate
if changes were made. The images or other third party material in this
article are included in the article’s Creative Commons licence, unless
indicated otherwise in a credit line to the material. If material is not

1 3
Page 25 of 26 471
Environmental Earth Sciences (2026) 85:471
included in the article’s Creative Commons licence and your intended
use is not permitted by statutory regulation or exceeds the permitted
use, you will need to obtain permission directly from the copyright
holder. To view a copy of this licence, visit h t t p : / / c r e a t i v e c o m m o n s . o
r g / l i c e n s e s / b y / 4 . 0 / .
References
Bikku T, Chappala R, Rao AN, Gaddam VG, Joseph S, Thota S (2025)
Explainable deep reinforcement learning for climate forecasting
with transfer learning. Environm Sci Pollut Res. h t t p s : / / d o i . o r g / 1
0 . 1 0 0 7 / s 1 1 3 5 6 - 0 2 5 - 3 7 0 9 4 - 9
Cai Y , Shi Q, Liu X (2024) Spatiotemporal mapping of surface water
using landsat images and spectral mixture analysis on google
earth engine. J Remote Sens 4:117. h t t p s : / / d o i . o r g / 1 0 . 3 3 9 0 / j r s 4
0 4 0 1 1 7
Chen F, Chen X, Van de V oorde T, Roberts D, Jiang H, Xu W (2020)
Open water detection in urban environments using high spa -
tial resolution remote sensing imagery. Remote Sens Environ
242:111706. https:/ /doi.or g/10.10 16/j. rse.2020.111706
Dong X, Hu C, Zhao Y (2023) Novel threshold self-regulating water
extraction method. J Hydrol Eng 28(8):04023020. h t t p s : / / d o i . o r g
/ 1 0 . 1 0 6 1 / J H Y E F F . H E E N G - 5 8 9 1
Embaby A, Youssef YM, Abu El-Magd SA (2024) Delineation of lin-
eaments for groundwater prospecting in hard rocks: inferences
from remote sensing and geophysical data. Environm Earth Sci
83:62. https:/ /doi.or g/10.10 07/s1 2665-023-11389-x
Francis A, Sidiropoulos P, Muller J-P (2019) CloudFCN: Accurate and
robust cloud detection for satellite imagery with deep learning.
Remote Sens 11(19):2312. https:/ /doi.or g/10.33 90/rs 11192312
Hussein F, Latifi H, Mojaradi B (2025) Leveraging sentinel-2 A time
series and integrated cellular automata-deep learning for accu -
rate land cover forecasts in Northern Iraq. Environm Earth Sci
84:586. https:/ /doi.or g/10.10 07/s1 2665-025-12561-1
Kang J, Guan H, Peng D, Chen Z (2021) Multi-scale context extractor
network for water-body extraction from high-resolution optical
remotely sensed images. Int J Appl Earth Obs Geoinf 103:102499.
https:/ /doi.or g/10.10 16/j. jag.2021.102499
Li A et al (2021a) Comparative analysis of machine learning algorithms
in automatic identification and extraction of water boundaries.
Appl Sci 11(21):10062. https:/ /doi.or g/10.33 90/ap p112110062
Li L, Su H, Du Q, Wu T (2021b) A novel surface water index using
local background information for long term and large-scale
Landsat images. ISPRS J Photogramm Remote Sens 172:59–78.
https:/ /doi.or g/10.10 16/j. isprsjprs.2020.12.013
Li H, Zech J, Ludwig C, Fendrich S, Shapiro A, Schultz M, Zipf A
(2021c) Automatic mapping of national surface water with Open-
StreetMap and Sentinel-2 MSI data using deep learning. Int J
Appl Earth Obs Geoinf 104:102571. h t t p s : / / d o i . o r g / 1 0 . 1 0 1 6 / j . j a
g . 2 0 2 1 . 1 0 2 5 7 1
Li Z, Zhang X, Xiao P (2022) Spectral index-driven FCN model train-
ing for water extraction from multispectral imagery. ISPRS J
Photogramm Remote Sens 192:344–360. h t t p s : / / d o i . o r g / 1 0 . 1 0 1 6
/ j . i s p r s j p r s . 2 0 2 2 . 0 1 . 0 0 1
Lilay MY , Taye GD (2023) Semantic segmentation model for land
cover classification from satellite images in Gambella National
Park, Ethiopia. SN Appl Sci 5:76. h t t p s : / / d o i . o r g / 1 0 . 1 0 0 7 / s 4 2 4 5
2 - 0 2 3 - 0 5 2 8 0 - 4
Liu S et al (2023) Comparing water indices for Landsat data for auto -
mated surface water body extraction under complex ground back-
ground: A case study in Jilin Province. Remote Sens 15(6):1678.
https:/ /doi.or g/10.33 90/rs 15061678
Luo X, Hu Z, Liu L (2023) Investigating the seasonal dynamics of
surface water over the Qinghai-Tibet Plateau using Sentinel-1
imagery and a novel gated multiscale ConvNet. Int J Digit Earth
16:1372–1394. https:/ /doi.or g/10.10 80/17 538947.2022.2091852
Mohanty SP et al (2020) Deep learning for understanding satellite
imagery: An experimental survey. Front Artif Intell 3:534696.
https:/ /doi.or g/10.33 89/fr ai.2020.534696
Ngamile S, Kganyago M, Madonsela S, Mvandaba V (2025) Charac -
terising the spatio-temporal patterns of water quality parameters
in the cradle of humankind world heritage site using Sentinel-2
and random forest regressor. Front Remote Sens 6:1631403
Pallavi M, Thivakaran TK, Ganapathi C (2022) Evaluation of land use/
land cover classification based on different bands of sentinel-2
satellite imagery using neural networks. Int J Adv Comput Sci
Appl 13:10. https:/ /doi.or g/10.14 569/I JACSA.2022.0131070
Pallavi M, Thivakaran TK, Ganapathi C (2022a) Evaluation of land
use/land cover classification based on different bands of senti -
nel-2 satellite imagery using neural networks. Intern J Adv Com-
put Sci Appl. https:/ /doi.or g/10.14 569/I JACSA.2022.0131070
Pallavi M, Thivakaran TK, Ganapathi C (2022b) "A tile-based
approach for the LULC classification of sentinel image using
deep learning techniques. In: 2022 International Conference for
Advancement in Technology (ICONAT). IEEE. h t t p s : / / d o i . o r g / 1 0
. 1 1 0 9 / I C O N A T 5 3 4 2 3 . 2 0 2 2 . 9 7 2 6 0 3 0
Pan Z et al (2020) Deep learning segmentation and classification for
urban village using a worldview satellite image based on U-Net.
Remote Sens 12(10):1574. https:/ /doi.or g/10.33 90/rs 12101574
Sathish Kumar M (2023) Extraction of surface water extent: Auto -
mated thresholding approaches. Environ Sci Proc 29(1):31.
https:/ /doi.or g/10.33 90/EC RS2023-15861
Sharma K, Tiwari R, Wadhwani AK, Chaturvedi S (2024) Evaluating
the impact of land use land cover changes on urban ecosystem
services in Nashik, India: a RS-GIS based approach. Environm
Earth Sci 83:683. https:/ /doi.or g/10.10 07/s1 2665-024-11965-9
Sun Q, Li J (2024) A method for extracting small water bodies based
on DEM and remote sensing images. Sci Rep 14:760. h t t p s : / / d o i .
o r g / 1 0 . 1 0 3 8 / s 4 1 5 9 8 - 0 2 4 - 0 1 9 3 1 - 2
Thota S, Bikku T, Mortezaagha P (2026a) Integrative approaches for
sustainable and resilient water-food systems under climate and
socio-economic pressures. Discov Environm 4:252. h t t p s : / / d o i . o r
g / 1 0 . 1 0 0 7 / s 4 4 2 7 4 - 0 2 6 - 0 0 7 1 9 - y
Thota S, Bikku T, Shanmugasundaram P (2026b) Quantum-enhanced
graph-based ecological modeling for food web stability analysis.
Mathematics Open 5:2650016. h t t p s : / / d o i . o r g / 1 0 . 1 1 4 2 / S 2 8 1 1 0 0 7
2 2 6 5 0 0 1 6 1
Thota S, Kumar C, Shanmugasundaram P (2026c) Twenty years of
mathematical modeling in ecology: a thematic architecture and
macroecological trend analysis. BMC Ecology and Evolution
26:97. https:/ /doi.or g/10.11 86/s1 2862-026-02605-w
Wang Z et al (2020) MSLWENet: A novel deep learning network for
lake water body extraction of Google remote sensing images.
Remote Sens 12(24):4140. https:/ /doi.or g/10.33 90/rs 12244140
Wieland M et al (2024) S1S2-Water: A global dataset for semantic seg-
mentation of water bodies from sentinel- 1 and sentinel-2 satellite
images. IEEE J Sel Top Appl Earth Obs Remote Sens 17:1084–
1099. https:/ /doi.or g/10.11 09/JS TARS.2023.3333969
Yan X, Song J, Liu Y , Lu S, Xu Y , Ma C, Zhu Y (2023) A Transformer-
based method to reduce cloud shadow interference in automatic
lake water surface extraction from Sentinel-2 imagery. J Hydrol
620:129561. https: //d oi. or g/10 . 10 16/j. jhyd rol.2023.129561
Publisher''s Note Springer Nature remains neutral with regard to juris-
dictional claims in published maps and institutional affiliations.

论文原文与出版信息

Environmental Earth Sciences
查看论文原文

继续探索