Rapid Flood Susceptibility Mapping in the Indian Himalayan Region using CNN-U-Net Segmentation: Insights from the 2025 Monsoon Events

Rapid Flood Susceptibility Mapping in the Indian Himalayan Region using CNN-U-Net Segmentation: Insights from the 2025 Monsoon Events | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Rapid Flood Susceptibility Mapping in the Indian Himalayan Region using CNN-U-Net Segmentation: Insights from the 2025 Monsoon Events Rachit ., Vaibhav Tripathi, Mohit Prakash Mohanty, Ashish Pandey, and 1 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-7715126/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted 5 You are reading this latest preprint version Abstract The Indian Himalayan Region (IHR) is increasingly threatened by hydro-meteorological hazards such as cloudbursts, flash floods, and landslides, driven by climatic extremes and rapid land-use changes. The June–August 2025 monsoon floods in the states of Himachal Pradesh and Uttarakhand in India highlighted this vulnerability, causing severe loss of life and infrastructure damage. To address the urgent need for rapid and reliable information tools during the crucial initial hours of disaster occurrence in complex, data-scarce environments, this study develops a rapidly deployable deep learning-based flood susceptibility mapping framework tailored for the IHR. The framework employs a Convolutional Neural Network (CNN) with U-Net architecture, integrating 14 hydro-geomorphological predictors (e.g., altitude, slope, TWI, NDVI). A novel Local Convexity Factor (LCF), adaptively calibrated using curvature and slope, enhances micro-topographic characterisation, improving hazard delineation in rugged landscapes and reducing overestimation in depositional zones. The model achieved strong predictive skill (accuracy = 97.12%, RMSE ≈ 0.145, CSI = 68.44%, AUC ROC ≈ 0.99), demonstrating high predictive reliability despite limited flood observations. Compared to Sentinel-1 SAR and the JRC Global Flood Hazard Map, the CNN–U-Net approach effectively captures both riverine and upland flood hotspots by understanding the hidden patterns in catchment physiography. Designed for rapid retraining and deployment within hours, the framework functions as an operational, stakeholder-ready information tool, supporting early warning, land-use planning, and climate-resilient infrastructure development. Beyond flood mapping, the proposed framework can be utilized for multi-hazard susceptibility mapping in other mountainous and data-scarce regions worldwide. CNN Deep learning Flood Indian Himalayas Multi-Hazard Susceptibility U-net Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Figure 7 1. Introduction Mountain regions worldwide are experiencing unprecedented vulnerability to hydro-meteorological disasters, with climate change intensifying flood risks across major mountain ranges globally (Rick et al. 2023 ). Mountains across the globe are emerging as climate hotspots, where enhanced warming and altered precipitation regimes threaten cryospheric reserves, water security, and ecosystem stability (Pepin et al. 2022 ). Ombadi et al. ( 2023 ) showed that mountains across the Northern Hemisphere are becoming increasingly susceptible to extreme rainfall events, with every 1°C of global warming leading to approximately 15%(Pepin et al., 2022 ) more rain at high elevations, substantially exceeding the rate expected from atmospheric moisture increases alone. The European Alps have documented significant increases in torrential flooding, with climate change and settlement growth contributing to rising losses, while glacial lake outburst floods (GLOFs) across mountain regions have increased in frequency, with documented events showing temporal increases globally (Schlögl et al., 2021 ). According to the Emergency Events Database (EM-DAT), there were 203 flood disasters worldwide in 2023, resulting in 7,763 deaths and economic losses exceeding US $ 20.4 billion (CRED 2023). The Indian Himalayan Region (IHR) is inherently vulnerable to various geological and hydro-meteorological disasters due to its geographical setting, which is characterised by steep, unstable terrain and the prevalence of erratic weather patterns (Nagamani et al. 2024 ; Shah and Malakar 2024 ; Roul et al. 2025 ). Over the past eight years, the state of Uttarakhand has experienced over 25,000 disaster incidents, resulting in 705 deaths from flash floods and landslides, including 389 deaths from flash floods alone (India Today 2025; The Times of India 2025 a). In 2021, the state of Himachal Pradesh reported 476 deaths and ₹1,151 crore in losses, while in 2023, there were 441 fatalities and ₹12,000 crore in damages (Down To Earth 2025 ). In 2024-25, Himachal Pradesh recorded 358 deaths from hydro-meteorological disasters, the second-highest in India. The state also reported the loss of over 7,000 cattle and damage to 1,004 houses (The Times of India 2024 ). In June–August 2025, the states of Himachal Pradesh and Uttarakhand in India experienced intense monsoon activity leading to cloudbursts, flash floods, and landslides, resulting in over 100 fatalities and extensive infrastructure damage (NDTV 2025). On 5th August 2025, the Dharali village located in the Uttarkashi district of Uttarakhand experienced significant destruction of residential and commercial structures. The incident led to the loss of numerous lives, with several people missing or crushed under mudflows. This devastation was primarily attributed to a substantial debris slide resulting from a potential cloudburst and ensuing flash floods (The Hindu 2025 ; The Times of India 2025 b). Such recurring incidences in the IHR are primarily driven by both natural and human-induced changes, which significantly contribute to flood occurrences (Nagamani et al. 2024 ; Alam et al. 2025 ). These areas are prone to multiple hazards that often co-occur and interact across space and time, thus elevating the need for comprehensive susceptibility maps to support integrated risk management (Tripathi et al. 2023 ; Sreevalsan-Nair and Mundayatt 2025 ). Despite the high propensity for disasters in the IHR, many regions continue to face severe data scarcity (Agrawal et al. 2018; Sultan et al. 2022 ). To overcome this challenge, satellite-based observations, particularly those derived from Sentinel-1 SAR, have become indispensable for inundation mapping, owing to their unique capability to capture imagery through clouds and in the absence of daylight (Uddin et al. 2019 ; Konapala et al. 2021 ). However, their efficacy in the Himalayan terrain remains limited. Sentinel-1 revisit times may miss short-lived floods (~ 58% detectability for events), while issues such as topographic distortions, vegetation interference, and rapidly changing flood conditions further reduce their reliability (Tarpanelli et al. 2022 ; Breznik et al. 2025 ). In such contexts, Hydraulic-cum-Hydrodynamic Modelling (HHM) offers a valuable alternative by simulating flood dynamics across diverse terrains and hydrological settings with high accuracy. However, its operational application is often constrained by the need for extensive multidimensional input data and substantial computational resources (Pareta 2024 ). These requirements pose a major hindrance for such regions, especially those existing in the low and middle-income nations, where financial and technical resources for data procurement are limited (Bentivoglio et al. 2022 ). Given these limitations, there is a pressing need for approaches that can overcome data scarcity while remaining operationally feasible in resource-constrained settings. In this regard, data-driven methods, particularly Deep Learning (DL) models, are emerging as powerful tools for flood mapping and risk assessment in data-scarce regions (Karim et al. 2023 ; Fereshtehpour et al. 2024 ). In disaster-prone regions like the IHR, where rapid and reliable information is essential during the crucial initial hours and data availability is limited, the rapid deployment capability of CNN-based models provide a decisive operational advantage for flood susceptibility mapping. While traditional hydrodynamic models require 40 + minutes to several days for complete flood simulations, CNN models can generate high-resolution flood susceptibility maps within seconds to minutes after training (Song et al. 2025 ; Taysi et al. 2025 ). This dramatic increase in speed, achieving computational efficiency several times greater than conventional approaches, enables near real-time hazard assessment, crucial for emergency response. Traditional HHM requires cross-sections, bathymetry, roughness coefficients, boundary conditions, discharge records, gauged calibration data, and meshing, along with extensive computational resources and continuous monitoring inputs for reliable simulation of flood dynamics (Karim et al. 2023 ). In contrast, DL models can operate using satellite or remote sensing data combined with minimal ground truth information, such as flood extent, making them particularly suited for regions that face data scarcity and resource limitations (Karim et al. 2023 ; Khosravi et al. 2023 ; Fereshtehpour et al. 2024 ; Biazar et al. 2025 ). DL models, especially Convolutional Neural Networks (CNNs), offer critical advantages in contexts where accurate representation of spatial flood characteristics is required, as they can capture complex feature interactions and produce more reliable susceptibility assessments and hazard maps than conventional machine learning and numerical approaches (Wang et al. 2020 ; Bentivoglio et al. 2022 ). They are particularly valuable for operational workflows in large mountainous regions such as the IHR, where they can support timely planning and early-stage risk zoning without the need for computationally intensive full hydrodynamic simulations across numerous scenarios (Karim et al. 2023 ). CNNs are adept at learning spatial patterns and non-linear interactions across geospatial grids, enabling reliable predictions in areas where satellite observations are limited. Ullah et al. ( 2022 ) found that CNN-based models outperform traditional machine learning methods, capturing spatial interconnectedness between neighbouring pixels more effectively. By integrating multi-source predictors such as digital elevation models, land cover, rainfall intensity, and hazard inventories, CNNs provide a robust framework for early-stage risk zoning and flood susceptibility mapping, without the prohibitive computational costs of full hydrodynamic simulations (Karim et al. 2023 ). Looking ahead, such models hold strong potential to serve as the backbone for rapid operational, AI-enabled decision-support systems tailored to the Himalayan context, ultimately advancing multi-hazard resilience and preparedness. In the context of the IHR, where disasters evolve rapidly and data scarcity hampers timely decision-making, the need for a rapid, reliable, and easily deployable information tool is paramount. Previous research has demonstrated that topographic variables, including elevation, slope, drainage density, and relative relief, constitute the most influential predictors in machine learning based flood susceptibility assessments in mountainous areas (Zhao et al. 2018 ). Furthermore, recent studies emphasize that even subtle micro-topographic variations in relatively flatter valley regions are critical for accurately delineating localized flood hazards, as concave and convex terrain features strongly influence water accumulation and flow dynamics (Safaei-Moghadam et al. 2023 ). These findings underscore the necessity of incorporating both large-scale morphometric controls and fine-scale topographic features to enhance the reliability of flood hazard mapping in data-scarce and topographically complex environments such as the Himalayas. This study addresses this gap and advances flood susceptibility mapping in complex, data-scarce mountainous regions through three key innovations. First, we integrate a Local Convexity Factor (LCF) into a CNN U-Net framework, enabling enhanced micro-topographic characterisation that improves hazard delineation in rugged landscapes and reduces overestimation in depositional zones representing a key advancement for flood risk modelling in the complex geomorphological context of the Indian Himalayas. Second, by combining 14 relevant hydro-geomorphological predictors, including terrain, land cover, soil, and rainfall climatology within a DL semantic segmentation model, we capture both riverine and upland flood hotspots that are often overlooked by conventional approaches in mountainous terrain settings. Third, the proposed framework is lightweight and rapidly deployable, capable of retraining within hours, making it operationally suitable for regions where near-real-time hazard mapping is critical but hydrodynamic simulations are impractical due to data or resource constraints. This method leverages learned terrain-hydrology patterns to deliver high-resolution susceptibility maps with superior spatial detail and predictive reliability. Beyond its immediate application in the Indian Himalayan Region, the framework is readily transferable to other mountainous and data-scarce environments worldwide, supporting integrated multi-hazard risk assessment. Its design enables a practical decision support tool for land-use planners, policymakers, and disaster management agencies. By informing targeted interventions, guiding infrastructure placement, and enhancing early-warning capabilities, the approach contributes to building climate-resilient cities and communities capable of adapting to increasingly unpredictable hazard regimes. 2. Study area The present study focuses on the northwestern segment of the Indian Himalayan Region (IHR), specifically the states of Himachal Pradesh (HP) and Uttarakhand (UK) ( Fig. 1 ) . These mountain states are highly vulnerable to hydro-meteorological extremes such as intense precipitation, flash floods, and glacial lake outburst floods, along with frequent landslides and riverbank erosion, owing to their fragile topography and active tectonics (Kansal and Singh 2022 ; Gupta et al. 2024 ). Uttarakhand is located between 28°43’ and 31°28’ N latitude and 77°34’ and 81°3’ E longitude, while Himachal Pradesh lies between 30°22’ and 33°16’ N latitude and 75°35’ and 79°01’ E longitude. The elevation in these regions ranges from 187 meters to approximately 7,124 meters, encompassing both states collectively. Studies and analyses indicate that over 45% of Himachal Pradesh is highly prone to floods, landslides, and avalanches, while 85% of the districts in Uttarakhand are vulnerable to extreme floods (CEEW 2021 ; PreventionWeb 2025 ). The region has experienced devastating flood events, including the 2013 Kedarnath disaster (6,054 fatalities) and July 2023 Himachal Pradesh floods (187 fatalities, economic losses of ₹5,620 crore) (UNDRR CRED 2022; Sphere India 2023 ). The increasing frequency and intensity of hydro-meteorological disasters since 1997, combined with rapid glacial lake expansion (1,048 glacial lakes in Himachal's Sutlej catchment as of 2023), establishes this region as a critical area for comprehensive flood susceptibility assessment (India Today 2024 ). 3. Materials and Methods The materials and methodology include the selection and processing of 14 flood-conditioning factors, the design and training of a CNN U-Net model for flood susceptibility prediction, evaluation of model performance using various metrics, analysis of factor contributions through permutation feature importance, and delineation of flood extent using Sentinel-1 SAR imagery. The overall workflow, integrating multi-source geospatial data, CNN U-Net modeling, and Sentinel-1 SAR flood mapping, is illustrated in Fig. 2 . 3.1 Flood Conditioning Factors A total of 14 flood-conditioning factors were selected based on their proven relevance in flood susceptibility modelling, as supported by previous studies (Saravanan et al. 2023 ; Amiri et al. 2024 ; Dey et al. 2024 ; Tripathi and Mohanty 2024 ). The rasters for the conditioning factors are stacked as a 3D tensor \(\:X\in\:{\mathbb{R}}^{H\times\:W\times\:N}\) where H , and W are spatial dimensions and N is the number of predictors; No Data values are converted to “NaN” at source read time and later filled for model input during patching. Each band is standardized using z-score normalization computed spatially, per band (Eq. 1 ): $$\:{X}_{norm}=\:\frac{X-\mu\:}{\sigma\:}$$ 1 where µ is computed for each band by averaging all pixels across the image height and width while ignoring missing values. The standard deviation σ is computed similarly for each band across the image height and width, again ignoring missing values (NaNs). This is the dispersion of pixel values over rows and columns for that band. The standardized multi-band array is written to a Zarr store (an open, chunked, and compressed storage format for large N -dimensional arrays) with carefully chosen chunk sizes, enabling block-wise, random access to small subarrays without loading the full raster into memory. This chunked, compressed layout accelerates sliding-window training and inference by reading only the tiles needed at each step, supports parallel input/output (multiple patches read or written concurrently), and scales efficiently to very large rasters on local disks or object storage. Details of chosen flood conditioning factors, viz., sources and spatio-temporal resolutions, are listed in Table 1 , with significance discussed in the following subsections. 3.1.1 Altitude In the Himalayan context, elevation directly influences precipitation patterns, temperature gradients, and snow-melt processes that contribute to flood generation (Patel et al. 2022 ). Lower elevations in the foothills and valley regions are more susceptible to flooding due to gravitational water flow and natural accumulation processes from surrounding higher terrain (Ullah et al. 2022 ; Dey et al. 2024 ; Nagamani et al. 2024 ; Kumar et al. 2024 ) ( Fig. 1 ) . 3.1.2 Slope The rugged Himalayan terrain exhibits steep slopes that promote rapid runoff generation, but areas with gentle slopes (0–5°) in valley bottoms and terraced agricultural regions are more prone to water accumulation and flooding (Patel et al. 2022 ). Slope characteristics in the study region significantly influence surface water flow velocity and flood generation patterns, with the region's diverse topography ranging from gentle valley floors to extremely steep mountain faces ( Fig. 3 a ) . 3.1.3 Aspect Aspect orientation in the Himachal Pradesh and Uttarakhand regions critically influences flood susceptibility through its control over monsoon precipitation patterns, solar radiation, and soil moisture conditions (Nagamani et al. 2024 ). The southwestern and southeastern facing slopes receive the maximum impact from monsoon winds carrying moisture from the Bay of Bengal, resulting in enhanced precipitation and increased flood risk (Sajwan and Sushil 2016 ; Nagamani et al. 2024 ) ( Fig. 3 b ) . 3.1.4 Plan curvature Plan curvature describes the curvature of the land surface perpendicular to the direction of maximum slope, indicating whether water flow converges or diverges across the landscape (Edamo et al. 2024 ). Concave surfaces tend to concentrate flow and promote water accumulation, increasing flood susceptibility, while convex surfaces facilitate flow dispersion and reduce flood risk (Aydin and Iban 2023). Flat areas with minimal curvature are particularly vulnerable to flooding as they provide limited natural drainage (Mehravar et al. 2023) ( Fig. 3 c ) . 3.1.5 Topographic wetness index (TWI) TWI quantifies the potential for water accumulation at any location based on the upstream contributing area and local slope gradient (Al-Kindi and Alabri 2024; Amiri et al. 2024 ). Higher TWI values indicate areas with greater potential for soil saturation and surface runoff generation, making them more susceptible to flooding (Aydin and Iban 2023). This index effectively captures the topographic control on hydrological processes and is calculated as the natural logarithm of the ratio between specific catchment area and slope tangent (Li and Hong 2023; Khosravi et al. 2019) ( Fig. 3 d ) . 3.1.6 Topographic roughness index (TRI) TRI measures the variability in elevation within a local area and indicates surface complexity and roughness (Aydin and Iban 2023). Areas with high TRI values typically exhibit complex terrain with varied elevation changes that can create local drainage complications and influence flood flow patterns (Al-Kindi and Alabri 2024) ( Fig. 3 e ) . 3.1.7 Local convexity factor (LCF) The LCF was used as a conditioning factor in the deep learning framework to enhance flood susceptibility mapping in the complex terrain of the Indian Himalayas. Proposed by Liu et al. ( 2025 ), LCF utilizes the correlation between normalized Digital Elevation Model (DEM) patches and a standard Gaussian surface to identify micro-depressions that concentrate runoff, capturing subtle terrain features often missed by conventional curvature and slope metrics. The LCF methodology proposed by Liu et al. ( 2025 ) iterates over a range of window sizes (2–50) and selects the maximum correlation value for each pixel, but this can lead to over or underemphasis of features, especially in heterogeneous terrain. The presented methodology introduces an adaptive windowing approach for pixel-wise calculation of LCF based on local slope and curvature extracted from a DEM. The procedure entails loading and preprocessing the DEM, normalizing elevation values, computing slope and curvature derivatives, and dynamically determining the local window size for each pixel based on these factors. For each pixel, the analysis window size ( W ) is dynamically selected based on terrain factors using the following formulation (Eq. 2 ): $$\:W=clip({k}_{1}\:.\:slope+\:{k}_{2}.\:\left|curvature\right|+\:{k}_{3},\:2,\:50)$$ 2 Where k 1 , k 2 and k 3 are user-set parameters, and W is clipped between 2 and 50 to remain numerically stable. Slope and curvature were first normalized to the range [0,1]. The constants k 1 , k 2 and k 3 control the influence of slope, curvature, and a baseline offset, respectively. We set k 1 = 30, k 2 = 20 and k 3 = 2. These values were chosen so that, for our ~ 90 m DEM, the resulting window sizes realistically capture the scale of the hilly landforms. The mean window size corresponded to about 1.3 km on the ground, which is large enough to capture meso-scale landform patterns (ridge-valley contrasts, slope breaks), with very few pixels reaching the lower (2 pixel) or upper (50 pixel) limits. In this formulation, slope has slightly more weight than curvature, meaning that steep gradients tend to produce larger windows, while curvature refines the adjustment in convex and concave terrain. The offset k 3 ​ ensures that even very flat regions retain a minimal window. We verified the suitability of these parameters through simple diagnostics, including histograms of window size distribution and inspection of spatial maps, and found that small changes in the parameter values did not substantially alter the results. Building on this foundation, the subsequent step involves extracting a normalized DEM patch of the adaptive window size and correlating it with a Gaussian reference surface. The resulting Pearson correlation coefficient serves as the Local Convexity Factor (LCF). Raster outputs are exported for further geomorphological analysis and hazard assessment. The adaptive nature of this method ensures precise, terrain-sensitive quantification of convexity, outperforming the standard fixed-window technique ( Fig. 3 f ) . 3.1.8 Drainage density Drainage density represents the total length of stream channels per unit area and indicates the drainage efficiency of a watershed (Al-Kindi and Alabri 2024). High drainage density typically facilitates rapid water evacuation and reduces flood risk, while low drainage density can lead to water accumulation and increased flood susceptibility (Abusarif et al. 2023) ( Fig. 3 g ) . 3.1.9 Distance from river Proximity to water bodies is a crucial factor in flood susceptibility assessment, with areas closer to rivers and streams exhibiting significantly higher flood risk (Al-Kindi and Alabri 2024; Dey et al. 2024 ). The relationship between distance and flood susceptibility typically follows an inverse exponential decay, where flood probability decreases rapidly with increasing distance from water sources (Aydin and Iban 2023) ( Fig. 3 h ) . 3.1.10. Lithology Geological formations control subsurface drainage, infiltration capacity, and groundwater flow patterns, significantly influencing surface flood behaviour (Al-Kindi and Alabri 2024; Khosravi et al. 2019). Impermeable rock types such as granite and schist reduce infiltration capacity and increase surface runoff, while permeable formations like limestone and sandstone promote groundwater recharge and reduce flood susceptibility (Aydin and Iban 2023). Lithology data from the Geological Survey of India (GSI) were used in this study. A total of 295 lithologic classes were identified in the same across Himachal Pradesh and Uttarakhand, as shown in Fig. 3 i. 3.1.11 Rainfall The rainfall data for the study area was derived by calculating the annual average from the IMD gridded time-series data spanning from 1990 to 2024. This annual average raster was then converted into point data, which were subsequently interpolated using the inverse distance weighting (IDW) method at a pixel size of 90 meters. This process resulted in a uniformly distributed rainfall raster for the region ( Fig. 3 j ) . 3.1.12 Land use land cover (LULC) LULC types determine surface permeability, roughness characteristics, and runoff coefficients, making them essential factors in flood modeling (Al-Kindi and Alabri 2024; Amiri et al. 2024 ). Urban areas with impervious surfaces exhibit high runoff coefficients and increased flood susceptibility, while forested areas provide natural flood protection through enhanced infiltration and flow retardation (Aydin and Iban 2023). Agricultural lands show intermediate flood susceptibility depending on crop type, soil conditions, and management practices (Ahmadlou et al. 2021) ( Fig. 3 k ) . 3.1.13 Normalized difference vegetation index (NDVI) NDVI quantifies vegetation density and health, which directly affects surface roughness, infiltration capacity, and runoff generation (Al-Kindi and Alabri 2024; Dey et al. 2024 ). Dense vegetation increases surface roughness, promotes infiltration, and reduces flood susceptibility through enhanced evapotranspiration and flow retardation (Amiri et al. 2024 ). Areas with low NDVI values, indicating sparse vegetation or bare soil, exhibit higher flood susceptibility due to reduced infiltration capacity and increased surface runoff (Aydin and Iban 2023; Mehravar et al. 2023) ( Fig. 3 l ) . 3.1.14 Soil Moisture Antecedent soil moisture strongly conditions the fraction of rainfall that becomes direct runoff, with higher soil moisture leading to substantially increased flood peaks because more precipitation is converted to surface flow rather than infiltrating the soil (Merz et al. 2021 ). This nonlinear relationship means that modest increases in precipitation can produce disproportionately large increases in runoff when soils are near saturation (Merz et al. 2021 ). The mean top layer (0–5 cm) soil moisture from March 2015 to May 2025 was computed and used as an input feature to the CNN ( Fig. 3 m ) . Since CNNs are supervised learning algorithms, they require labelled data for training. To provide these labels, we utilised the Joint Research Centre (JRC) Global River Flood Hazard Map (100-year return period) (Baugh et al. 2024 ) as a reference flood hazard inventory, which served as the ground truth for model training and evaluation. Table 1 Relevant conditioning factors for flood in the study area, their sources and resolution. S. No. Flood Conditioning Factors Scale/Resolution Data Source 1 Altitude (DEM) 90 meters MERIT ( https://developers.google.com/earth-engine/datasets/catalog/MERIT_DEM_v1_0_3 ) 2 Slope 3 Aspect 4 Plan curvature 5 TWI 6 TRI 7 LCF 8 Drainage density Vector data (Projection: WGS 1984 LCC) India WRIS rivers shapefile ( https://indiawris.gov.in/wris/#/geoSpatialData ) 9 Distance from the river 10 Lithology 1:50,000 GSI Bhukosh ( https://bhukosh.gsi.gov.in/Bhukosh/ ) 11 Rainfall 0.25° × 0.25°; annual average (1990–2024) IMD ( https://www.imdpune.gov.in/lrfindex.php ) 12 LULC 500 m MODIS data ( https://developers.google.com/earth-engine/datasets/catalog/MODIS_061_MCD12Q1 ) 13 NDVI 250 m MODIS data ( https://developers.google.com/earth-engine/datasets/catalog/MODIS_061_MYD13Q1 ) 14 Soil Moisture 11 km SMAP mission ( https://nsidc.org/data/spl4smgp/versions/7 ) * MERIT: Multi-Error-Removed Improved-Terrain; WRIS: Water Resources Information System; GSI : Geological Survey of India; IMD: India Meteorological Department; MODIS: Moderate Resolution Imaging Spectroradiometer; SMAP: Soil Moisture Active Passive; WGS: World Geodetic System; LCC: Lambert Conformal Conic . The CNN’s predicted outputs, representing pixel-wise flood probability, were visually and quantitatively compared with observed flood-affected areas in both states. The employed CNN U-net architecture is explained in the following sub-sections. 3.2 U-Net architecture adopted for CNN model The Convolutional Neural Network (CNN) U-Net architecture was chosen due to its strong ability to capture spatial dependencies and multiscale features that are critical for flood susceptibility mapping in heterogeneous and mountainous terrains (Fakhri and Gkanatsios 2025 ). Unlike conventional pixel-based classifiers, U-Net employs an encoder-decoder structure with skip connections, enabling it to retain both global context and fine-grained spatial details. The details are provided in sections below and the representative model structure is shown in Fig. 2 . To further refine the design, hyperparameter tuning was conducted using the Keras Tuner framework with a Random Search strategy. This enabled systematic exploration of filter depth, dropout rates, and learning rate, ensuring that the selected architecture was not only theoretically appropriate but also empirically optimised for the dataset. 3.2.1 Encoder Path The encoder comprises a series of convolutional blocks with 3×3 filters, each followed by Rectified Linear Unit (ReLU) activation and batch normalization to accelerate convergence and prevent internal covariate shifts. Each block is followed by a 2×2 max-pooling layer, which progressively downsamples the feature maps while capturing high-level contextual information. Here, a custom encoder built from multiple Keras Convolution 2 Dimensional + Maximum Pooling 2D layers has been used, designed specifically for patch-wise hazard susceptibility analysis. Hyperparameter search identified an optimal first-layer filter depth of 32 filters, which balanced predictive accuracy and computational efficiency. 3.2.2 Decoder Path The decoder mirrors the encoder, with transposed convolutions (up-convolutions) used to upsample the feature maps. Skip connections between encoder and decoder layers ensure that spatial features lost during pooling are reintroduced, enhancing boundary delineation of flood-prone zones. 3.2.3 Output Layer The final layer uses a 1×1 convolution with a sigmoid activation function, generating pixel-wise flood susceptibility probabilities ranging from 0 to 1. 3.2.4 Training Configuration Loss Function: A custom loss function using a weighted mask with binary cross-entropy (BCE) for binary classification was implemented in TensorFlow. This approach is particularly effective for probability-based predictions. The weighted binary cross-entropy loss calculates the loss using class weights that are inversely proportional to class frequency, which helps to emphasize the minority class (in this case, the flood class) and address class imbalance issues. Optimizer: The Adam optimizer with a tuned learning rate of ~ 0.001 was employed for faster convergence. Data Split: Input datasets were randomly split into 70% for training, 15% for validation, and 15% for testing. Batch Size and Epochs: A batch size of 64 was employed, which offered a practical balance between gradient stability and computational efficiency on CPU. Smaller batch sizes tend to produce noisier gradient updates, while excessively large batches increase memory demand without improving convergence speed on CPU. The maximum number of training epochs was set to 100 to provide sufficient opportunity for the model to converge. However, to avoid unnecessary computation and potential overfitting, Early Stopping was applied with a patience value of 10 (Nemni et al. 2020 ; Tavus et al. 2022 ). This means that training was halted if the validation loss did not improve for ten consecutive epochs, ensuring that the model stopped at an optimal point without wasting resources. Hyperparameters, including filter depth, dropout, and learning rate, were tuned using Random Search with Keras Tuner, and the final configuration was selected based on validation accuracy and computational efficiency. 3.2.5 Evaluation Metrics The performance of the model was evaluated at each stage of the CNN process: training, validation, and testing. We used several metrics to ensure a thorough assessment of its predictive capabilities, including Area Under the Receiver Operating Characteristic Curve (AUC-ROC), Precision, Recall, F1-score, Mean Absolute Error (MAE), Coefficient of determination (R 2 ), Root Mean Square Error (RMSE), and the Critical Success Index (CSI). CNN’s ability to learn non-linear terrain-hydrology interactions allows it to predict susceptibility even in areas occluded by cloud cover or shadowed by steep terrain, where optical or SAR imagery fails. By leveraging DEM-based derivatives and multi-source geospatial layers, the model effectively generalizes flood risk across unobserved zones, offering a significant improvement over threshold-based or traditional ML methods. 3.3 Permutation Feature Importance To understand how much each flood-conditioning factor influenced the CNN predictions, we used the Permutation Feature Importance (PFI) method. In this approach, the values of each factor were shuffled (permuted) across the test set, so that the link between that factor and flooding was broken. We then measured how much the model’s accuracy dropped compared to its performance on the unshuffled data. A large drop meant the factor was important; a small drop meant it contributed less. Finally, the scores were normalized so they added up to 1, giving the relative importance of each factor. 3.4 Sentinel-1 SAR based Flood Extent Mapping To check the accuracy of our proposed CNN model, a flood extent mapping was also performed using Sentinel-1 Synthetic Aperture Radar (SAR) data. For this purpose, Sentinel-1 imagery was accessed and processed through the Google Earth Engine platform to delineate flood-affected areas across Himachal Pradesh and Uttarakhand. The analysis focused on the VH polarization of descending-pass Sentinel-1 Ground Range Detected (GRD) imagery, offering a spatial resolution of 10 meters. Two distinct temporal windows were defined to capture the pre- and post-flood conditions: 1–15 June 2025 and 1–10 July 2025, respectively. Image mosaics corresponding to each period were generated and clipped to the administrative boundary of Himachal Pradesh and Uttarakhand. To reduce speckle noise commonly present in SAR datasets, a Refined Lee filter was applied through a custom function adapted from the ESA SNAP toolbox. This filtering was conducted in natural backscatter units, following conversion from decibel (dB) values, and the output was reconverted to dB post-processing. Flood extent was derived by calculating the pixel-wise ratio between post-flood and pre-flood backscatter intensities. Areas exhibiting a backscatter ratio greater than 1.20 were classified as inundated, consistent with established SAR-based flood detection thresholds. To improve classification accuracy, permanently or seasonally inundated water bodies (seasonality ≥ 5 months) were masked using the JRC Global Surface Water (GSW) dataset. In addition, terrain with slope gradients exceeding 5°, as determined from HydroSHEDS elevation data, was excluded to eliminate potential false positives associated with radar shadowing or terrain-induced distortions. Spurious noise elements, defined as isolated pixel groups smaller than four connected pixels, were also removed to produce a cleaner flood mask. The final flood extent product was exported as a GeoTIFF at 90-meter resolution, enabling integration into subsequent geospatial analyses and visualization workflows. 4. Results The 2025 monsoon season, characterized by intense rainfall and multiple cloudburst events, severely impacted settlements, transportation networks, and infrastructure along river valleys in Himachal Pradesh and Uttarakhand. In Himachal Pradesh, some of the most affected areas include stretches of the Beas River in Mandi district and the Sainj Valley (The Logical Indian 2025 ). In Uttarakhand, significant flooding and landslides were reported near Yamunotri and along the Barkot–Yamunotri road in Uttarkashi, primarily triggered by cloudbursts and flash floods (The Times of India 2025 c). In gradient and hilly terrains, it is difficult to assess this probability due to the challenging landscape. 4.1 Detection of Flooded Hotspots by the Model The CNN-based flood susceptibility model effectively captured these high-risk areas, demonstrating its ability to learn complex hydro-geomorphological relationships. The probability map of flood occurrence ( Fig. 3 a to 3 g ) highlights regions with high to very high susceptibility, aligning with actual flood-affected zones reported during the event. In the Beas River stretch near the Kangra–Hamirpur–Mandi boundary ( Fig. 3 a ) , the model successfully delineated the flooded regions as reported in news articles (The Economic Times 2025 ; The Logical Indian 2025 ), validated by the overlap with Sentinel‑1 SAR-derived water-logged areas. In Mandi ( Fig. 3 b ) and Sainj Valley near Sainj Bridge ( Fig. 3 c ) , the CNN predictions were consistent with observed damage, correctly classified as high susceptible zones. Along the Barkot–Yamunotri road ( Fig. 3 d ) , susceptibility was notably high, reflecting the increased risk to settlements and infrastructure near the Yamunotri pilgrimage route. The Balganga–Budhakedar (Bhilangana block) region also displayed a wide range of high to very high susceptibility levels ( Fig. 3 e ) . In Rudraprayag, the Belni Bridge area exhibited very high susceptibility values ( Fig. 3 f ) . 4.2 Feature Importance of Flood Conditioning Factors The analysis of input features revealed that NDVI, TWI, and altitude are the most critical contributors to flood susceptibility in the study region ( Fig. 5 ) . The high rank of TWI aligns with Dhote et al. ( 2023 ), emphasizing the role of flow accumulation in Himalayan catchments, and has been particularly well-documented in flood susceptibility mapping, as shown by Uca et al. (2022) and Khoirunisa et al. ( 2021 ). LCF ranked 7th among 14 predictors in permutation importance (relative importance = 0.024, compared to 0.296 for NDVI). In practice, LCF improved local delineation accuracy, especially in narrow valleys and interfluve zones, and reduced false positives in flatter depositional reaches, as seen in Fig. 4 and Fig. 7 b. This refinement represents a methodological step forward for CNN-based hazard mapping in data-scarce and geomorphologically complex regions. 4.3 Performance Evaluation of Models on the Test Dataset The flood susceptibility prediction CNN model achieved an overall accuracy of 97.12% with a Critical Success Index (CSI) being 68.44% ( Table 2 ) . This indicates that the model provides reliable event detection performance, especially under conditions of class imbalance, as in the case of hilly terrains, where flood data is scarce, posing a challenge in identifying actual flood impact locations. Other evaluation metrics also confirm the model’s robustness: Root Mean Square Error (RMSE) ~ 0.145, Coefficient of Determination (R²) ~ 0.719, Precision ~ 0.872, Recall ~ 0.761, F1‑Score ~ 0.813, AUC‑ROC ≈ 0.99 ( Table 2 ) . The high AUC-ROC value reflects the model’s strong discriminative ability, even in heterogeneous terrain with limited event data. Notably, the CNN model, trained on multi-factor hazard and local causative factors, overcomes these limitations by learning the underlying geospatial patterns of flood susceptibility, rather than relying solely on real-time flood observations. Figure 6 a to 6 f present the comprehensive performance evaluation of the proposed flood susceptibility CNN U-Net model, illustrating the progression of training loss and accuracy, the ROC curves assessed at different stages of execution, and the testing confusion matrix, collectively validating the robustness and predictive effectiveness of the model. Table 2 Performance evaluation metrics used in the study. Performance Metric Value (approx.) with LCF Value (approx.) without LCF Precision 0.872 0.825 Recall 0.761 0.771 F-1 Score 0.813 0.797 CSI 68.44% 66.32% RMSE 0.145 0.153 R 2 0.719 0.685 MAE 0.038 0.043 AUC 0.988 0.986 Overall Accuracy 97.12% 96.83% 4.4 Lightweight CNN U-Net Performance for Large-Scale Flood Mapping The proposed CNN U-Net demonstrates high efficiency while effectively handling large-scale flood susceptibility mapping. The model operates on a study area covering over 1,00,000 km² (Himachal Pradesh and Uttarakhand), with each input factor raster of ~ 135 MB at 90 m resolution. Despite the large input size, the network remains lightweight, requiring only 5.54 MB of storage for the model weights and architecture. Training on CPU for the full dataset completes in approximately 3 hours, facilitated by a patch-wise approach (16×16 pixels, stride 8) and Zarr-based out-of-core data handling. This compact design ensures computational efficiency, practical deployability in resource-constrained settings, and scalability for operational flood hazard assessment over large regions. Prior U-Net flood works emphasize accuracy and richer inputs but rely on GPU training and larger stacks, for example, Li and Demir ( 2023 ) adjusted U-Net on Sentinel-1 with terrain cues was trained on NVIDIA Titan X GPUs and explored 10–30 m inputs, cross-validation, and transfer learning to boost evaluation, highlighting compute-intensive regimes that our CPU-only pipeline sidesteps via patch-wise optimization and input/output design. While comparative susceptibility studies employing U-Net families report strong accuracies using curated datasets and GPU-oriented training regimes, their workflows emphasize model variants and hyperparameter tuning rather than end-to-end efficiency on modest infrastructure (Melgar-García et al. 2023 ). In contrast, advanced transformer-augmented architectures like TransUNet demonstrate superior segmentation capabilities but with higher parameters count and incur substantial computational costs and GPU memory usage, while fusion approaches combining U-Net with CNNs achieve high accuracies but typically operate on smaller datasets without addressing large-scale geographic mapping challenges or resource-constrained deployment scenarios that our compact architecture explicitly targets (Anand et al. 2023 ; Chen et al. 2024 ). 4.5 Comparison with Global Flood Model Product When compared with the JRC Global River Flood Hazard Map (100-year return period product derived from large‑scale hydrological and hydrodynamic modelling), the proposed CNN framework showed a critical advantage in detecting flood-prone zones beyond the primary floodplains as shown in Fig. 7 a. The JRC model is constrained by river-network hydrodynamics and often misses non-riverine hazard zones such as confined gullies, alluvial cones, and slope channel junctions that frequently flood in Himalayan settings (Nobile et al., 2025 ) The CNN model, by contrast, integrates both hydrological proximity variables (e.g., distance to rivers, drainage density) and topographic land cover features (e.g., NDVI, TWI, LCF, soil moisture), allowing it to capture subtle and highly localised hazard hotspots missed by purely physics‑driven approaches. Most crucially, the model’s susceptibility map closely matched actual 2025 flood locations, including in steep-sided narrow valleys where neither SAR nor hydrodynamic products provided definitive hazard depiction. 5. Discussion 5.1 Comparison with Satellite-Based and Global Flood Hazard products Compared with Sentinel-1 SAR, the proposed CNN U-Net model provided more consistent detection across reported flood zones and avoided false negatives caused by radar limitations. Although most mapped locations, using Sentinel-1 SAR, were indeed reported to have faced severe flooding during the 2025 monsoon, the satellite-based analysis could not capture several of these events ( Fig. 3 b to 3 f ) . This discrepancy is consistent with known limitations of C-band SAR in mountainous environments. The main issues are: (i) the six-day revisit cycle of Sentinel-1 (Sentinel-1 User Handbook 2013), which often misses short-lived flood peaks; (ii) terrain occlusion and shadowing in steep valleys (Shi et al. 2024 ); (iii) disturbance from vegetation and rough water surfaces that scatter radar signals (Chen and Zhao 2022 ; Risling et al. 2024 ); and (iv) additional uncertainties arising from environmental conditions such as rainfall, snowmelt, or landslide activity that further degrade SAR performance (Fakhri and Gkanatsios 2025 ). These constraints explain the under-representation of localized floods and highlight the need for complementary susceptibility mapping approaches. When compared to the JRC Global Flood Hazard maps ( Fig. 7 ) , the model offered finer resolution and more realistic identification of Himalayan hotspots such as Dehradun, Mandi, Kangra, and Shimla. The ability to match actual 2025 monsoon flood flashpoints further validate the model’s suitability for localised disaster risk reduction, where global hazard products often fail to capture small-scale variability (Tripathi and Mohanty 2024 ). 5.2 Contribution of LCF and Operational Relevance of CNN U-Net Although the LCF ranked 7th among 14 predictors in permutation importance (relative importance = 0.024, compared to 0.296 for NDVI), its contribution is disproportionately valuable in rugged mountain terrain. By detecting micro-depressions and subtle convex–concave transitions that concentrate overland flow, LCF captures terrain signatures often underestimated when relying on slope or curvature alone (Liu et al. 2025 ). Figure 7 and Table 2 depict the noteworthy and innovative contribution of this factor in enhancing the spatially informed predictability of the deep learning model. This improvement is reflected in the performance of the CNN U-Net, which, by leveraging dominant predictors such as NDVI, TWI, and altitude alongside LCF, achieved high classification accuracy (97.12%), CSI (68.44%), and low RMSE (0.145). The application of deep learning for flood susceptibility mapping has experienced rapid advancement globally, with numerous studies demonstrating the superiority of CNN-based approaches over traditional statistical and shallow machine learning methods (Ramayanti et al. 2022 ; Ouma and Omai 2023 ; Jamali et al. 2024 ). However, most existing CNN applications in flood mapping have focused primarily on inundation detection using SAR imagery or on susceptibility mapping in plains and deltaic environments, with limited attention to the unique challenges posed by complex mountainous terrain (Li and Demir 2023 ; Fakhri and Gkanatsios 2025 ; Wang and Feng 2025 ). Unlike single-time satellite captures, the CNN U-Net learns underlying spatial patterns of flood susceptibility, which allows identification of hazard-prone areas even under persistent cloud cover, in topographic shadows, or in zones beyond riverine inundation. This performance underscores the model’s advantage in mountainous regions where hydro-geomorphological conditioning dominates flood behaviour. Its operational practicality is equally noteworthy and innovative, as the lightweight architecture of the CNN-U-Net, combined with easily retrainable inputs, positions the framework as highly operational. Once updated predictor datasets are available, the model can be retrained and deployed within hours, allowing susceptibility updates at scales useful for district and state disaster management agencies. Such rapid workflows are particularly relevant for supporting early warning systems, pre-monsoon preparedness, climate-resilient infrastructure planning, and land-use regulation. In contrast, hydrodynamic models often require extensive data, computational resources, and time, making them less practical for rapid decision-making in dynamic Himalayan environments (Karim et al. 2023 ). 5.3 Limitations and Future Directions Despite its strong performance, several limitations remain. The model’s accuracy is partly dependent on the quality of the flood inventory used for training, which may contain model biases or uncertainties. While LCF improves representation of topography, the current 90 metre resolution may not fully capture very localised floods. Transferability to other physiographic regions, such as plains or deltaic settings, requires further testing. Future work should focus on: Event-based forecasting by integrating dynamic rainfall and discharge inputs. Benchmarking against multiple AI models and CNN architectures to evaluate robustness and identify the optimal framework for flood susceptibility mapping in IHR. Multi-hazard extension, linking flood susceptibility with co-occurring hazards such as landslides and GLOFs. Transfer learning approaches to reduce retraining needs across regions. Higher resolution datasets (30 m or LiDAR-based DEMs) for local planning. Coupling susceptibility with exposure and vulnerability layers to produce holistic risk assessments. The study also highlights broader implications for hazard research and practice. The CNN U-Net segmentation framework, augmented with geomorphologically relevant predictors such as the LCF, helps bridge the gap between coarse global flood models and data-intensive hydrodynamic simulations. By generating susceptibility maps that are both accurate and operationally feasible, the framework supports disaster risk reduction efforts under the Sendai Framework and responds to the growing challenge of extreme rainfall and flooding in a changing Himalayan climate. These maps can directly inform early warning systems, land-use planning, and risk mitigation strategies. Moreover, the approach holds considerable potential for advancing multi-hazard susceptibility mapping in other mountainous regions worldwide. 6. Conclusion This study demonstrates the potential of DL-based flood susceptibility mapping in the complex and data-scarce terrains of the Indian Himalayan Region, with a specific focus on Himachal Pradesh and Uttarakhand. Motivated by the devastating June–August 2025 monsoon-induced hydro-meteorological events, the study underscores the high flood susceptibility of low-lying valleys and riverine corridors, where settlements and infrastructure are exposed to floods, cloudbursts, and cascading hazards. By employing a CNN with a U-Net architecture and integrating 14 hydro-geomorphological conditioning factors, the approach successfully produced spatially explicit, high-resolution flood susceptibility maps that showed strong alignment with observed flood impacts. Although LCF ranked mid‑tier in overall importance (0.024 relative score), it proved critical for refining hazard delineation in rugged terrain and reducing overestimation in depositional areas, representing a methodological innovation in data‑driven flood risk modelling. Importantly, the CNN framework overcomes limitations inherent to observation‑driven methods such as Sentinel‑1 SAR rapid mapping, which in Himalayan contexts suffers from infrequent revisit intervals, terrain occlusion, dense vegetation interference, and rapidly changing flood hydrodynamics. This makes the approach inherently operational, supporting early warning systems, land‑use regulation, climate‑resilient infrastructure planning, and pre‑monsoon preparedness. Beyond flood risk applications, the proposed framework holds promise as a scalable multi‑hazard susceptibility mapping platform for other hydro‑meteorological and geomorphic hazards in mountainous regions. Future progress will hinge on integrating higher‑quality ground observations and more systematic disaster reporting to further strengthen model reliability. In summary, this research underscores the importance of pattern‑informed, stakeholder‑ready flood susceptibility modelling for the IHR, reducing reliance on temporally constrained remote sensing, bridging gaps in traditional hydrodynamic hazard assessments, and enabling precise, high‑resolution risk mapping that is vital for disaster preparedness and long‑term climate resilience. By enabling faster and more accurate susceptibility mapping, the CNN U-Net approach offers a practical decision-support tool for mitigating flood risk in one of the world’s most disaster-prone mountain systems. Declarations Acknowledgements: The authors gratefully acknowledge the ESA Copernicus Hub for Sentinel-1A/B data, IMD for rainfall data, and JRC for providing the river flood hazard map and surface water information. The work presented here is supported by the Department of Science and Technology (Integrated Centre for Adaptation to Climate Change, Disaster Risk Reduction and Sustainability (ICARS)-IIT Roorkee); Project No. DST-2211-WRC/23-24 sanctioned by DST (OM No DST/CCP/NMSKCC/CoE/236/2024(G)] and Indian Space Research Organisation (Project No. STC-2278-WRC-CNA/23-24). We also thank the Indian Institute of Technology Roorkee (IIT Roorkee) and the Department of Water Resources Development and Management (WRD&M)for computational support, especially the PARAM Ganga high-performance computing facility. Declaration of interests: The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper. Author contributions: Rachit: Conceptualization, Methodology, Data Collection, Data Analysis, Writing—original draft, Writing—review and editing; Vaibhav Tripathi: Conceptualization, Methodology, Data Analysis, Writing—review and editing; Mohit Prakash Mohanty: Conceptualization, Methodology, Writing—review and editing; Supervision; Ashish Pandey: Supervision, Project Administration; Anil Kumar Gupta: Supervision, Project Administration. References Alam MG, Tripathi V, Bhatt CM, Mohanty MP (2025) A Novel Framework Embedding Bayesian-optimized Ensemble Machine Learning and Explainable Artificial Intelligence (XAI) to Improve Flood Prediction in Complex Watersheds. Environ Sustain Indicat 100760. https://doi.org/10.1016/j.indic.2025.100760 Amiri A, Soltani K, Ebtehaj I, Bonakdari H (2024) A novel machine learning tool for current and future flood susceptibility mapping by integrating remote sensing and geographic information systems. J Hydrol 632:130936. https://doi.org/10.1016/j.jhydrol.2024.130936 Anand V, Gupta S, Koundal D, Singh K (2023) Fusion of U-Net and CNN model for segmentation and classification of skin lesion from dermoscopy images. Expert Syst Appl 213:119230. https://doi.org/10.1016/j.eswa.2022.119230 Baugh C, Colonese J, D'Angelo C, Dottori F, Neal J, Prudhomme C, Salamon P (2024) Global river flood hazard maps. European Commission, Joint Research Centre (JRC). http://data.europa.eu/89h/jrc-floods-floodmapgl_rp50y-tif Bentivoglio R, Isufi E, Jonkman SN, Taormina R (2022) Deep learning methods for flood mapping: A review of existing applications and future research directions. Hydrol Earth Syst Sci 26(16):4345–4378. https://doi.org/10.5194/hess-26-4345-2022 Biazar SM, Golmohammadi G, Nedhunuri RR, Shaghaghi S, Mohammadi K (2025) Artificial Intelligence in Hydrology: Advancements in Soil, Water Resource Management, and Sustainable Development. Sustainability 17(5):2250. https://doi.org/10.3390/su17052250 Breznik J, Oštir K, Rak G (2025) The potential of Sentinel-1 imagery for flood event detection: A satellite vs. hydraulic model comparison. J Hydrol 651:132587. https://doi.org/10.1016/j.jhydrol.2024.132587 CEEW (2021) 85% Districts in Uttarakhand Vulnerable to Extreme Floods. https://www.ceew.in/press-releases/85-districts-uttarakhand-vulnerable-extreme-floods-ceew. Accessed 18 Aug 2025 Chen J, Mei J, Li X, Lu Y, Yu Q, Wei Q, Lu X, Xie Y, Adeli E, Wang Y, Lungren MP, Zhang S, Xing L, Lu L, Yuille A, Zhou Y (2024) TransUNet: Rethinking the U-Net architecture design for medical image segmentation through the lens of transformers. Med Image Anal 97:103280. https://doi.org/10.1016/j.media.2024.103280 Chen Z, Zhao S (2022) Automatic monitoring of surface water dynamics using Sentinel-1 and Sentinel-2 data with Google Earth Engine. Int J Appl Earth Obs Geoinf 113:103010. https://doi.org/10.1016/j.jag.2022.103010 Dey H, Haque MM, Shao W, VanDyke M, Hao F (2024) Simulating flood risk in Tampa Bay using a machine learning driven approach. npj Nat Hazards 1(1):40. https://doi.org/10.1038/s44304-024-00045-4 Dhote PR, Joshi Y, Rajib A, Thakur PK, Nikam BR, Aggarwal SP (2023) Evaluating topography-based approaches for fast floodplain mapping in data-scarce complex-terrain regions: Findings from a Himalayan basin. J Hydrol 620: 129309. https://doi.org/10.1016/j.jhydrol.2023.129309 Down To Earth (2025) With two months of monsoon left, will Himachal create new records of natural disasters? https://www.downtoearth.org.in/natural-disasters/with-two-months-of-monsoon-left-will-himachal-create-new-records-of-natural-disasters. Accessed 18 Aug 2025 Edamo ML, Ayele EG, Yisihak Ukumo T, Alemayehu Kassaye A, Paulos Haile A (2024) Capability of logistic regression in identifying flood-susceptible areas in a small watershed. H2Open J 7 (5):351–374. https://doi.org/10.2166/h2oj.2024.024 Fakhri F, Gkanatsios I (2025) Quantitative evaluation of flood extent detection using attention U-Net case studies from Eastern South Wales Australia in March 2021 and July 2022. Sci Rep 15(1):12377. https://doi.org/10.1038/s41598-025-92734-x Fereshtehpour M, Esmaeilzadeh M, Alipour RS, Burian SJ (2024) Impacts of DEM type and resolution on deep learning-based flood inundation mapping. Earth Sci Inform 17(2):1125–1145. https://doi.org/10.1007/s12145-024-01239-0 Gupta V, Syed B, Pathania A, Raaj S, Nanda A, Awasthi S, Shukla DP (2024) Hydrometeorological analysis of July-2023 floods in Himachal Pradesh, India. Nat Hazards 120(8):7549–7574. https://doi.org/10.1007/s11069-024-06520-5 India Today (2024) Flash flood risk in Himachal, Uttarakhand as glacial lakes in Himalayas rise: Report. https://www.indiatoday.in/environment/story/flash-flood-risk-himachal-uttarakhand-glacial-lakes-himalayas-rise-2601325-2024-09-17 Accessed 30 Aug 2025 India Today (2025) Nature’s fury in numbers: A look at Uttarakhand’s disaster record. https://www.indiatoday.in/diu/story/natures-fury-in-numbers-a-look-at-uttarakhand-disaster-record-2769669-2025-08-11. Accessed 18 Aug 2025 Jamali A, Roy SK, Beni LH, Pradhan B, Li J, Ghamisi P (2024) Residual wave vision U-Net for flood mapping using dual polarization Sentinel-1 SAR imagery. Int J Appl Earth Obs Geoinf 127:103662. https://doi.org/10.1016/j.jag.2024.103662 Kansal ML, Singh S (2022) Flood Management Issues in Hilly Regions of Uttarakhand (India) under Changing Climatic Conditions. Water 14(12):1879. https://doi.org/10.3390/w14121879 Karim F, Armin MA, Ahmedt-Aristizabal D, Tychsen-Smith L, Petersson L (2023) A Review of Hydrodynamic and Machine Learning Approaches for Flood Inundation Modeling. Water 15(3):566. https://doi.org/10.3390/w15030566 Khoirunisa N, Ku CY, Liu CY (2021) A GIS-Based Artificial Neural Network Model for Flood Susceptibility Assessment. Int J Environ Res Public Health 18(3):1072. https://doi.org/10.3390/ijerph18031072 Khosravi K, Rezaie F, Cooper JR, Kalantari Z, Abolfathi S, Hatamiafkoueieh J (2023) Soil water erosion susceptibility assessment using deep learning algorithms. J Hydrol 618:129229 https://doi.org/10.1016/j.jhydrol.2023.129229 Konapala G, Kumar SV, Khalique Ahmad S (2021) Exploring Sentinel-1 and Sentinel-2 diversity for flood inundation mapping using deep learning. ISPRS J Photogramm Remote Sens 180:163–173. https://doi.org/10.1016/j.isprsjprs.2021.08.016 Kumar V, Solanki YP, Sharma KV, Patel A, Tiwari DK, Mehta DJ (2024) India’s flood risk assessment and mapping with multi-criteria decision analysis and GIS integration. J Water Clim Change. https://doi.org/10.2166/wcc.2024.054 Li Z, Demir I (2023). U-net-based semantic classification for flood extent extraction using SAR imagery and GEE platform: A case study for 2019 central US flooding. Sci Total Environ 869:161757. https://doi.org/10.1016/j.scitotenv.2023.161757 Liu J, Zhao X, Chen Y, Sun H, Gu Y, Xu S (2025) A novel flood conditioning factor based on topography for flood susceptibility modeling. Geosci Front 16(1):101960. https://doi.org/10.1016/j.gsf.2024.101960 Melgar-García L, Martínez-Álvarez F, Bui DT, Troncoso A (2023) A novel semantic segmentation approach based on U-Net, WU-Net, and U-Net++ deep learning for predicting areas sensitive to pluvial flood at tropical area. International Journal of Digital Earth. https://doi.org/10.1080/17538947.2023.2252401 Merz B, Blöschl G, Vorogushyn S, Dottori F, Aerts JCJH, Bates P, Bertola M, Kemter M, Kreibich H, Lall U, Macdonald E (2021) Causes, impacts and patterns of disastrous river floods. Nat Rev Earth Environ 2(9):592–609. https://doi.org/10.1038/s43017-021-00195-3 Nagamani K, Mishra AK, Meer MS, Das J (2024) Understanding flash flooding in the Himalayan Region: A case study. Sci Rep 14(1):7060. https://doi.org/10.1038/s41598-024-53535-w Nemni E, Bullock J, Belabbes S, Bromley L (2020) Fully Convolutional Neural Network for Rapid Flood Segmentation in Synthetic Aperture Radar Imagery. Remote Sens 12(16):2532. https://doi.org/10.3390/rs12162532 Nobile EGL, Figueiredo R, Arrighi C, Romão X, Martina MLV (2025) Flood risk assessment of cultural heritage across countries and spatial scales. Int J Disaster Risk Reduct 118:105236. https://doi.org/10.1016/j.ijdrr.2025.105236 Ombadi M, Risser MD, Rhoades AM, Varadharajan C (2023) A warming-induced reduction in snow fraction amplifies rainfall extremes. Nature 619(7969):305–310. https://doi.org/10.1038/s41586-023-06092-7 Ouma YO, Omai L (2023) Flood Susceptibility Mapping Using Image-Based 2D-CNN Deep Learning: Overview and Case Study Application Using Multiparametric Spatial Data in Data-Scarce Urban Environments. Int J Intell Syst 2023(1):5672401. https://doi.org/10.1155/2023/5672401 Pareta K (2024) 1D-2D hydrodynamic and sediment transport modelling using MIKE models. Discover Water 4(1):94. https://doi.org/10.1007/s43832-024-00130-9 Patel P, Thakur PK, Aggarwal SP, Garg V, Dhote PR, Nikam BR, Swain S, Al-Ansari N (2022) Revisiting 2013 Uttarakhand flash floods through hydrological evaluation of precipitation data sources and morphometric prioritization. Geomat Nat Hazards Risk 13(1):646-666. https://doi.org/10.1080/19475705.2022.2038696 Pepin NC, Arnone E, Gobiet A, Haslinger K, Kotlarski S, Notarnicola C, Palazzi E, Seibert P, Serafin S, Schöner W, Terzago S, Thornton JM, Vuille M, Adler C (2022) Climate Changes and Their Elevational Patterns in the Mountains of the World. Rev Geophys 60(1):e2020RG000730. https://doi.org/10.1029/2020RG000730 PreventionWeb (2025) Over 45% of Himachal Pradesh at risk of landslides, floods. https://www.preventionweb.net/news/over-45-himachal-pradesh-risk-landslides-floods. Accessed 18 Aug 2025 Ramayanti S, Nur AS, Syifa M, Panahi M, Achmad AR, Park S, Lee, CW (2022) Performance comparison of two deep learning models for flood susceptibility map in Beira area, Mozambique. Egypt J Remote Sens Space Sci 25(4):1025–1036. https://doi.org/10.1016/j.ejrs.2022.11.003 Rick B, McGrath D, McCoy SW, Armstrong WH (2023) Unchanged frequency and decreasing magnitude of outbursts from ice-dammed lakes in Alaska. Nature Commun 14(1):6138. https://doi.org/10.1038/s41467-023-41794-6 Risling A, Lindersson S, Brandimarte L (2024) A comparison of global flood models using Sentinel-1 and a change detection approach. Nat Hazards 120(12):11133–11152. https://doi.org/10.1007/s11069-024-06629-7 Roul AR, Behera AK, Pradhan RM, Pradhan SP (2025) Navigating Hydrological Factors in Understanding and Mitigating Landslide Hazards in the Indian Himalayas: A Short Review. Landslides: Analysis, Modeling and Mitigation 3–24. https://doi.org/10.1007/978-3-031-78896-3_1 Safaei-Moghadam A, Tarboton D, Minsker B (2023) Estimating the likelihood of roadway pluvial flood based on crowdsourced traffic data and depression-based DEM analysis. Nat Hazards Earth Syst Sci 23(1):1–19. https://doi.org/10.5194/nhess-23-1-2023 Saravanan S, Abijith D, Reddy NM, Kss P, Janardhanam N, Sathiyamurthi S, Sivakumar V (2023) Flood susceptibility mapping using machine learning boosting algorithms techniques in Idukki district of Kerala India. Urban Clim 49:101503. https://doi.org/10.1016/j.uclim.2023.101503 Sajwan KS, Sushil K (2016) A Geological Appraisal of Slope Instability in Upper Alaknanda Valley, Uttarakhand Himalaya, India. J Geol Geophys 5:1-7. https://doi.org/10.4172/2381-8719.1000258 Schlögl M, Fuchs S, Scheidl C, Heiser M (2021) Trends in torrential flooding in the Austrian Alps: A combination of climate change, exposure dynamics, and mitigation measures. Clim Risk Manag 32:100294. https://doi.org/10.1016/j.crm.2021.100294 Sentinel-1 Team (2013) Sentinel-1 User Handbook. https://ftp.itc.nl/pub/Dragon4_Lecturer_2018/ESA%20EEs%20and%20Sentinels%20Brochures%20pdf/Copernicus%20Sentinels%201,2,3/Sentinel-1%20User%20Handbook.pdf. Accessed 17 Aug 2025 Shah A, Malakar K (2024) Climate-change-induced risk mapping of the Indian Himalayan districts using the latest IPCC framework. Int J Disaster Risk Reduct 102:104283. https://doi.org/10.1016/j.ijdrr.2024.104283 Shi C, Zuo X, Zhang J, Zhu D, Li Y, Bu J (2024) Accuracy Assessment of Geometric-Distortion Identification Methods for Sentinel-1 Synthetic Aperture Radar Imagery in Highland Mountainous Regions. Sensors (Basel, Switzerland) 24(9):2834. https://doi.org/10.3390/s24092834 Song W, Guan M, Guo K, Yu D (2025) Rapid flood inundation mapping by integrating deep learning-based image super-resolution with coarse-grid hydrodynamic modeling. Eng Appl Comput Fluid Mech 19(1):2481115. https://doi.org/10.1080/19942060.2025.2481115 Sphere India (2023) Situation Assessment Report Himachal Pradesh Floods Jul 2023. https://www.sphereindia.org.in/sites/default/files/2023-10/Situation%20Report_Himachal%20Pradesh_Floods_Jul23.pdf. Accessed 17 Aug 2025 Sreevalsan-Nair J, Mundayatt A (2025) Evolution of Data-driven Single- and Multi-Hazard Susceptibility Mapping and Emergence of Deep Learning Methods (No. arXiv:2502.09045). arXiv. https://doi.org/10.48550/arXiv.2502.09045 Sultan H, Zhan J, Rashid W, Chu X, Bohnett E (2022) Systematic Review of Multi-Dimensional Vulnerabilities in the Himalayas. Int J Environ Res Public Health 19(19):12177. https://doi.org/10.3390/ijerph191912177 Tarpanelli A, Mondini AC, Camici S (2022) Effectiveness of Sentinel-1 and Sentinel-2 for flood detection assessment in Europe. Nat Hazards Earth Syst Sci 22(8):2473–2489. https://doi.org/10.5194/nhess-22-2473-2022 Tavus B, Can, R, Kocaman S (2022) A cnn-based flood mapping approach using Sentinel-1 data. ISPRS Ann Photogramm Remote Sens Spatial Inf Sci 3:549-556. https://doi.org/10.5194/isprs-annals-V-3-2022-549-2022 Taysi H, Yang YCE, Gangrade S, Chegini T, Kao SC, Li HY (2025) Enhancing 2D hydrodynamic flood models through machine learning and urban drainage integration. J Hydrol 659:133258. https://doi.org/10.1016/j.jhydrol.2025.133258 The Economic Times (2025) Severe floods experienced in Beas river following heavy rainfall in Himachal Pradesh’s Mandi.https://economictimes.indiatimes.com/news/india/severe-floods-experienced-in-beas-river-following-heavy-rainfall-in-himachal-pradeshs-mandi/articleshow/122196209.cms?from=mdr. Accessed 17 Aug 2025 The Hindu(2025) Uttarkashi flash floods highlights: Rescuers retrieve two bodies in flood-hit Dharali; at least 190 rescued. https://www.thehindu.com/news/national/uttarakhand/uttarakhand-uttarkashi-cloudburst-flash-floods-highlights-august-6-2025/article69896766.ece. Accessed 17 Aug 2025 The Logical Indian(2025) Himachal Pradesh: Cloudburst in Kullu’s Sainj Valley Sparks Flash Floods, Cuts Off Villages - No Casualties So Far. https://thelogicalindian.com/himachal-pradesh-cloudburst-in-kullus-sainj-valley-sparks-flash-floods-cuts-off-villages-no-casualties-so-far/. Accessed 17 Aug 2025 The Times of India (2025a) 705 deaths in 10 years: Multiple studies warn of flash floods emerging as major killer in Uttarakhand; climate change, extreme weather rewriting Himalayan hydrology. https://timesofindia.indiatimes.com/city/dehradun/705-deaths-in-10-yrs-multiple-studies-warn-of-flash-floods-emerging-as-major-killer-in-ukhand/articleshow/123206705.cms. Accessed 18 Aug 2025 The Times of India(2025b) Cloudburst, heavy rains wreak havoc in Uttarakhand: Char Dham yatra halted, workers missing, landslides hit highways – 10 key updates. https://timesofindia.indiatimes.com/india/cloudburst-heavy-rains-wreak-havoc-in-uttarakhand-char-dham-yatra-halted-workers-missing-landslides-hit-highways-10-key-updates/articleshow/122138901.cms. Accessed 17 Aug 2025 The Times of India (2025c) Uttarkashi cloudburst havoc: At least 4 dead as flash floods swallow homes, hotels in Dharali; entire village washed away. https://timesofindia.indiatimes.com/city/dehradun/uttarakhand-cloudburst-flash-floods-hit-uttarkashi-several-villagers-washed-away/articleshow/123115132.cms. Accessed 17 Aug 2025 The Times of India (2024) Natural disasters killed 358 across HP this year. https://timesofindia.indiatimes.com/city/shimla/himachal-pradesh-faces-catastrophic-year-with-358-lives-lost-to-natural-disasters/articleshow/115950309.cms. Accessed 18 Aug 2025 Tripathi V, Mohanty MP, Singh H (2023) Fidelity of machine learning models in capturing flood inundation through geomorphic descriptors over Ganga sub‑basin, India. 12th World Con‑gress on Water Resources and Environment (EWRA 2023), Thessaloniki, Greece. Tripathi V, Mohanty MP (2024) Can geomorphic flood descriptors coupled with machine learning models enhance in quantifying flood risks over data-scarce catchments? Development of a hybrid framework for Ganga basin (India). Environ. Sci Pollut Res 1-20. https://doi.org/10.1007/s11356-024-33507-3 Uca, Lamada M, Nyompa S, Haris NA (2022) Morphometry and Topographic Wetness Index Analysis for flood inundation mapping in Mata Allo watershed (South Sulawesi, Indonesia). Forum Geogr XXI(2):133–145. https://doi.org/10.5775/fg.2022.105.d Uddin K, Matin MA, Meyer FJ (2019) Operational Flood Mapping Using Multi-Temporal Sentinel-1 SAR Images: A Case Study from Bangladesh. Remote Sens 11(13):1581. https://doi.org/10.3390/rs11131581 Ullah K, Wang Y, Fang Z, Wang L, Rahman M (2022) Multi-hazard susceptibility mapping based on Convolutional Neural Networks. Geosci Front 13(5):101425. https://doi.org/10.1016/j.gsf.2022.101425 UN Office for Disaster Risk Reduction (UNDRR), Centre for Research on the Epidemiology of Disasters (CRED)(2022) "Human cost of disasters - An overview of the last 20 years 2000-2019". https://reliefweb.int/sites/reliefweb.int/files/resources/Human%20Cost%20of%20Disasters%202000-2019%20Report%20-%20UN%20Office%20for%20Disaster%20Risk%20Reduction.pdf. Accessed 18 Aug 2025 Wang F, Feng X (2025) Flood change detection model based on an improved U-net network and multi-head attention mechanism. Sci Rep 15(1):3295. https://doi.org/10.1038/s41598-025-87851-6 Wang Y, Fang Z, Hong H, Peng L (2020) Flood susceptibility mapping using convolutional neural network frameworks. J Hydrol 582:124482. https://doi.org/10.1016/j.jhydrol.2019.124482 Zhao G, Pang B, Xu Z, Yue J, Tu T (2018) Mapping flood susceptibility in mountainous areas on a national scale in China. Sci Total Environ 615:1133–1142. https://doi.org/10.1016/j.scitotenv.2017.10.037 Cite Share Download PDF Status: Under Review Version 1 posted Editorial decision: Major revisions 22 Feb, 2026 Reviewers agreed at journal 15 Oct, 2025 Reviewers invited by journal 14 Oct, 2025 Editor assigned by journal 29 Sep, 2025 First submitted to journal 25 Sep, 2025 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-7715126","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":529733607,"identity":"d054c788-eeda-4c2d-9523-15b895072055","order_by":0,"name":"Rachit .","email":"","orcid":"","institution":"Indian Institute of Technology Roorkee","correspondingAuthor":false,"prefix":"","firstName":"Rachit","middleName":"","lastName":".","suffix":""},{"id":529733608,"identity":"846fda49-00dc-4e61-908f-ebf9f47c6fc7","order_by":1,"name":"Vaibhav Tripathi","email":"","orcid":"","institution":"Indian Institute of Technology Roorkee","correspondingAuthor":false,"prefix":"","firstName":"Vaibhav","middleName":"","lastName":"Tripathi","suffix":""},{"id":529733609,"identity":"cc8201d3-25b8-457e-8d27-1afef2d6d0cf","order_by":2,"name":"Mohit Prakash Mohanty","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAABBklEQVRIiWNgGAWjYFACHoYDIMoAzKkAYmbGBogMMx4tByBagErPgLU0NhDSwgDXwtgGFoJZgx3otvcePPyBYZucOfvZ44955x3OZ2Bnbn9cwWAnz8DOewCbFrMz5xKADrttbNmTl9jMu+2wZQPQYY1nGJING5j5ErBquZFjANKSuOFAjiFQy20DkF+AnmFOYGDmMcCv5fwboJY5cC31RGi5AbKlAa7lMG4tZ84YHDhjcNvY4MYbw5lzjv03YANqmdlgcNywDZeW4z3GHyoqbssZnM8x+PCmJs2An//4g48NFdXy/PxnsGqBAKgUEzCOGNhgImy41SMA4w9iVI2CUTAKRsGIAwCznGEncVup2AAAAABJRU5ErkJggg==","orcid":"https://orcid.org/0009-0009-0060-3227","institution":"Indian Institute of Technology Roorkee","correspondingAuthor":true,"prefix":"","firstName":"Mohit","middleName":"Prakash","lastName":"Mohanty","suffix":""},{"id":529733610,"identity":"9f3a9eef-d5a2-4242-834a-6ba2b2a754e4","order_by":3,"name":"Ashish Pandey","email":"","orcid":"","institution":"Indian Institute of Technology Roorkee","correspondingAuthor":false,"prefix":"","firstName":"Ashish","middleName":"","lastName":"Pandey","suffix":""},{"id":529733611,"identity":"890b948c-6cf4-40fc-b26a-dc674fb222ba","order_by":4,"name":"Anil Kumar Gupta","email":"","orcid":"","institution":"Indian Institute of Technology Roorkee","correspondingAuthor":false,"prefix":"","firstName":"Anil","middleName":"Kumar","lastName":"Gupta","suffix":""}],"badges":[],"createdAt":"2025-09-25 16:54:41","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-7715126/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-7715126/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":94640553,"identity":"653aa583-66c2-465c-b95b-c2551b734820","added_by":"auto","created_at":"2025-10-29 07:49:50","extension":"xml","order_by":2,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":11001,"visible":true,"origin":"","legend":"","description":"","filename":"nhazNHAZD2503301.xml","url":"https://assets-eu.researchsquare.com/files/rs-7715126/v1/a3f11d48624e6ea1f8bce942.xml"},{"id":94637908,"identity":"c6fc44d0-db0f-4ba7-85fc-587261219a3d","added_by":"auto","created_at":"2025-10-29 07:26:05","extension":"xml","order_by":3,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1007,"visible":true,"origin":"","legend":"","description":"","filename":"NHAZD250330154179.go.xml","url":"https://assets-eu.researchsquare.com/files/rs-7715126/v1/f2a2d546ba35dac24474213e.xml"},{"id":94640664,"identity":"a2e1f927-97e2-48cd-ad11-d32aa40082a9","added_by":"auto","created_at":"2025-10-29 07:50:03","extension":"xml","order_by":4,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":890,"visible":true,"origin":"","legend":"","description":"","filename":"NHAZD2503301Import.xml","url":"https://assets-eu.researchsquare.com/files/rs-7715126/v1/64c05bff5c3fccb2a9d4de6e.xml"},{"id":94640179,"identity":"c3564f10-bbd4-4942-99d5-359c1f8572f6","added_by":"auto","created_at":"2025-10-29 07:48:38","extension":"xml","order_by":5,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":185487,"visible":true,"origin":"","legend":"","description":"","filename":"NHAZD25033010enriched.xml","url":"https://assets-eu.researchsquare.com/files/rs-7715126/v1/1892c4fc177882486da6a27c.xml"},{"id":94640131,"identity":"cbcbc2d7-7e58-4abc-8d66-a2ccb96dcdbd","added_by":"auto","created_at":"2025-10-29 07:48:27","extension":"png","order_by":6,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":872311,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-7715126/v1/49e3ac88e87676b09364447a.png"},{"id":94637916,"identity":"86df5b58-74b2-4afe-87fd-cf558781996f","added_by":"auto","created_at":"2025-10-29 07:26:05","extension":"png","order_by":7,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":2908392,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-7715126/v1/4a109baf5d3cb838d38daf6d.png"},{"id":94640683,"identity":"9490a2a2-a44c-416d-9367-12dc1b2e9f97","added_by":"auto","created_at":"2025-10-29 07:50:04","extension":"jpeg","order_by":8,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1982881,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage3.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7715126/v1/f6a0a158e1d16af7bcba07b2.jpeg"},{"id":94640405,"identity":"ed7d5dfc-e2e0-49d9-a59b-e204b65eddcb","added_by":"auto","created_at":"2025-10-29 07:49:23","extension":"jpeg","order_by":9,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1006478,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage4.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7715126/v1/b51c2657b8346384adff226a.jpeg"},{"id":94640643,"identity":"292ab0bb-1cdd-4cb6-97d5-8a2764cfcbd1","added_by":"auto","created_at":"2025-10-29 07:50:00","extension":"png","order_by":10,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":32793,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-7715126/v1/c5366ff170659f3fbbb70302.png"},{"id":94640945,"identity":"e64d2ecb-cde1-4170-ab59-0086afd841ae","added_by":"auto","created_at":"2025-10-29 07:50:24","extension":"png","order_by":11,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":236454,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage6.png","url":"https://assets-eu.researchsquare.com/files/rs-7715126/v1/70594346626ed34a76f44ad5.png"},{"id":94637928,"identity":"e8e9600c-dd9b-4efc-8ac2-f3c2ebe7c60f","added_by":"auto","created_at":"2025-10-29 07:26:05","extension":"jpeg","order_by":12,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1563534,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage7.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7715126/v1/b1bc591b8103b3e40367308b.jpeg"},{"id":94640428,"identity":"a903615a-da0a-435e-94be-6dc6ab39be6b","added_by":"auto","created_at":"2025-10-29 07:49:28","extension":"png","order_by":13,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":189076,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-7715126/v1/0ee9c2bb7e33c009bc444cdf.png"},{"id":94637925,"identity":"d6f0aea7-ea31-4866-bbc3-2c5c6168f115","added_by":"auto","created_at":"2025-10-29 07:26:05","extension":"png","order_by":14,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":477068,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-7715126/v1/032d9ea119b21c6b46163275.png"},{"id":94637917,"identity":"fe7f71bd-7433-4cfd-a64d-00e7f13b7431","added_by":"auto","created_at":"2025-10-29 07:26:05","extension":"png","order_by":15,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":860026,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-7715126/v1/71d3bb07a2c416111d3035e3.png"},{"id":94637919,"identity":"95bf3f61-ec79-4500-bc02-3c76e8da973d","added_by":"auto","created_at":"2025-10-29 07:26:05","extension":"png","order_by":16,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":526296,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-7715126/v1/a5d34e72f7a5df7dc9529f30.png"},{"id":94640375,"identity":"720c67f6-d059-49a8-8c55-188b156ca2e9","added_by":"auto","created_at":"2025-10-29 07:49:19","extension":"png","order_by":17,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":29848,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-7715126/v1/73e0f05cb84295a68c9c2616.png"},{"id":94637927,"identity":"a8ce8a06-b35f-4a21-b9d3-8c5ecc29df4b","added_by":"auto","created_at":"2025-10-29 07:26:05","extension":"png","order_by":18,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":54222,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage6.png","url":"https://assets-eu.researchsquare.com/files/rs-7715126/v1/1aa7ac4bb1c73106b79ebff5.png"},{"id":94640489,"identity":"39477be3-0a9b-4e72-8a4d-1940f74dd2d6","added_by":"auto","created_at":"2025-10-29 07:49:41","extension":"png","order_by":19,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":383879,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage7.png","url":"https://assets-eu.researchsquare.com/files/rs-7715126/v1/9db8a4275ca470e5db38896a.png"},{"id":94637929,"identity":"d224d42e-8674-4d56-8595-57046e235523","added_by":"auto","created_at":"2025-10-29 07:26:05","extension":"xml","order_by":20,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":184500,"visible":true,"origin":"","legend":"","description":"","filename":"NHAZD25033010structuring.xml","url":"https://assets-eu.researchsquare.com/files/rs-7715126/v1/1015bac13a17c21e8456f6ea.xml"},{"id":94637931,"identity":"defdc45c-0aba-4622-a4f8-18eda2898163","added_by":"auto","created_at":"2025-10-29 07:26:06","extension":"html","order_by":21,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":194195,"visible":true,"origin":"","legend":"","description":"","filename":"earlyproof.html","url":"https://assets-eu.researchsquare.com/files/rs-7715126/v1/7b515145bc040355a696247d.html"},{"id":94640825,"identity":"5f57aad9-967d-4594-8a6c-55d22fe85d29","added_by":"auto","created_at":"2025-10-29 07:50:17","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":436059,"visible":true,"origin":"","legend":"\u003cp\u003eStudy area map representing the geographical location of Himachal Pradesh and Uttarakhand in India, along with the elevationcharacteristics of the region.\u003c/p\u003e","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-7715126/v1/ee59ecb227d58b28c5b2c728.png"},{"id":94640862,"identity":"8c7623a4-23a2-40ab-8a7d-c18295aa65fd","added_by":"auto","created_at":"2025-10-29 07:50:19","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":1669504,"visible":true,"origin":"","legend":"\u003cp\u003eConceptual framework for the CNN U-net-based flood susceptibility prediction\u003c/p\u003e","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-7715126/v1/570977ee9a1844f699aa88d7.png"},{"id":94640955,"identity":"bf546e05-75d4-48fb-ae59-2e1cfd774918","added_by":"auto","created_at":"2025-10-29 07:50:24","extension":"jpeg","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":1982881,"visible":true,"origin":"","legend":"\u003cp\u003eFlood conditioning factors maps used for flood susceptibility mapping \u003cstrong\u003e(a)\u003c/strong\u003eslope, \u003cstrong\u003e(b)\u003c/strong\u003e aspect, \u003cstrong\u003e(c)\u003c/strong\u003e plan curvature, \u003cstrong\u003e(d)\u003c/strong\u003e TWI, \u003cstrong\u003e(e)\u003c/strong\u003eTRI, \u003cstrong\u003e(f)\u003c/strong\u003e LCF, \u003cstrong\u003e(g)\u003c/strong\u003e drainage density, \u003cstrong\u003e(h)\u003c/strong\u003e distance from river, \u003cstrong\u003e(i)\u003c/strong\u003e lithology, \u003cstrong\u003e(j)\u003c/strong\u003e average annual rainfall, \u003cstrong\u003e(k)\u003c/strong\u003eLULC, \u003cstrong\u003e(l)\u003c/strong\u003e NDVI and \u003cstrong\u003e(m)\u003c/strong\u003e soil moisture.\u003c/p\u003e","description":"","filename":"floatimage3.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7715126/v1/97cda33a39fb3e46d47ff1c0.jpeg"},{"id":94637905,"identity":"86efb6b2-471f-4629-89ae-4e0cf74cbc4c","added_by":"auto","created_at":"2025-10-29 07:26:05","extension":"jpeg","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":1006478,"visible":true,"origin":"","legend":"\u003cp\u003eLocations of flooded areas in the June-July 2025 monsoon in Himachal Pradesh and Uttarakhand, along with the predicted flood susceptibility map from CNN. Overlapping black pixels show the flooded area as retrieved from GEE using Sentinel-1 C-band SAR. Images \u003cstrong\u003e(a to g) \u003c/strong\u003eshow the zoomed location extent of the flood probabilities of locations reported as flooded during the specified period and thereafter.\u003c/p\u003e","description":"","filename":"floatimage4.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7715126/v1/df5f87b0984427b12f7815e2.jpeg"},{"id":94640630,"identity":"54158299-cb4a-4d23-b5cb-6dda040b4e5a","added_by":"auto","created_at":"2025-10-29 07:49:57","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":192260,"visible":true,"origin":"","legend":"\u003cp\u003eRelative importance of each conditioning/causing factor of flood susceptibility\u003c/p\u003e","description":"","filename":"floatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-7715126/v1/7cba79ce119f97811ec8dd7c.png"},{"id":94637913,"identity":"c6547de0-3e83-4b2b-b771-9828ce001dc4","added_by":"auto","created_at":"2025-10-29 07:26:05","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":280891,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003e(a) \u003c/strong\u003eLoss, \u003cstrong\u003e(b)\u003c/strong\u003e accuracy, \u003cstrong\u003e(c to e) \u003c/strong\u003eROC curves evaluated during each stage of the execution and \u003cstrong\u003e(f)\u003c/strong\u003etesting confusion matrix for flood susceptibility CNN U-net model.\u003c/p\u003e","description":"","filename":"floatimage6.png","url":"https://assets-eu.researchsquare.com/files/rs-7715126/v1/22b0ff557f1cbd4dda0cf2de.png"},{"id":94637910,"identity":"2d33e85d-a180-4131-a16b-e0b1748da61d","added_by":"auto","created_at":"2025-10-29 07:26:05","extension":"jpeg","order_by":7,"title":"Figure 7","display":"","copyAsset":false,"role":"figure","size":583355,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003e(a) \u003c/strong\u003eLocalised flood susceptible areas at a flooding location identified by the CNN U-net model with LCF and \u003cstrong\u003e(b)\u003c/strong\u003e without LCF as a conditioning factor\u003c/p\u003e","description":"","filename":"floatimage7.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7715126/v1/8d71176d580613c930036006.jpeg"},{"id":94641198,"identity":"a384bd37-bf48-4e98-8a53-e97f4ba2f433","added_by":"auto","created_at":"2025-10-29 07:51:42","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":7331048,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7715126/v1/187ab3d5-46ed-4361-afbe-2966fef0185f.pdf"}],"financialInterests":"","formattedTitle":"Rapid Flood Susceptibility Mapping in the Indian Himalayan Region using CNN-U-Net Segmentation: Insights from the 2025 Monsoon Events","fulltext":[{"header":"1. Introduction","content":"\u003cp\u003eMountain regions worldwide are experiencing unprecedented vulnerability to hydro-meteorological disasters, with climate change intensifying flood risks across major mountain ranges globally (Rick et al. \u003cspan citationid=\"CR41\" class=\"CitationRef\"\u003e2023\u003c/span\u003e). Mountains across the globe are emerging as climate hotspots, where enhanced warming and altered precipitation regimes threaten cryospheric reserves, water security, and ecosystem stability (Pepin et al. \u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e2022\u003c/span\u003e). Ombadi et al. (\u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e2023\u003c/span\u003e) showed that mountains across the Northern Hemisphere are becoming increasingly susceptible to extreme rainfall events, with every 1\u0026deg;C of global warming leading to approximately 15%(Pepin et al., \u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e2022\u003c/span\u003e) more rain at high elevations, substantially exceeding the rate expected from atmospheric moisture increases alone. The European Alps have documented significant increases in torrential flooding, with climate change and settlement growth contributing to rising losses, while glacial lake outburst floods (GLOFs) across mountain regions have increased in frequency, with documented events showing temporal increases globally (Schl\u0026ouml;gl et al., \u003cspan citationid=\"CR47\" class=\"CitationRef\"\u003e2021\u003c/span\u003e). According to the Emergency Events Database (EM-DAT), there were 203 flood disasters worldwide in 2023, resulting in 7,763 deaths and economic losses exceeding US\u003cspan\u003e$\u003c/span\u003e20.4\u0026nbsp;billion (CRED 2023).\u003c/p\u003e\u003cp\u003eThe Indian Himalayan Region (IHR) is inherently vulnerable to various geological and hydro-meteorological disasters due to its geographical setting, which is characterised by steep, unstable terrain and the prevalence of erratic weather patterns (Nagamani et al. \u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Shah and Malakar \u003cspan citationid=\"CR49\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Roul et al. \u003cspan citationid=\"CR43\" class=\"CitationRef\"\u003e2025\u003c/span\u003e). Over the past eight years, the state of Uttarakhand has experienced over 25,000 disaster incidents, resulting in 705 deaths from flash floods and landslides, including 389 deaths from flash floods alone (India Today 2025; The Times of India \u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e2025\u003c/span\u003ea). In 2021, the state of Himachal Pradesh reported 476 deaths and ₹1,151 crore in losses, while in 2023, there were 441 fatalities and ₹12,000 crore in damages (Down To Earth \u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e2025\u003c/span\u003e). In 2024-25, Himachal Pradesh recorded 358 deaths from hydro-meteorological disasters, the second-highest in India. The state also reported the loss of over 7,000 cattle and damage to 1,004 houses (The Times of India \u003cspan citationid=\"CR64\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). In June\u0026ndash;August 2025, the states of Himachal Pradesh and Uttarakhand in India experienced intense monsoon activity leading to cloudbursts, flash floods, and landslides, resulting in over 100 fatalities and extensive infrastructure damage (NDTV 2025). On 5th August 2025, the Dharali village located in the Uttarkashi district of Uttarakhand experienced significant destruction of residential and commercial structures. The incident led to the loss of numerous lives, with several people missing or crushed under mudflows. This devastation was primarily attributed to a substantial debris slide resulting from a potential cloudburst and ensuing flash floods (The Hindu \u003cspan citationid=\"CR59\" class=\"CitationRef\"\u003e2025\u003c/span\u003e; The Times of India \u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e2025\u003c/span\u003eb). Such recurring incidences in the IHR are primarily driven by both natural and human-induced changes, which significantly contribute to flood occurrences (Nagamani et al. \u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Alam et al. \u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e2025\u003c/span\u003e). These areas are prone to multiple hazards that often co-occur and interact across space and time, thus elevating the need for comprehensive susceptibility maps to support integrated risk management (Tripathi et al. \u003cspan citationid=\"CR65\" class=\"CitationRef\"\u003e2023\u003c/span\u003e; Sreevalsan-Nair and Mundayatt \u003cspan citationid=\"CR53\" class=\"CitationRef\"\u003e2025\u003c/span\u003e).\u003c/p\u003e\u003cp\u003eDespite the high propensity for disasters in the IHR, many regions continue to face severe data scarcity (Agrawal et al. 2018; Sultan et al. \u003cspan citationid=\"CR54\" class=\"CitationRef\"\u003e2022\u003c/span\u003e). To overcome this challenge, satellite-based observations, particularly those derived from Sentinel-1 SAR, have become indispensable for inundation mapping, owing to their unique capability to capture imagery through clouds and in the absence of daylight (Uddin et al. \u003cspan citationid=\"CR68\" class=\"CitationRef\"\u003e2019\u003c/span\u003e; Konapala et al. \u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e2021\u003c/span\u003e). However, their efficacy in the Himalayan terrain remains limited. Sentinel-1 revisit times may miss short-lived floods (~\u0026thinsp;58% detectability for events), while issues such as topographic distortions, vegetation interference, and rapidly changing flood conditions further reduce their reliability (Tarpanelli et al. \u003cspan citationid=\"CR55\" class=\"CitationRef\"\u003e2022\u003c/span\u003e; Breznik et al. \u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e2025\u003c/span\u003e). In such contexts, Hydraulic-cum-Hydrodynamic Modelling (HHM) offers a valuable alternative by simulating flood dynamics across diverse terrains and hydrological settings with high accuracy. However, its operational application is often constrained by the need for extensive multidimensional input data and substantial computational resources (Pareta \u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). These requirements pose a major hindrance for such regions, especially those existing in the low and middle-income nations, where financial and technical resources for data procurement are limited (Bentivoglio et al. \u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e2022\u003c/span\u003e).\u003c/p\u003e\u003cp\u003eGiven these limitations, there is a pressing need for approaches that can overcome data scarcity while remaining operationally feasible in resource-constrained settings. In this regard, data-driven methods, particularly Deep Learning (DL) models, are emerging as powerful tools for flood mapping and risk assessment in data-scarce regions (Karim et al. \u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e2023\u003c/span\u003e; Fereshtehpour et al. \u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). In disaster-prone regions like the IHR, where rapid and reliable information is essential during the crucial initial hours and data availability is limited, the rapid deployment capability of CNN-based models provide a decisive operational advantage for flood susceptibility mapping. While traditional hydrodynamic models require 40\u0026thinsp;+\u0026thinsp;minutes to several days for complete flood simulations, CNN models can generate high-resolution flood susceptibility maps within seconds to minutes after training (Song et al. \u003cspan citationid=\"CR51\" class=\"CitationRef\"\u003e2025\u003c/span\u003e; Taysi et al. \u003cspan citationid=\"CR57\" class=\"CitationRef\"\u003e2025\u003c/span\u003e). This dramatic increase in speed, achieving computational efficiency several times greater than conventional approaches, enables near real-time hazard assessment, crucial for emergency response. Traditional HHM requires cross-sections, bathymetry, roughness coefficients, boundary conditions, discharge records, gauged calibration data, and meshing, along with extensive computational resources and continuous monitoring inputs for reliable simulation of flood dynamics (Karim et al. \u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e2023\u003c/span\u003e). In contrast, DL models can operate using satellite or remote sensing data combined with minimal ground truth information, such as flood extent, making them particularly suited for regions that face data scarcity and resource limitations (Karim et al. \u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e2023\u003c/span\u003e; Khosravi et al. \u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e2023\u003c/span\u003e; Fereshtehpour et al. \u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Biazar et al. \u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e2025\u003c/span\u003e).\u003c/p\u003e\u003cp\u003eDL models, especially Convolutional Neural Networks (CNNs), offer critical advantages in contexts where accurate representation of spatial flood characteristics is required, as they can capture complex feature interactions and produce more reliable susceptibility assessments and hazard maps than conventional machine learning and numerical approaches (Wang et al. \u003cspan citationid=\"CR72\" class=\"CitationRef\"\u003e2020\u003c/span\u003e; Bentivoglio et al. \u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e2022\u003c/span\u003e). They are particularly valuable for operational workflows in large mountainous regions such as the IHR, where they can support timely planning and early-stage risk zoning without the need for computationally intensive full hydrodynamic simulations across numerous scenarios (Karim et al. \u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e2023\u003c/span\u003e). CNNs are adept at learning spatial patterns and non-linear interactions across geospatial grids, enabling reliable predictions in areas where satellite observations are limited. Ullah et al. (\u003cspan citationid=\"CR69\" class=\"CitationRef\"\u003e2022\u003c/span\u003e) found that CNN-based models outperform traditional machine learning methods, capturing spatial interconnectedness between neighbouring pixels more effectively. By integrating multi-source predictors such as digital elevation models, land cover, rainfall intensity, and hazard inventories, CNNs provide a robust framework for early-stage risk zoning and flood susceptibility mapping, without the prohibitive computational costs of full hydrodynamic simulations (Karim et al. \u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e2023\u003c/span\u003e). Looking ahead, such models hold strong potential to serve as the backbone for rapid operational, AI-enabled decision-support systems tailored to the Himalayan context, ultimately advancing multi-hazard resilience and preparedness.\u003c/p\u003e\u003cp\u003eIn the context of the IHR, where disasters evolve rapidly and data scarcity hampers timely decision-making, the need for a rapid, reliable, and easily deployable information tool is paramount. Previous research has demonstrated that topographic variables, including elevation, slope, drainage density, and relative relief, constitute the most influential predictors in machine learning based flood susceptibility assessments in mountainous areas (Zhao et al. \u003cspan citationid=\"CR73\" class=\"CitationRef\"\u003e2018\u003c/span\u003e). Furthermore, recent studies emphasize that even subtle micro-topographic variations in relatively flatter valley regions are critical for accurately delineating localized flood hazards, as concave and convex terrain features strongly influence water accumulation and flow dynamics (Safaei-Moghadam et al. \u003cspan citationid=\"CR44\" class=\"CitationRef\"\u003e2023\u003c/span\u003e). These findings underscore the necessity of incorporating both large-scale morphometric controls and fine-scale topographic features to enhance the reliability of flood hazard mapping in data-scarce and topographically complex environments such as the Himalayas. This study addresses this gap and advances flood susceptibility mapping in complex, data-scarce mountainous regions through three key innovations. First, we integrate a Local Convexity Factor (LCF) into a CNN U-Net framework, enabling enhanced micro-topographic characterisation that improves hazard delineation in rugged landscapes and reduces overestimation in depositional zones representing a key advancement for flood risk modelling in the complex geomorphological context of the Indian Himalayas. Second, by combining 14 relevant hydro-geomorphological predictors, including terrain, land cover, soil, and rainfall climatology within a DL semantic segmentation model, we capture both riverine and upland flood hotspots that are often overlooked by conventional approaches in mountainous terrain settings. Third, the proposed framework is lightweight and rapidly deployable, capable of retraining within hours, making it operationally suitable for regions where near-real-time hazard mapping is critical but hydrodynamic simulations are impractical due to data or resource constraints. This method leverages learned terrain-hydrology patterns to deliver high-resolution susceptibility maps with superior spatial detail and predictive reliability. Beyond its immediate application in the Indian Himalayan Region, the framework is readily transferable to other mountainous and data-scarce environments worldwide, supporting integrated multi-hazard risk assessment. Its design enables a practical decision support tool for land-use planners, policymakers, and disaster management agencies. By informing targeted interventions, guiding infrastructure placement, and enhancing early-warning capabilities, the approach contributes to building climate-resilient cities and communities capable of adapting to increasingly unpredictable hazard regimes.\u003c/p\u003e"},{"header":"2. Study area","content":"\u003cp\u003eThe present study focuses on the northwestern segment of the Indian Himalayan Region (IHR), specifically the states of Himachal Pradesh (HP) and Uttarakhand (UK) \u003cb\u003e(\u003c/b\u003eFig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e\u003cb\u003e)\u003c/b\u003e. These mountain states are highly vulnerable to hydro-meteorological extremes such as intense precipitation, flash floods, and glacial lake outburst floods, along with frequent landslides and riverbank erosion, owing to their fragile topography and active tectonics (Kansal and Singh \u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e2022\u003c/span\u003e; Gupta et al. \u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). Uttarakhand is located between 28\u0026deg;43\u0026rsquo; and 31\u0026deg;28\u0026rsquo; N latitude and 77\u0026deg;34\u0026rsquo; and 81\u0026deg;3\u0026rsquo; E longitude, while Himachal Pradesh lies between 30\u0026deg;22\u0026rsquo; and 33\u0026deg;16\u0026rsquo; N latitude and 75\u0026deg;35\u0026rsquo; and 79\u0026deg;01\u0026rsquo; E longitude. The elevation in these regions ranges from 187 meters to approximately 7,124 meters, encompassing both states collectively. Studies and analyses indicate that over 45% of Himachal Pradesh is highly prone to floods, landslides, and avalanches, while 85% of the districts in Uttarakhand are vulnerable to extreme floods (CEEW \u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e2021\u003c/span\u003e; PreventionWeb \u003cspan citationid=\"CR39\" class=\"CitationRef\"\u003e2025\u003c/span\u003e). The region has experienced devastating flood events, including the 2013 Kedarnath disaster (6,054 fatalities) and July 2023 Himachal Pradesh floods (187 fatalities, economic losses of ₹5,620 crore) (UNDRR CRED 2022; Sphere India \u003cspan citationid=\"CR52\" class=\"CitationRef\"\u003e2023\u003c/span\u003e). The increasing frequency and intensity of hydro-meteorological disasters since 1997, combined with rapid glacial lake expansion (1,048 glacial lakes in Himachal's Sutlej catchment as of 2023), establishes this region as a critical area for comprehensive flood susceptibility assessment (India Today \u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e2024\u003c/span\u003e).\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003e\u003c/p\u003e"},{"header":"3. Materials and Methods","content":"\u003cp\u003eThe materials and methodology include the selection and processing of 14 flood-conditioning factors, the design and training of a CNN U-Net model for flood susceptibility prediction, evaluation of model performance using various metrics, analysis of factor contributions through permutation feature importance, and delineation of flood extent using Sentinel-1 SAR imagery. The overall workflow, integrating multi-source geospatial data, CNN U-Net modeling, and Sentinel-1 SAR flood mapping, is illustrated in Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e.\u003c/p\u003e\u003cdiv id=\"Sec4\" class=\"Section2\"\u003e\u003ch2\u003e3.1 Flood Conditioning Factors\u003c/h2\u003e\u003cp\u003eA total of 14 flood-conditioning factors were selected based on their proven relevance in flood susceptibility modelling, as supported by previous studies (Saravanan et al. \u003cspan citationid=\"CR45\" class=\"CitationRef\"\u003e2023\u003c/span\u003e; Amiri et al. \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Dey et al. \u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Tripathi and Mohanty \u003cspan citationid=\"CR66\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). The rasters for the conditioning factors are stacked as a 3D tensor \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:X\\in\\:{\\mathbb{R}}^{H\\times\\:W\\times\\:N}\\)\u003c/span\u003e\u003c/span\u003e where \u003cem\u003eH\u003c/em\u003e, and \u003cem\u003eW\u003c/em\u003e are spatial dimensions and \u003cem\u003eN\u003c/em\u003e is the number of predictors; No Data values are converted to \u0026ldquo;NaN\u0026rdquo; at source read time and later filled for model input during patching. Each band is standardized using z-score normalization computed spatially, per band (Eq.\u0026nbsp;\u003cspan refid=\"Equ1\" class=\"InternalRef\"\u003e1\u003c/span\u003e):\u003cdiv id=\"Equ1\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ1\" name=\"EquationSource\"\u003e\n$$\\:{X}_{norm}=\\:\\frac{X-\\mu\\:}{\\sigma\\:}$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e1\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003ewhere \u003cem\u003e\u0026micro;\u003c/em\u003e is computed for each band by averaging all pixels across the image height and width while ignoring missing values. The standard deviation \u003cem\u003eσ\u003c/em\u003e is computed similarly for each band across the image height and width, again ignoring missing values (NaNs). This is the dispersion of pixel values over rows and columns for that band. The standardized multi-band array is written to a Zarr store (an open, chunked, and compressed storage format for large \u003cem\u003eN\u003c/em\u003e-dimensional arrays) with carefully chosen chunk sizes, enabling block-wise, random access to small subarrays without loading the full raster into memory. This chunked, compressed layout accelerates sliding-window training and inference by reading only the tiles needed at each step, supports parallel input/output (multiple patches read or written concurrently), and scales efficiently to very large rasters on local disks or object storage. Details of chosen flood conditioning factors, viz., sources and spatio-temporal resolutions, are listed in Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e, with significance discussed in the following subsections.\u003c/p\u003e\u003cdiv id=\"Sec5\" class=\"Section3\"\u003e\u003ch2\u003e3.1.1 Altitude\u003c/h2\u003e\u003cp\u003eIn the Himalayan context, elevation directly influences precipitation patterns, temperature gradients, and snow-melt processes that contribute to flood generation (Patel et al. \u003cspan citationid=\"CR37\" class=\"CitationRef\"\u003e2022\u003c/span\u003e). Lower elevations in the foothills and valley regions are more susceptible to flooding due to gravitational water flow and natural accumulation processes from surrounding higher terrain (Ullah et al. \u003cspan citationid=\"CR69\" class=\"CitationRef\"\u003e2022\u003c/span\u003e; Dey et al. \u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Nagamani et al. \u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Kumar et al. \u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e2024\u003c/span\u003e) \u003cb\u003e(\u003c/b\u003eFig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e\u003cb\u003e)\u003c/b\u003e.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec6\" class=\"Section3\"\u003e\u003ch2\u003e3.1.2 Slope\u003c/h2\u003e\u003cp\u003eThe rugged Himalayan terrain exhibits steep slopes that promote rapid runoff generation, but areas with gentle slopes (0\u0026ndash;5\u0026deg;) in valley bottoms and terraced agricultural regions are more prone to water accumulation and flooding (Patel et al. \u003cspan citationid=\"CR37\" class=\"CitationRef\"\u003e2022\u003c/span\u003e). Slope characteristics in the study region significantly influence surface water flow velocity and flood generation patterns, with the region's diverse topography ranging from gentle valley floors to extremely steep mountain faces \u003cb\u003e(\u003c/b\u003eFig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003ea\u003cb\u003e)\u003c/b\u003e.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec7\" class=\"Section3\"\u003e\u003ch2\u003e3.1.3 Aspect\u003c/h2\u003e\u003cp\u003eAspect orientation in the Himachal Pradesh and Uttarakhand regions critically influences flood susceptibility through its control over monsoon precipitation patterns, solar radiation, and soil moisture conditions (Nagamani et al. \u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). The southwestern and southeastern facing slopes receive the maximum impact from monsoon winds carrying moisture from the Bay of Bengal, resulting in enhanced precipitation and increased flood risk (Sajwan and Sushil \u003cspan citationid=\"CR46\" class=\"CitationRef\"\u003e2016\u003c/span\u003e; Nagamani et al. \u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e2024\u003c/span\u003e) \u003cb\u003e(\u003c/b\u003eFig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eb\u003cb\u003e)\u003c/b\u003e.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec8\" class=\"Section3\"\u003e\u003ch2\u003e3.1.4 Plan curvature\u003c/h2\u003e\u003cp\u003ePlan curvature describes the curvature of the land surface perpendicular to the direction of maximum slope, indicating whether water flow converges or diverges across the landscape (Edamo et al. \u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). Concave surfaces tend to concentrate flow and promote water accumulation, increasing flood susceptibility, while convex surfaces facilitate flow dispersion and reduce flood risk (Aydin and Iban 2023). Flat areas with minimal curvature are particularly vulnerable to flooding as they provide limited natural drainage (Mehravar et al. 2023) \u003cb\u003e(\u003c/b\u003eFig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003ec\u003cb\u003e)\u003c/b\u003e.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec9\" class=\"Section3\"\u003e\u003ch2\u003e3.1.5 Topographic wetness index (TWI)\u003c/h2\u003e\u003cp\u003eTWI quantifies the potential for water accumulation at any location based on the upstream contributing area and local slope gradient (Al-Kindi and Alabri 2024; Amiri et al. \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). Higher TWI values indicate areas with greater potential for soil saturation and surface runoff generation, making them more susceptible to flooding (Aydin and Iban 2023). This index effectively captures the topographic control on hydrological processes and is calculated as the natural logarithm of the ratio between specific catchment area and slope tangent (Li and Hong 2023; Khosravi et al. 2019) \u003cb\u003e(\u003c/b\u003eFig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003ed\u003cb\u003e)\u003c/b\u003e.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec10\" class=\"Section3\"\u003e\u003ch2\u003e3.1.6 Topographic roughness index (TRI)\u003c/h2\u003e\u003cp\u003eTRI measures the variability in elevation within a local area and indicates surface complexity and roughness (Aydin and Iban 2023). Areas with high TRI values typically exhibit complex terrain with varied elevation changes that can create local drainage complications and influence flood flow patterns (Al-Kindi and Alabri 2024) \u003cb\u003e(\u003c/b\u003eFig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003ee\u003cb\u003e)\u003c/b\u003e.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec11\" class=\"Section3\"\u003e\u003ch2\u003e3.1.7 Local convexity factor (LCF)\u003c/h2\u003e\u003cp\u003eThe LCF was used as a conditioning factor in the deep learning framework to enhance flood susceptibility mapping in the complex terrain of the Indian Himalayas. Proposed by Liu et al. (\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e2025\u003c/span\u003e), LCF utilizes the correlation between normalized Digital Elevation Model (DEM) patches and a standard Gaussian surface to identify micro-depressions that concentrate runoff, capturing subtle terrain features often missed by conventional curvature and slope metrics. The LCF methodology proposed by Liu et al. (\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e2025\u003c/span\u003e) iterates over a range of window sizes (2\u0026ndash;50) and selects the maximum correlation value for each pixel, but this can lead to over or underemphasis of features, especially in heterogeneous terrain. The presented methodology introduces an adaptive windowing approach for pixel-wise calculation of LCF based on local slope and curvature extracted from a DEM. The procedure entails loading and preprocessing the DEM, normalizing elevation values, computing slope and curvature derivatives, and dynamically determining the local window size for each pixel based on these factors. For each pixel, the analysis window size (\u003cem\u003eW\u003c/em\u003e) is dynamically selected based on terrain factors using the following formulation (Eq.\u0026nbsp;\u003cspan refid=\"Equ2\" class=\"InternalRef\"\u003e2\u003c/span\u003e):\u003cdiv id=\"Equ2\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ2\" name=\"EquationSource\"\u003e\n$$\\:W=clip({k}_{1}\\:.\\:slope+\\:{k}_{2}.\\:\\left|curvature\\right|+\\:{k}_{3},\\:2,\\:50)$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e2\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003eWhere \u003cem\u003ek\u003c/em\u003e\u003csub\u003e1\u003c/sub\u003e, \u003cem\u003ek\u003c/em\u003e\u003csub\u003e2\u003c/sub\u003e and \u003cem\u003ek\u003c/em\u003e\u003csub\u003e3\u003c/sub\u003e are user-set parameters, and \u003cem\u003eW\u003c/em\u003e is clipped between 2 and 50 to remain numerically stable. Slope and curvature were first normalized to the range [0,1]. The constants \u003cem\u003ek\u003c/em\u003e\u003csub\u003e1\u003c/sub\u003e, \u003cem\u003ek\u003c/em\u003e\u003csub\u003e2\u003c/sub\u003e and \u003cem\u003ek\u003c/em\u003e\u003csub\u003e3\u003c/sub\u003e control the influence of slope, curvature, and a baseline offset, respectively. We set \u003cem\u003ek\u003c/em\u003e\u003csub\u003e1\u003c/sub\u003e\u0026thinsp;=\u0026thinsp;30, \u003cem\u003ek\u003c/em\u003e\u003csub\u003e2\u003c/sub\u003e\u0026thinsp;=\u0026thinsp;20 and \u003cem\u003ek\u003c/em\u003e\u003csub\u003e3\u003c/sub\u003e\u0026thinsp;=\u0026thinsp;2. These values were chosen so that, for our\u0026thinsp;~\u0026thinsp;90 m DEM, the resulting window sizes realistically capture the scale of the hilly landforms. The mean window size corresponded to about 1.3 km on the ground, which is large enough to capture meso-scale landform patterns (ridge-valley contrasts, slope breaks), with very few pixels reaching the lower (2 pixel) or upper (50 pixel) limits. In this formulation, slope has slightly more weight than curvature, meaning that steep gradients tend to produce larger windows, while curvature refines the adjustment in convex and concave terrain. The offset \u003cem\u003ek\u003c/em\u003e\u003csub\u003e3\u003c/sub\u003e​ ensures that even very flat regions retain a minimal window. We verified the suitability of these parameters through simple diagnostics, including histograms of window size distribution and inspection of spatial maps, and found that small changes in the parameter values did not substantially alter the results. Building on this foundation, the subsequent step involves extracting a normalized DEM patch of the adaptive window size and correlating it with a Gaussian reference surface. The resulting Pearson correlation coefficient serves as the Local Convexity Factor (LCF). Raster outputs are exported for further geomorphological analysis and hazard assessment. The adaptive nature of this method ensures precise, terrain-sensitive quantification of convexity, outperforming the standard fixed-window technique \u003cb\u003e(\u003c/b\u003eFig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003ef\u003cb\u003e)\u003c/b\u003e.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec12\" class=\"Section3\"\u003e\u003ch2\u003e3.1.8 Drainage density\u003c/h2\u003e\u003cp\u003eDrainage density represents the total length of stream channels per unit area and indicates the drainage efficiency of a watershed (Al-Kindi and Alabri 2024). High drainage density typically facilitates rapid water evacuation and reduces flood risk, while low drainage density can lead to water accumulation and increased flood susceptibility (Abusarif et al. 2023) \u003cb\u003e(\u003c/b\u003eFig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eg\u003cb\u003e)\u003c/b\u003e.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec13\" class=\"Section3\"\u003e\u003ch2\u003e3.1.9 Distance from river\u003c/h2\u003e\u003cp\u003eProximity to water bodies is a crucial factor in flood susceptibility assessment, with areas closer to rivers and streams exhibiting significantly higher flood risk (Al-Kindi and Alabri 2024; Dey et al. \u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). The relationship between distance and flood susceptibility typically follows an inverse exponential decay, where flood probability decreases rapidly with increasing distance from water sources (Aydin and Iban 2023) \u003cb\u003e(\u003c/b\u003eFig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eh\u003cb\u003e)\u003c/b\u003e.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec14\" class=\"Section3\"\u003e\u003ch2\u003e3.1.10. Lithology\u003c/h2\u003e\u003cp\u003eGeological formations control subsurface drainage, infiltration capacity, and groundwater flow patterns, significantly influencing surface flood behaviour (Al-Kindi and Alabri 2024; Khosravi et al. 2019). Impermeable rock types such as granite and schist reduce infiltration capacity and increase surface runoff, while permeable formations like limestone and sandstone promote groundwater recharge and reduce flood susceptibility (Aydin and Iban 2023). Lithology data from the Geological Survey of India (GSI) were used in this study. A total of 295 lithologic classes were identified in the same across Himachal Pradesh and Uttarakhand, as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003ei.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec15\" class=\"Section3\"\u003e\u003ch2\u003e3.1.11 Rainfall\u003c/h2\u003e\u003cp\u003eThe rainfall data for the study area was derived by calculating the annual average from the IMD gridded time-series data spanning from 1990 to 2024. This annual average raster was then converted into point data, which were subsequently interpolated using the inverse distance weighting (IDW) method at a pixel size of 90 meters. This process resulted in a uniformly distributed rainfall raster for the region \u003cb\u003e(\u003c/b\u003eFig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003ej\u003cb\u003e)\u003c/b\u003e.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec16\" class=\"Section3\"\u003e\u003ch2\u003e3.1.12 Land use land cover (LULC)\u003c/h2\u003e\u003cp\u003eLULC types determine surface permeability, roughness characteristics, and runoff coefficients, making them essential factors in flood modeling (Al-Kindi and Alabri 2024; Amiri et al. \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). Urban areas with impervious surfaces exhibit high runoff coefficients and increased flood susceptibility, while forested areas provide natural flood protection through enhanced infiltration and flow retardation (Aydin and Iban 2023). Agricultural lands show intermediate flood susceptibility depending on crop type, soil conditions, and management practices (Ahmadlou et al. 2021) \u003cb\u003e(\u003c/b\u003eFig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003ek\u003cb\u003e)\u003c/b\u003e.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec17\" class=\"Section3\"\u003e\u003ch2\u003e3.1.13 Normalized difference vegetation index (NDVI)\u003c/h2\u003e\u003cp\u003eNDVI quantifies vegetation density and health, which directly affects surface roughness, infiltration capacity, and runoff generation (Al-Kindi and Alabri 2024; Dey et al. \u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). Dense vegetation increases surface roughness, promotes infiltration, and reduces flood susceptibility through enhanced evapotranspiration and flow retardation (Amiri et al. \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). Areas with low NDVI values, indicating sparse vegetation or bare soil, exhibit higher flood susceptibility due to reduced infiltration capacity and increased surface runoff (Aydin and Iban 2023; Mehravar et al. 2023) \u003cb\u003e(\u003c/b\u003eFig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003el\u003cb\u003e)\u003c/b\u003e.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec18\" class=\"Section3\"\u003e\u003ch2\u003e3.1.14 Soil Moisture\u003c/h2\u003e\u003cp\u003eAntecedent soil moisture strongly conditions the fraction of rainfall that becomes direct runoff, with higher soil moisture leading to substantially increased flood peaks because more precipitation is converted to surface flow rather than infiltrating the soil (Merz et al. \u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e2021\u003c/span\u003e). This nonlinear relationship means that modest increases in precipitation can produce disproportionately large increases in runoff when soils are near saturation (Merz et al. \u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e2021\u003c/span\u003e). The mean top layer (0\u0026ndash;5 cm) soil moisture from March 2015 to May 2025 was computed and used as an input feature to the CNN \u003cb\u003e(\u003c/b\u003eFig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003em\u003cb\u003e)\u003c/b\u003e.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eSince CNNs are supervised learning algorithms, they require labelled data for training. To provide these labels, we utilised the Joint Research Centre (JRC) Global River Flood Hazard Map (100-year return period) (Baugh et al. \u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e2024\u003c/span\u003e) as a reference flood hazard inventory, which served as the ground truth for model training and evaluation.\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003eRelevant conditioning factors for flood in the study area, their sources and resolution.\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"4\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u003cp\u003eS. No.\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eFlood Conditioning Factors\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eScale/Resolution\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e\u003cp\u003eData Source\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e1\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eAltitude (DEM)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\" morerows=\"6\" rowspan=\"7\"\u003e\u003cp\u003e90 meters\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\" morerows=\"6\" rowspan=\"7\"\u003e\u003cp\u003eMERIT\u003c/p\u003e\u003cp\u003e(\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://developers.google.com/earth-engine/datasets/catalog/MERIT_DEM_v1_0_3\u003c/span\u003e\u003cspan address=\"https://developers.google.com/earth-engine/datasets/catalog/MERIT_DEM_v1_0_3\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e2\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eSlope\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e3\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eAspect\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e4\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003ePlan curvature\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e5\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eTWI\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e6\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eTRI\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e7\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eLCF\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e8\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eDrainage density\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\" morerows=\"1\" rowspan=\"2\"\u003e\u003cp\u003eVector data (Projection: WGS 1984 LCC)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\" morerows=\"1\" rowspan=\"2\"\u003e\u003cp\u003eIndia WRIS rivers shapefile (\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://indiawris.gov.in/wris/#/geoSpatialData\u003c/span\u003e\u003cspan address=\"https://indiawris.gov.in/wris/#/geoSpatialData\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e9\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eDistance from the river\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e10\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eLithology\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e1:50,000\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003eGSI Bhukosh (\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://bhukosh.gsi.gov.in/Bhukosh/\u003c/span\u003e\u003cspan address=\"https://bhukosh.gsi.gov.in/Bhukosh/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e11\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eRainfall\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e0.25\u0026deg; \u0026times; 0.25\u0026deg;; annual average (1990\u0026ndash;2024)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003eIMD (\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://www.imdpune.gov.in/lrfindex.php\u003c/span\u003e\u003cspan address=\"https://www.imdpune.gov.in/lrfindex.php\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e12\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eLULC\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e500 m\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003eMODIS data (\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://developers.google.com/earth-engine/datasets/catalog/MODIS_061_MCD12Q1\u003c/span\u003e\u003cspan address=\"https://developers.google.com/earth-engine/datasets/catalog/MODIS_061_MCD12Q1\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e13\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eNDVI\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e250 m\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003eMODIS data (\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://developers.google.com/earth-engine/datasets/catalog/MODIS_061_MYD13Q1\u003c/span\u003e\u003cspan address=\"https://developers.google.com/earth-engine/datasets/catalog/MODIS_061_MYD13Q1\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e14\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eSoil Moisture\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e11 km\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003eSMAP mission (\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://nsidc.org/data/spl4smgp/versions/7\u003c/span\u003e\u003cspan address=\"https://nsidc.org/data/spl4smgp/versions/7\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003e*\u003cem\u003eMERIT: Multi-Error-Removed Improved-Terrain; WRIS: Water Resources Information System; GSI\u003c/em\u003e: \u003cem\u003eGeological Survey of India; IMD: India Meteorological Department; MODIS: Moderate Resolution Imaging Spectroradiometer; SMAP: Soil Moisture Active Passive; WGS: World Geodetic System; LCC: Lambert Conformal Conic\u003c/em\u003e.\u003c/p\u003e\u003cp\u003eThe CNN\u0026rsquo;s predicted outputs, representing pixel-wise flood probability, were visually and quantitatively compared with observed flood-affected areas in both states. The employed CNN U-net architecture is explained in the following sub-sections.\u003c/p\u003e\u003c/div\u003e\u003c/div\u003e\u003cdiv id=\"Sec19\" class=\"Section2\"\u003e\u003ch2\u003e3.2 U-Net architecture adopted for CNN model\u003c/h2\u003e\u003cp\u003eThe Convolutional Neural Network (CNN) U-Net architecture was chosen due to its strong ability to capture spatial dependencies and multiscale features that are critical for flood susceptibility mapping in heterogeneous and mountainous terrains (Fakhri and Gkanatsios \u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e2025\u003c/span\u003e). Unlike conventional pixel-based classifiers, U-Net employs an encoder-decoder structure with skip connections, enabling it to retain both global context and fine-grained spatial details. The details are provided in sections below and the representative model structure is shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e. To further refine the design, hyperparameter tuning was conducted using the Keras Tuner framework with a Random Search strategy. This enabled systematic exploration of filter depth, dropout rates, and learning rate, ensuring that the selected architecture was not only theoretically appropriate but also empirically optimised for the dataset.\u003c/p\u003e\u003cdiv id=\"Sec20\" class=\"Section3\"\u003e\u003ch2\u003e3.2.1 Encoder Path\u003c/h2\u003e\u003cp\u003eThe encoder comprises a series of convolutional blocks with 3\u0026times;3 filters, each followed by Rectified Linear Unit (ReLU) activation and batch normalization to accelerate convergence and prevent internal covariate shifts. Each block is followed by a 2\u0026times;2 max-pooling layer, which progressively downsamples the feature maps while capturing high-level contextual information. Here, a custom encoder built from multiple Keras Convolution 2 Dimensional\u0026thinsp;+\u0026thinsp;Maximum Pooling 2D layers has been used, designed specifically for patch-wise hazard susceptibility analysis. Hyperparameter search identified an optimal first-layer filter depth of 32 filters, which balanced predictive accuracy and computational efficiency.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec21\" class=\"Section3\"\u003e\u003ch2\u003e3.2.2 Decoder Path\u003c/h2\u003e\u003cp\u003eThe decoder mirrors the encoder, with transposed convolutions (up-convolutions) used to upsample the feature maps. Skip connections between encoder and decoder layers ensure that spatial features lost during pooling are reintroduced, enhancing boundary delineation of flood-prone zones.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec22\" class=\"Section3\"\u003e\u003ch2\u003e3.2.3 Output Layer\u003c/h2\u003e\u003cp\u003eThe final layer uses a 1\u0026times;1 convolution with a sigmoid activation function, generating pixel-wise flood susceptibility probabilities ranging from 0 to 1.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec23\" class=\"Section3\"\u003e\u003ch2\u003e3.2.4 Training Configuration\u003c/h2\u003e\u003cp\u003e\u003cul\u003e\u003cli\u003e\u003cp\u003eLoss Function: A custom loss function using a weighted mask with binary cross-entropy (BCE) for binary classification was implemented in TensorFlow. This approach is particularly effective for probability-based predictions. The weighted binary cross-entropy loss calculates the loss using class weights that are inversely proportional to class frequency, which helps to emphasize the minority class (in this case, the flood class) and address class imbalance issues.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eOptimizer: The Adam optimizer with a tuned learning rate of ~\u0026thinsp;0.001 was employed for faster convergence.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eData Split: Input datasets were randomly split into 70% for training, 15% for validation, and 15% for testing.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eBatch Size and Epochs: A batch size of 64 was employed, which offered a practical balance between gradient stability and computational efficiency on CPU. Smaller batch sizes tend to produce noisier gradient updates, while excessively large batches increase memory demand without improving convergence speed on CPU. The maximum number of training epochs was set to 100 to provide sufficient opportunity for the model to converge. However, to avoid unnecessary computation and potential overfitting, Early Stopping was applied with a patience value of 10 (Nemni et al. \u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e2020\u003c/span\u003e; Tavus et al. \u003cspan citationid=\"CR56\" class=\"CitationRef\"\u003e2022\u003c/span\u003e). This means that training was halted if the validation loss did not improve for ten consecutive epochs, ensuring that the model stopped at an optimal point without wasting resources.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eHyperparameters, including filter depth, dropout, and learning rate, were tuned using Random Search with Keras Tuner, and the final configuration was selected based on validation accuracy and computational efficiency.\u003c/p\u003e\u003c/li\u003e\u003c/ul\u003e\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec24\" class=\"Section3\"\u003e\u003ch2\u003e3.2.5 Evaluation Metrics\u003c/h2\u003e\u003cp\u003eThe performance of the model was evaluated at each stage of the CNN process: training, validation, and testing. We used several metrics to ensure a thorough assessment of its predictive capabilities, including Area Under the Receiver Operating Characteristic Curve (AUC-ROC), Precision, Recall, F1-score, Mean Absolute Error (MAE), Coefficient of determination (R\u003csup\u003e2\u003c/sup\u003e), Root Mean Square Error (RMSE), and the Critical Success Index (CSI). CNN\u0026rsquo;s ability to learn non-linear terrain-hydrology interactions allows it to predict susceptibility even in areas occluded by cloud cover or shadowed by steep terrain, where optical or SAR imagery fails. By leveraging DEM-based derivatives and multi-source geospatial layers, the model effectively generalizes flood risk across unobserved zones, offering a significant improvement over threshold-based or traditional ML methods.\u003c/p\u003e\u003c/div\u003e\u003c/div\u003e\u003cdiv id=\"Sec25\" class=\"Section2\"\u003e\u003ch2\u003e3.3 Permutation Feature Importance\u003c/h2\u003e\u003cp\u003eTo understand how much each flood-conditioning factor influenced the CNN predictions, we used the Permutation Feature Importance (PFI) method. In this approach, the values of each factor were shuffled (permuted) across the test set, so that the link between that factor and flooding was broken. We then measured how much the model\u0026rsquo;s accuracy dropped compared to its performance on the unshuffled data. A large drop meant the factor was important; a small drop meant it contributed less. Finally, the scores were normalized so they added up to 1, giving the relative importance of each factor.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec26\" class=\"Section2\"\u003e\u003ch2\u003e3.4 Sentinel-1 SAR based Flood Extent Mapping\u003c/h2\u003e\u003cp\u003eTo check the accuracy of our proposed CNN model, a flood extent mapping was also performed using Sentinel-1 Synthetic Aperture Radar (SAR) data. For this purpose, Sentinel-1 imagery was accessed and processed through the Google Earth Engine platform to delineate flood-affected areas across Himachal Pradesh and Uttarakhand. The analysis focused on the VH polarization of descending-pass Sentinel-1 Ground Range Detected (GRD) imagery, offering a spatial resolution of 10 meters. Two distinct temporal windows were defined to capture the pre- and post-flood conditions: 1\u0026ndash;15 June 2025 and 1\u0026ndash;10 July 2025, respectively. Image mosaics corresponding to each period were generated and clipped to the administrative boundary of Himachal Pradesh and Uttarakhand. To reduce speckle noise commonly present in SAR datasets, a Refined Lee filter was applied through a custom function adapted from the ESA SNAP toolbox. This filtering was conducted in natural backscatter units, following conversion from decibel (dB) values, and the output was reconverted to dB post-processing. Flood extent was derived by calculating the pixel-wise ratio between post-flood and pre-flood backscatter intensities. Areas exhibiting a backscatter ratio greater than 1.20 were classified as inundated, consistent with established SAR-based flood detection thresholds. To improve classification accuracy, permanently or seasonally inundated water bodies (seasonality\u0026thinsp;\u0026ge;\u0026thinsp;5 months) were masked using the JRC Global Surface Water (GSW) dataset. In addition, terrain with slope gradients exceeding 5\u0026deg;, as determined from HydroSHEDS elevation data, was excluded to eliminate potential false positives associated with radar shadowing or terrain-induced distortions. Spurious noise elements, defined as isolated pixel groups smaller than four connected pixels, were also removed to produce a cleaner flood mask. The final flood extent product was exported as a GeoTIFF at 90-meter resolution, enabling integration into subsequent geospatial analyses and visualization workflows.\u003c/p\u003e\u003c/div\u003e"},{"header":"4. Results","content":"\u003cp\u003eThe 2025 monsoon season, characterized by intense rainfall and multiple cloudburst events, severely impacted settlements, transportation networks, and infrastructure along river valleys in Himachal Pradesh and Uttarakhand. In Himachal Pradesh, some of the most affected areas include stretches of the Beas River in Mandi district and the Sainj Valley (The Logical Indian \u003cspan citationid=\"CR60\" class=\"CitationRef\"\u003e2025\u003c/span\u003e). In Uttarakhand, significant flooding and landslides were reported near Yamunotri and along the Barkot\u0026ndash;Yamunotri road in Uttarkashi, primarily triggered by cloudbursts and flash floods (The Times of India \u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e2025\u003c/span\u003ec). In gradient and hilly terrains, it is difficult to assess this probability due to the challenging landscape.\u003c/p\u003e\u003cdiv id=\"Sec28\" class=\"Section2\"\u003e\u003ch2\u003e4.1 Detection of Flooded Hotspots by the Model\u003c/h2\u003e\u003cp\u003eThe CNN-based flood susceptibility model effectively captured these high-risk areas, demonstrating its ability to learn complex hydro-geomorphological relationships. The probability map of flood occurrence \u003cb\u003e(\u003c/b\u003eFig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003ea to \u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eg\u003cb\u003e)\u003c/b\u003e highlights regions with high to very high susceptibility, aligning with actual flood-affected zones reported during the event. In the Beas River stretch near the Kangra\u0026ndash;Hamirpur\u0026ndash;Mandi boundary \u003cb\u003e(\u003c/b\u003eFig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003ea\u003cb\u003e)\u003c/b\u003e, the model successfully delineated the flooded regions as reported in news articles (The Economic Times \u003cspan citationid=\"CR58\" class=\"CitationRef\"\u003e2025\u003c/span\u003e; The Logical Indian \u003cspan citationid=\"CR60\" class=\"CitationRef\"\u003e2025\u003c/span\u003e), validated by the overlap with Sentinel‑1 SAR-derived water-logged areas. In Mandi \u003cb\u003e(\u003c/b\u003eFig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eb\u003cb\u003e)\u003c/b\u003e and Sainj Valley near Sainj Bridge \u003cb\u003e(\u003c/b\u003eFig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003ec\u003cb\u003e)\u003c/b\u003e, the CNN predictions were consistent with observed damage, correctly classified as high susceptible zones. Along the Barkot\u0026ndash;Yamunotri road \u003cb\u003e(\u003c/b\u003eFig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003ed\u003cb\u003e)\u003c/b\u003e, susceptibility was notably high, reflecting the increased risk to settlements and infrastructure near the Yamunotri pilgrimage route. The Balganga\u0026ndash;Budhakedar (Bhilangana block) region also displayed a wide range of high to very high susceptibility levels \u003cb\u003e(\u003c/b\u003eFig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003ee\u003cb\u003e)\u003c/b\u003e. In Rudraprayag, the Belni Bridge area exhibited very high susceptibility values \u003cb\u003e(\u003c/b\u003eFig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003ef\u003cb\u003e)\u003c/b\u003e.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec29\" class=\"Section2\"\u003e\u003ch2\u003e4.2 Feature Importance of Flood Conditioning Factors\u003c/h2\u003e\u003cp\u003eThe analysis of input features revealed that NDVI, TWI, and altitude are the most critical contributors to flood susceptibility in the study region \u003cb\u003e(\u003c/b\u003eFig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e\u003cb\u003e)\u003c/b\u003e. The high rank of TWI aligns with Dhote et al. (\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e2023\u003c/span\u003e), emphasizing the role of flow accumulation in Himalayan catchments, and has been particularly well-documented in flood susceptibility mapping, as shown by Uca et al. (2022) and Khoirunisa et al. (\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e2021\u003c/span\u003e). LCF ranked 7th among 14 predictors in permutation importance (relative importance\u0026thinsp;=\u0026thinsp;0.024, compared to 0.296 for NDVI).\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eIn practice, LCF improved local delineation accuracy, especially in narrow valleys and interfluve zones, and reduced false positives in flatter depositional reaches, as seen in Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e \u003cb\u003eand\u003c/b\u003e Fig.\u0026nbsp;\u003cspan refid=\"Fig7\" class=\"InternalRef\"\u003e7\u003c/span\u003eb. This refinement represents a methodological step forward for CNN-based hazard mapping in data-scarce and geomorphologically complex regions.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec30\" class=\"Section2\"\u003e\u003ch2\u003e4.3 Performance Evaluation of Models on the Test Dataset\u003c/h2\u003e\u003cp\u003eThe flood susceptibility prediction CNN model achieved an overall accuracy of 97.12% with a Critical Success Index (CSI) being 68.44% \u003cb\u003e(\u003c/b\u003eTable\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e\u003cb\u003e)\u003c/b\u003e. This indicates that the model provides reliable event detection performance, especially under conditions of class imbalance, as in the case of hilly terrains, where flood data is scarce, posing a challenge in identifying actual flood impact locations. Other evaluation metrics also confirm the model\u0026rsquo;s robustness: Root Mean Square Error (RMSE)\u0026thinsp;~\u0026thinsp;0.145, Coefficient of Determination (R\u0026sup2;)\u0026thinsp;~\u0026thinsp;0.719, Precision\u0026thinsp;~\u0026thinsp;0.872, Recall\u0026thinsp;~\u0026thinsp;0.761, F1‑Score\u0026thinsp;~\u0026thinsp;0.813, AUC‑ROC\u0026thinsp;\u0026asymp;\u0026thinsp;0.99 \u003cb\u003e(\u003c/b\u003eTable\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e\u003cb\u003e)\u003c/b\u003e. The high AUC-ROC value reflects the model\u0026rsquo;s strong discriminative ability, even in heterogeneous terrain with limited event data. Notably, the CNN model, trained on multi-factor hazard and local causative factors, overcomes these limitations by learning the underlying geospatial patterns of flood susceptibility, rather than relying solely on real-time flood observations.\u003c/p\u003e\u003cp\u003eFigure\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003ea to \u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003ef present the comprehensive performance evaluation of the proposed flood susceptibility CNN U-Net model, illustrating the progression of training loss and accuracy, the ROC curves assessed at different stages of execution, and the testing confusion matrix, collectively validating the robustness and predictive effectiveness of the model.\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003ePerformance evaluation metrics used in the study.\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"3\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u003cp\u003ePerformance Metric\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eValue (approx.) with LCF\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eValue (approx.) without LCF\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003ePrecision\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.872\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.825\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eRecall\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.761\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.771\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eF-1 Score\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.813\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.797\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eCSI\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e68.44%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e66.32%\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eRMSE\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.145\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.153\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eR\u003csup\u003e2\u003c/sup\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.719\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.685\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eMAE\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.038\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.043\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eAUC\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.988\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.986\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eOverall Accuracy\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e97.12%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e96.83%\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec31\" class=\"Section2\"\u003e\u003ch2\u003e4.4 Lightweight CNN U-Net Performance for Large-Scale Flood Mapping\u003c/h2\u003e\u003cp\u003eThe proposed CNN U-Net demonstrates high efficiency while effectively handling large-scale flood susceptibility mapping. The model operates on a study area covering over 1,00,000 km\u0026sup2; (Himachal Pradesh and Uttarakhand), with each input factor raster of ~\u0026thinsp;135 MB at 90 m resolution. Despite the large input size, the network remains lightweight, requiring only 5.54 MB of storage for the model weights and architecture. Training on CPU for the full dataset completes in approximately 3 hours, facilitated by a patch-wise approach (16\u0026times;16 pixels, stride 8) and Zarr-based out-of-core data handling. This compact design ensures computational efficiency, practical deployability in resource-constrained settings, and scalability for operational flood hazard assessment over large regions. Prior U-Net flood works emphasize accuracy and richer inputs but rely on GPU training and larger stacks, for example, Li and Demir (\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e2023\u003c/span\u003e) adjusted U-Net on Sentinel-1 with terrain cues was trained on NVIDIA Titan X GPUs and explored 10\u0026ndash;30 m inputs, cross-validation, and transfer learning to boost evaluation, highlighting compute-intensive regimes that our CPU-only pipeline sidesteps via patch-wise optimization and input/output design. While comparative susceptibility studies employing U-Net families report strong accuracies using curated datasets and GPU-oriented training regimes, their workflows emphasize model variants and hyperparameter tuning rather than end-to-end efficiency on modest infrastructure (Melgar-Garc\u0026iacute;a et al. \u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e2023\u003c/span\u003e). In contrast, advanced transformer-augmented architectures like TransUNet demonstrate superior segmentation capabilities but with higher parameters count and incur substantial computational costs and GPU memory usage, while fusion approaches combining U-Net with CNNs achieve high accuracies but typically operate on smaller datasets without addressing large-scale geographic mapping challenges or resource-constrained deployment scenarios that our compact architecture explicitly targets (Anand et al. \u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e2023\u003c/span\u003e; Chen et al. \u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e2024\u003c/span\u003e).\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec32\" class=\"Section2\"\u003e\u003ch2\u003e4.5 Comparison with Global Flood Model Product\u003c/h2\u003e\u003cp\u003eWhen compared with the JRC Global River Flood Hazard Map (100-year return period product derived from large‑scale hydrological and hydrodynamic modelling), the proposed CNN framework showed a critical advantage in detecting flood-prone zones beyond the primary floodplains as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig7\" class=\"InternalRef\"\u003e7\u003c/span\u003ea. The JRC model is constrained by river-network hydrodynamics and often misses non-riverine hazard zones such as confined gullies, alluvial cones, and slope channel junctions that frequently flood in Himalayan settings (Nobile et al., \u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e2025\u003c/span\u003e)\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eThe CNN model, by contrast, integrates both hydrological proximity variables (e.g., distance to rivers, drainage density) and topographic land cover features (e.g., NDVI, TWI, LCF, soil moisture), allowing it to capture subtle and highly localised hazard hotspots missed by purely physics‑driven approaches. Most crucially, the model\u0026rsquo;s susceptibility map closely matched actual 2025 flood locations, including in steep-sided narrow valleys where neither SAR nor hydrodynamic products provided definitive hazard depiction.\u003c/p\u003e\u003c/div\u003e"},{"header":"5. Discussion","content":"\u003cdiv id=\"Sec34\" class=\"Section2\"\u003e\u003ch2\u003e5.1 Comparison with Satellite-Based and Global Flood Hazard products\u003c/h2\u003e\u003cp\u003eCompared with Sentinel-1 SAR, the proposed CNN U-Net model provided more consistent detection across reported flood zones and avoided false negatives caused by radar limitations. Although most mapped locations, using Sentinel-1 SAR, were indeed reported to have faced severe flooding during the 2025 monsoon, the satellite-based analysis could not capture several of these events \u003cb\u003e(\u003c/b\u003eFig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eb to \u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003ef\u003cb\u003e)\u003c/b\u003e. This discrepancy is consistent with known limitations of C-band SAR in mountainous environments. The main issues are: (i) the six-day revisit cycle of Sentinel-1 (Sentinel-1 User Handbook 2013), which often misses short-lived flood peaks; (ii) terrain occlusion and shadowing in steep valleys (Shi et al. \u003cspan citationid=\"CR50\" class=\"CitationRef\"\u003e2024\u003c/span\u003e); (iii) disturbance from vegetation and rough water surfaces that scatter radar signals (Chen and Zhao \u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e2022\u003c/span\u003e; Risling et al. \u003cspan citationid=\"CR42\" class=\"CitationRef\"\u003e2024\u003c/span\u003e); and (iv) additional uncertainties arising from environmental conditions such as rainfall, snowmelt, or landslide activity that further degrade SAR performance (Fakhri and Gkanatsios \u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e2025\u003c/span\u003e). These constraints explain the under-representation of localized floods and highlight the need for complementary susceptibility mapping approaches.\u003c/p\u003e\u003cp\u003eWhen compared to the JRC Global Flood Hazard maps \u003cb\u003e(\u003c/b\u003eFig.\u0026nbsp;\u003cspan refid=\"Fig7\" class=\"InternalRef\"\u003e7\u003c/span\u003e\u003cb\u003e)\u003c/b\u003e, the model offered finer resolution and more realistic identification of Himalayan hotspots such as Dehradun, Mandi, Kangra, and Shimla. The ability to match actual 2025 monsoon flood flashpoints further validate the model\u0026rsquo;s suitability for localised disaster risk reduction, where global hazard products often fail to capture small-scale variability (Tripathi and Mohanty \u003cspan citationid=\"CR66\" class=\"CitationRef\"\u003e2024\u003c/span\u003e).\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec35\" class=\"Section2\"\u003e\u003ch2\u003e5.2 Contribution of LCF and Operational Relevance of CNN U-Net\u003c/h2\u003e\u003cp\u003eAlthough the LCF ranked 7th among 14 predictors in permutation importance (relative importance\u0026thinsp;=\u0026thinsp;0.024, compared to 0.296 for NDVI), its contribution is disproportionately valuable in rugged mountain terrain. By detecting micro-depressions and subtle convex\u0026ndash;concave transitions that concentrate overland flow, LCF captures terrain signatures often underestimated when relying on slope or curvature alone (Liu et al. \u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e2025\u003c/span\u003e). Figure\u0026nbsp;\u003cspan refid=\"Fig7\" class=\"InternalRef\"\u003e7\u003c/span\u003e and Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e depict the noteworthy and innovative contribution of this factor in enhancing the spatially informed predictability of the deep learning model. This improvement is reflected in the performance of the CNN U-Net, which, by leveraging dominant predictors such as NDVI, TWI, and altitude alongside LCF, achieved high classification accuracy (97.12%), CSI (68.44%), and low RMSE (0.145).\u003c/p\u003e\u003cp\u003eThe application of deep learning for flood susceptibility mapping has experienced rapid advancement globally, with numerous studies demonstrating the superiority of CNN-based approaches over traditional statistical and shallow machine learning methods (Ramayanti et al. \u003cspan citationid=\"CR40\" class=\"CitationRef\"\u003e2022\u003c/span\u003e; Ouma and Omai \u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e2023\u003c/span\u003e; Jamali et al. \u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). However, most existing CNN applications in flood mapping have focused primarily on inundation detection using SAR imagery or on susceptibility mapping in plains and deltaic environments, with limited attention to the unique challenges posed by complex mountainous terrain (Li and Demir \u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e2023\u003c/span\u003e; Fakhri and Gkanatsios \u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e2025\u003c/span\u003e; Wang and Feng \u003cspan citationid=\"CR71\" class=\"CitationRef\"\u003e2025\u003c/span\u003e). Unlike single-time satellite captures, the CNN U-Net learns underlying spatial patterns of flood susceptibility, which allows identification of hazard-prone areas even under persistent cloud cover, in topographic shadows, or in zones beyond riverine inundation. This performance underscores the model\u0026rsquo;s advantage in mountainous regions where hydro-geomorphological conditioning dominates flood behaviour. Its operational practicality is equally noteworthy and innovative, as the lightweight architecture of the CNN-U-Net, combined with easily retrainable inputs, positions the framework as highly operational. Once updated predictor datasets are available, the model can be retrained and deployed within hours, allowing susceptibility updates at scales useful for district and state disaster management agencies. Such rapid workflows are particularly relevant for supporting early warning systems, pre-monsoon preparedness, climate-resilient infrastructure planning, and land-use regulation. In contrast, hydrodynamic models often require extensive data, computational resources, and time, making them less practical for rapid decision-making in dynamic Himalayan environments (Karim et al. \u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e2023\u003c/span\u003e).\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec36\" class=\"Section2\"\u003e\u003ch2\u003e5.3 Limitations and Future Directions\u003c/h2\u003e\u003cp\u003eDespite its strong performance, several limitations remain. The model\u0026rsquo;s accuracy is partly dependent on the quality of the flood inventory used for training, which may contain model biases or uncertainties. While LCF improves representation of topography, the current 90 metre resolution may not fully capture very localised floods. Transferability to other physiographic regions, such as plains or deltaic settings, requires further testing.\u003c/p\u003e\u003cp\u003eFuture work should focus on:\u003c/p\u003e\u003cp\u003e\u003cul\u003e\u003cli\u003e\u003cp\u003eEvent-based forecasting by integrating dynamic rainfall and discharge inputs.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eBenchmarking against multiple AI models and CNN architectures to evaluate robustness and identify the optimal framework for flood susceptibility mapping in IHR.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eMulti-hazard extension, linking flood susceptibility with co-occurring hazards such as landslides and GLOFs.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eTransfer learning approaches to reduce retraining needs across regions.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eHigher resolution datasets (30 m or LiDAR-based DEMs) for local planning.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eCoupling susceptibility with exposure and vulnerability layers to produce holistic risk assessments.\u003c/p\u003e\u003c/li\u003e\u003c/ul\u003e\u003c/p\u003e\u003cp\u003eThe study also highlights broader implications for hazard research and practice. The CNN U-Net segmentation framework, augmented with geomorphologically relevant predictors such as the LCF, helps bridge the gap between coarse global flood models and data-intensive hydrodynamic simulations. By generating susceptibility maps that are both accurate and operationally feasible, the framework supports disaster risk reduction efforts under the Sendai Framework and responds to the growing challenge of extreme rainfall and flooding in a changing Himalayan climate. These maps can directly inform early warning systems, land-use planning, and risk mitigation strategies. Moreover, the approach holds considerable potential for advancing multi-hazard susceptibility mapping in other mountainous regions worldwide.\u003c/p\u003e\u003c/div\u003e"},{"header":"6. Conclusion","content":"\u003cp\u003eThis study demonstrates the potential of DL-based flood susceptibility mapping in the complex and data-scarce terrains of the Indian Himalayan Region, with a specific focus on Himachal Pradesh and Uttarakhand. Motivated by the devastating June\u0026ndash;August 2025 monsoon-induced hydro-meteorological events, the study underscores the high flood susceptibility of low-lying valleys and riverine corridors, where settlements and infrastructure are exposed to floods, cloudbursts, and cascading hazards. By employing a CNN with a U-Net architecture and integrating 14 hydro-geomorphological conditioning factors, the approach successfully produced spatially explicit, high-resolution flood susceptibility maps that showed strong alignment with observed flood impacts. Although LCF ranked mid‑tier in overall importance (0.024 relative score), it proved critical for refining hazard delineation in rugged terrain and reducing overestimation in depositional areas, representing a methodological innovation in data‑driven flood risk modelling.\u003c/p\u003e\u003cp\u003eImportantly, the CNN framework overcomes limitations inherent to observation‑driven methods such as Sentinel‑1 SAR rapid mapping, which in Himalayan contexts suffers from infrequent revisit intervals, terrain occlusion, dense vegetation interference, and rapidly changing flood hydrodynamics. This makes the approach inherently operational, supporting early warning systems, land‑use regulation, climate‑resilient infrastructure planning, and pre‑monsoon preparedness.\u003c/p\u003e\u003cp\u003eBeyond flood risk applications, the proposed framework holds promise as a scalable multi‑hazard susceptibility mapping platform for other hydro‑meteorological and geomorphic hazards in mountainous regions. Future progress will hinge on integrating higher‑quality ground observations and more systematic disaster reporting to further strengthen model reliability. In summary, this research underscores the importance of pattern‑informed, stakeholder‑ready flood susceptibility modelling for the IHR, reducing reliance on temporally constrained remote sensing, bridging gaps in traditional hydrodynamic hazard assessments, and enabling precise, high‑resolution risk mapping that is vital for disaster preparedness and long‑term climate resilience. By enabling faster and more accurate susceptibility mapping, the CNN U-Net approach offers a practical decision-support tool for mitigating flood risk in one of the world\u0026rsquo;s most disaster-prone mountain systems.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eAcknowledgements:\u0026nbsp;\u003c/strong\u003eThe authors gratefully acknowledge the ESA Copernicus Hub for Sentinel-1A/B data, IMD for rainfall data, and JRC for providing the river flood hazard map and surface water information. The work presented here is supported by the Department of Science and Technology (Integrated Centre for Adaptation to Climate Change, Disaster Risk Reduction and Sustainability (ICARS)-IIT Roorkee); Project No. DST-2211-WRC/23-24 sanctioned by DST (OM No DST/CCP/NMSKCC/CoE/236/2024(G)] and Indian Space Research Organisation (Project No. STC-2278-WRC-CNA/23-24). We also thank the Indian Institute of Technology Roorkee (IIT Roorkee) and the Department of Water Resources Development and Management (WRD\u0026amp;M)for computational support, especially the PARAM Ganga high-performance computing facility.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eDeclaration of interests:\u0026nbsp;\u003c/strong\u003eThe authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthor contributions:\u0026nbsp;\u003c/strong\u003eRachit: Conceptualization, Methodology, Data Collection, Data Analysis, Writing—original draft, Writing—review and editing; \u0026nbsp;Vaibhav Tripathi: Conceptualization, Methodology, Data Analysis, Writing—review and editing; Mohit Prakash Mohanty: Conceptualization, Methodology, Writing—review and editing; Supervision; Ashish Pandey: Supervision, Project Administration; Anil Kumar Gupta: Supervision, Project Administration.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eAlam MG, Tripathi V, Bhatt CM, Mohanty MP (2025) A Novel Framework Embedding Bayesian-optimized Ensemble Machine Learning and Explainable Artificial Intelligence (XAI) to Improve Flood Prediction in Complex Watersheds. Environ Sustain Indicat 100760. https://doi.org/10.1016/j.indic.2025.100760\u003c/li\u003e\n\u003cli\u003eAmiri A, Soltani K, Ebtehaj I, Bonakdari H (2024) A novel machine learning tool for current and future flood susceptibility mapping by integrating remote sensing and geographic information systems. J Hydrol 632:130936. https://doi.org/10.1016/j.jhydrol.2024.130936\u003c/li\u003e\n\u003cli\u003eAnand V, Gupta S, Koundal D, Singh K (2023) Fusion of U-Net and CNN model for segmentation and classification of skin lesion from dermoscopy images. Expert Syst Appl 213:119230. https://doi.org/10.1016/j.eswa.2022.119230\u003c/li\u003e\n\u003cli\u003eBaugh C, Colonese J, D\u0026apos;Angelo C, Dottori F, Neal J, Prudhomme C, Salamon P (2024) Global river flood hazard maps. European Commission, Joint Research Centre (JRC). http://data.europa.eu/89h/jrc-floods-floodmapgl_rp50y-tif \u003c/li\u003e\n\u003cli\u003eBentivoglio R, Isufi E, Jonkman SN, Taormina R (2022) Deep learning methods for flood mapping: A review of existing applications and future research directions. Hydrol Earth Syst Sci 26(16):4345\u0026ndash;4378. https://doi.org/10.5194/hess-26-4345-2022\u003c/li\u003e\n\u003cli\u003eBiazar SM, Golmohammadi G, Nedhunuri RR, Shaghaghi S, Mohammadi K (2025) Artificial Intelligence in Hydrology: Advancements in Soil, Water Resource Management, and Sustainable Development. Sustainability 17(5):2250. https://doi.org/10.3390/su17052250\u003c/li\u003e\n\u003cli\u003eBreznik J, O\u0026scaron;tir K, Rak G (2025) The potential of Sentinel-1 imagery for flood event detection: A satellite vs. hydraulic model comparison. J Hydrol 651:132587. https://doi.org/10.1016/j.jhydrol.2024.132587\u003c/li\u003e\n\u003cli\u003eCEEW (2021) 85% Districts in Uttarakhand Vulnerable to Extreme Floods. https://www.ceew.in/press-releases/85-districts-uttarakhand-vulnerable-extreme-floods-ceew. Accessed 18 Aug 2025\u003c/li\u003e\n\u003cli\u003eChen J, Mei J, Li X, Lu Y, Yu Q, Wei Q, Lu X, Xie Y, Adeli E, Wang Y, Lungren MP, Zhang S, Xing L, Lu L, Yuille A, Zhou Y (2024) TransUNet: Rethinking the U-Net architecture design for medical image segmentation through the lens of transformers. Med Image Anal 97:103280. https://doi.org/10.1016/j.media.2024.103280\u003c/li\u003e\n\u003cli\u003eChen Z, Zhao S (2022) Automatic monitoring of surface water dynamics using Sentinel-1 and Sentinel-2 data with Google Earth Engine. Int J Appl Earth Obs Geoinf 113:103010. https://doi.org/10.1016/j.jag.2022.103010\u003c/li\u003e\n\u003cli\u003eDey H, Haque MM, Shao W, VanDyke M, Hao F (2024) Simulating flood risk in Tampa Bay using a machine learning driven approach. npj Nat Hazards 1(1):40. https://doi.org/10.1038/s44304-024-00045-4\u003c/li\u003e\n\u003cli\u003eDhote PR, Joshi Y, Rajib A, Thakur PK, Nikam BR, Aggarwal SP (2023) Evaluating topography-based approaches for fast floodplain mapping in data-scarce complex-terrain regions: Findings from a Himalayan basin. J Hydrol\u003cem\u003e 620:\u003c/em\u003e129309. https://doi.org/10.1016/j.jhydrol.2023.129309\u003c/li\u003e\n\u003cli\u003eDown To Earth (2025) With two months of monsoon left, will Himachal create new records of natural disasters? https://www.downtoearth.org.in/natural-disasters/with-two-months-of-monsoon-left-will-himachal-create-new-records-of-natural-disasters. Accessed 18 Aug 2025\u003c/li\u003e\n\u003cli\u003eEdamo ML, Ayele EG, Yisihak Ukumo T, Alemayehu Kassaye A, Paulos Haile A (2024) Capability of logistic regression in identifying flood-susceptible areas in a small watershed. H2Open J \u003cem\u003e7\u003c/em\u003e(5):351\u0026ndash;374. https://doi.org/10.2166/h2oj.2024.024\u003c/li\u003e\n\u003cli\u003eFakhri F, Gkanatsios I (2025) Quantitative evaluation of flood extent detection using attention U-Net case studies from Eastern South Wales Australia in March 2021 and July 2022. Sci Rep 15(1):12377. https://doi.org/10.1038/s41598-025-92734-x\u003c/li\u003e\n\u003cli\u003eFereshtehpour M, Esmaeilzadeh M, Alipour RS, Burian SJ (2024) Impacts of DEM type and resolution on deep learning-based flood inundation mapping. Earth Sci Inform 17(2):1125\u0026ndash;1145. https://doi.org/10.1007/s12145-024-01239-0\u003c/li\u003e\n\u003cli\u003eGupta V, Syed B, Pathania A, Raaj S, Nanda A, Awasthi S, Shukla DP (2024) Hydrometeorological analysis of July-2023 floods in Himachal Pradesh, India. Nat Hazards 120(8):7549\u0026ndash;7574. https://doi.org/10.1007/s11069-024-06520-5\u003c/li\u003e\n\u003cli\u003eIndia Today (2024) Flash flood risk in Himachal, Uttarakhand as glacial lakes in Himalayas rise: Report. https://www.indiatoday.in/environment/story/flash-flood-risk-himachal-uttarakhand-glacial-lakes-himalayas-rise-2601325-2024-09-17 Accessed 30 Aug 2025\u003c/li\u003e\n\u003cli\u003eIndia Today (2025) Nature\u0026rsquo;s fury in numbers: A look at Uttarakhand\u0026rsquo;s disaster record. https://www.indiatoday.in/diu/story/natures-fury-in-numbers-a-look-at-uttarakhand-disaster-record-2769669-2025-08-11. Accessed 18 Aug 2025\u003c/li\u003e\n\u003cli\u003eJamali A, Roy SK, Beni LH, Pradhan B, Li J, Ghamisi P (2024) Residual wave vision U-Net for flood mapping using dual polarization Sentinel-1 SAR imagery. Int J Appl Earth Obs Geoinf 127:103662. https://doi.org/10.1016/j.jag.2024.103662\u003c/li\u003e\n\u003cli\u003eKansal ML, Singh S (2022) Flood Management Issues in Hilly Regions of Uttarakhand (India) under Changing Climatic Conditions. Water 14(12):1879. https://doi.org/10.3390/w14121879\u003c/li\u003e\n\u003cli\u003eKarim F, Armin MA, Ahmedt-Aristizabal D, Tychsen-Smith L, Petersson L (2023) A Review of Hydrodynamic and Machine Learning Approaches for Flood Inundation Modeling. Water 15(3):566. https://doi.org/10.3390/w15030566\u003c/li\u003e\n\u003cli\u003eKhoirunisa N, Ku CY, Liu CY (2021) A GIS-Based Artificial Neural Network Model for Flood Susceptibility Assessment. Int J Environ Res Public Health 18(3):1072. https://doi.org/10.3390/ijerph18031072\u003c/li\u003e\n\u003cli\u003eKhosravi K, Rezaie F, Cooper JR, Kalantari Z, Abolfathi S, Hatamiafkoueieh J (2023) Soil water erosion susceptibility assessment using deep learning algorithms. J Hydrol 618:129229 https://doi.org/10.1016/j.jhydrol.2023.129229\u003c/li\u003e\n\u003cli\u003eKonapala G, Kumar SV, Khalique Ahmad S (2021) Exploring Sentinel-1 and Sentinel-2 diversity for flood inundation mapping using deep learning. ISPRS J Photogramm Remote Sens 180:163\u0026ndash;173. https://doi.org/10.1016/j.isprsjprs.2021.08.016\u003c/li\u003e\n\u003cli\u003eKumar V, Solanki YP, Sharma KV, Patel A, Tiwari DK, Mehta DJ (2024) India\u0026rsquo;s flood risk assessment and mapping with multi-criteria decision analysis and GIS integration. J Water Clim Change. https://doi.org/10.2166/wcc.2024.054\u003c/li\u003e\n\u003cli\u003eLi Z, Demir I (2023). U-net-based semantic classification for flood extent extraction using SAR imagery and GEE platform: A case study for 2019 central US flooding. Sci Total Environ 869:161757. https://doi.org/10.1016/j.scitotenv.2023.161757\u003c/li\u003e\n\u003cli\u003eLiu J, Zhao X, Chen Y, Sun H, Gu Y, Xu S (2025) A novel flood conditioning factor based on topography for flood susceptibility modeling. Geosci Front 16(1):101960. https://doi.org/10.1016/j.gsf.2024.101960\u003c/li\u003e\n\u003cli\u003eMelgar-Garc\u0026iacute;a L, Mart\u0026iacute;nez-\u0026Aacute;lvarez F, Bui DT, Troncoso A (2023) A novel semantic segmentation approach based on U-Net, WU-Net, and U-Net++ deep learning for predicting areas sensitive to pluvial flood at tropical area. International Journal of Digital Earth. https://doi.org/10.1080/17538947.2023.2252401\u003c/li\u003e\n\u003cli\u003eMerz B, Bl\u0026ouml;schl G, Vorogushyn S, Dottori F, Aerts JCJH, Bates P, Bertola M, Kemter M, Kreibich H, Lall U, Macdonald E (2021) Causes, impacts and patterns of disastrous river floods. Nat Rev Earth Environ 2(9):592\u0026ndash;609. https://doi.org/10.1038/s43017-021-00195-3\u003c/li\u003e\n\u003cli\u003eNagamani K, Mishra AK, Meer MS, Das J (2024) Understanding flash flooding in the Himalayan Region: A case study. Sci Rep 14(1):7060. https://doi.org/10.1038/s41598-024-53535-w\u003c/li\u003e\n\u003cli\u003eNemni E, Bullock J, Belabbes S, Bromley L (2020) Fully Convolutional Neural Network for Rapid Flood Segmentation in Synthetic Aperture Radar Imagery. Remote Sens 12(16):2532. https://doi.org/10.3390/rs12162532\u003c/li\u003e\n\u003cli\u003eNobile EGL, Figueiredo R, Arrighi C, Rom\u0026atilde;o X, Martina MLV (2025) Flood risk assessment of cultural heritage across countries and spatial scales. Int J Disaster Risk Reduct 118:105236. https://doi.org/10.1016/j.ijdrr.2025.105236\u003c/li\u003e\n\u003cli\u003eOmbadi M, Risser MD, Rhoades AM, Varadharajan C (2023) A warming-induced reduction in snow fraction amplifies rainfall extremes. Nature 619(7969):305\u0026ndash;310. https://doi.org/10.1038/s41586-023-06092-7\u003c/li\u003e\n\u003cli\u003eOuma YO, Omai L (2023) Flood Susceptibility Mapping Using Image-Based 2D-CNN Deep Learning: Overview and Case Study Application Using Multiparametric Spatial Data in Data-Scarce Urban Environments. Int J Intell Syst 2023(1):5672401. https://doi.org/10.1155/2023/5672401\u003c/li\u003e\n\u003cli\u003ePareta K (2024) 1D-2D hydrodynamic and sediment transport modelling using MIKE models. Discover Water 4(1):94. https://doi.org/10.1007/s43832-024-00130-9\u003c/li\u003e\n\u003cli\u003ePatel P, Thakur PK, Aggarwal SP, Garg V, Dhote PR, Nikam BR, Swain S, Al-Ansari N (2022) Revisiting 2013 Uttarakhand flash floods through hydrological evaluation of precipitation data sources and morphometric prioritization. Geomat Nat Hazards Risk 13(1):646-666. https://doi.org/10.1080/19475705.2022.2038696\u003c/li\u003e\n\u003cli\u003ePepin NC, Arnone E, Gobiet A, Haslinger K, Kotlarski S, Notarnicola C, Palazzi E, Seibert P, Serafin S, Sch\u0026ouml;ner W, Terzago S, Thornton JM, Vuille M, Adler C (2022) Climate Changes and Their Elevational Patterns in the Mountains of the World. Rev Geophys 60(1):e2020RG000730. https://doi.org/10.1029/2020RG000730\u003c/li\u003e\n\u003cli\u003ePreventionWeb (2025) Over 45% of Himachal Pradesh at risk of landslides, floods. https://www.preventionweb.net/news/over-45-himachal-pradesh-risk-landslides-floods. Accessed 18 Aug 2025\u003c/li\u003e\n\u003cli\u003eRamayanti S, Nur AS, Syifa M, Panahi M, Achmad AR, Park S, Lee, CW (2022) Performance comparison of two deep learning models for flood susceptibility map in Beira area, Mozambique. Egypt J Remote Sens Space Sci 25(4):1025\u0026ndash;1036. https://doi.org/10.1016/j.ejrs.2022.11.003\u003c/li\u003e\n\u003cli\u003eRick B, McGrath D, McCoy SW, Armstrong WH (2023) Unchanged frequency and decreasing magnitude of outbursts from ice-dammed lakes in Alaska. Nature Commun 14(1):6138. https://doi.org/10.1038/s41467-023-41794-6\u003c/li\u003e\n\u003cli\u003eRisling A, Lindersson S, Brandimarte L (2024) A comparison of global flood models using Sentinel-1 and a change detection approach. Nat Hazards 120(12):11133\u0026ndash;11152. https://doi.org/10.1007/s11069-024-06629-7\u003c/li\u003e\n\u003cli\u003eRoul AR, Behera AK, Pradhan RM, Pradhan SP (2025) Navigating Hydrological Factors in Understanding and Mitigating Landslide Hazards in the Indian Himalayas: A Short Review. Landslides: Analysis, Modeling and Mitigation 3\u0026ndash;24. https://doi.org/10.1007/978-3-031-78896-3_1\u003c/li\u003e\n\u003cli\u003eSafaei-Moghadam A, Tarboton D, Minsker B (2023) Estimating the likelihood of roadway pluvial flood based on crowdsourced traffic data and depression-based DEM analysis. Nat Hazards Earth Syst Sci 23(1):1\u0026ndash;19. https://doi.org/10.5194/nhess-23-1-2023\u003c/li\u003e\n\u003cli\u003eSaravanan S, Abijith D, Reddy NM, Kss P, Janardhanam N, Sathiyamurthi S, Sivakumar V (2023) Flood susceptibility mapping using machine learning boosting algorithms techniques in Idukki district of Kerala India. Urban Clim 49:101503. https://doi.org/10.1016/j.uclim.2023.101503\u003c/li\u003e\n\u003cli\u003eSajwan KS, Sushil K (2016) A Geological Appraisal of Slope Instability in Upper Alaknanda Valley, Uttarakhand Himalaya, India. J Geol Geophys 5:1-7. https://doi.org/10.4172/2381-8719.1000258\u003c/li\u003e\n\u003cli\u003eSchl\u0026ouml;gl M, Fuchs S, Scheidl C, Heiser M (2021) Trends in torrential flooding in the Austrian Alps: A combination of climate change, exposure dynamics, and mitigation measures. Clim Risk Manag 32:100294. https://doi.org/10.1016/j.crm.2021.100294\u003c/li\u003e\n\u003cli\u003eSentinel-1 Team (2013) Sentinel-1 User Handbook. https://ftp.itc.nl/pub/Dragon4_Lecturer_2018/ESA%20EEs%20and%20Sentinels%20Brochures%20pdf/Copernicus%20Sentinels%201,2,3/Sentinel-1%20User%20Handbook.pdf. Accessed 17 Aug 2025\u003c/li\u003e\n\u003cli\u003eShah A, Malakar K (2024) Climate-change-induced risk mapping of the Indian Himalayan districts using the latest IPCC framework. Int J Disaster Risk Reduct 102:104283. https://doi.org/10.1016/j.ijdrr.2024.104283\u003c/li\u003e\n\u003cli\u003eShi C, Zuo X, Zhang J, Zhu D, Li Y, Bu J (2024) Accuracy Assessment of Geometric-Distortion Identification Methods for Sentinel-1 Synthetic Aperture Radar Imagery in Highland Mountainous Regions. Sensors (Basel, Switzerland) 24(9):2834. https://doi.org/10.3390/s24092834\u003c/li\u003e\n\u003cli\u003eSong W, Guan M, Guo K, Yu D (2025) Rapid flood inundation mapping by integrating deep learning-based image super-resolution with coarse-grid hydrodynamic modeling. Eng Appl Comput Fluid Mech 19(1):2481115. https://doi.org/10.1080/19942060.2025.2481115\u003c/li\u003e\n\u003cli\u003eSphere India (2023) Situation Assessment Report Himachal Pradesh Floods Jul 2023. https://www.sphereindia.org.in/sites/default/files/2023-10/Situation%20Report_Himachal%20Pradesh_Floods_Jul23.pdf. Accessed 17 Aug 2025\u003c/li\u003e\n\u003cli\u003eSreevalsan-Nair J, Mundayatt A (2025) Evolution of Data-driven Single- and Multi-Hazard Susceptibility Mapping and Emergence of Deep Learning Methods (No. arXiv:2502.09045). arXiv. https://doi.org/10.48550/arXiv.2502.09045 \u003c/li\u003e\n\u003cli\u003eSultan H, Zhan J, Rashid W, Chu X, Bohnett E (2022) Systematic Review of Multi-Dimensional Vulnerabilities in the Himalayas. Int J Environ Res Public Health 19(19):12177. https://doi.org/10.3390/ijerph191912177\u003c/li\u003e\n\u003cli\u003eTarpanelli A, Mondini AC, Camici S (2022) Effectiveness of Sentinel-1 and Sentinel-2 for flood detection assessment in Europe. Nat Hazards Earth Syst Sci 22(8):2473\u0026ndash;2489. https://doi.org/10.5194/nhess-22-2473-2022\u003c/li\u003e\n\u003cli\u003eTavus B, Can, R, Kocaman S (2022) A cnn-based flood mapping approach using Sentinel-1 data. ISPRS Ann Photogramm Remote Sens Spatial Inf Sci 3:549-556. https://doi.org/10.5194/isprs-annals-V-3-2022-549-2022\u003c/li\u003e\n\u003cli\u003eTaysi H, Yang YCE, Gangrade S, Chegini T, Kao SC, Li HY (2025) Enhancing 2D hydrodynamic flood models through machine learning and urban drainage integration. J Hydrol 659:133258. https://doi.org/10.1016/j.jhydrol.2025.133258\u003c/li\u003e\n\u003cli\u003eThe Economic Times (2025) Severe floods experienced in Beas river following heavy rainfall in Himachal Pradesh\u0026rsquo;s Mandi.https://economictimes.indiatimes.com/news/india/severe-floods-experienced-in-beas-river-following-heavy-rainfall-in-himachal-pradeshs-mandi/articleshow/122196209.cms?from=mdr. Accessed 17 Aug 2025\u003c/li\u003e\n\u003cli\u003eThe Hindu(2025) Uttarkashi flash floods highlights: Rescuers retrieve two bodies in flood-hit Dharali; at least 190 rescued. https://www.thehindu.com/news/national/uttarakhand/uttarakhand-uttarkashi-cloudburst-flash-floods-highlights-august-6-2025/article69896766.ece. Accessed 17 Aug 2025\u003c/li\u003e\n\u003cli\u003eThe Logical Indian(2025) Himachal Pradesh: Cloudburst in Kullu\u0026rsquo;s Sainj Valley Sparks Flash Floods, Cuts Off Villages - No Casualties So Far. https://thelogicalindian.com/himachal-pradesh-cloudburst-in-kullus-sainj-valley-sparks-flash-floods-cuts-off-villages-no-casualties-so-far/. Accessed 17 Aug 2025\u003c/li\u003e\n\u003cli\u003eThe Times of India (2025a) 705 deaths in 10 years: Multiple studies warn of flash floods emerging as major killer in Uttarakhand; climate change, extreme weather rewriting Himalayan hydrology. https://timesofindia.indiatimes.com/city/dehradun/705-deaths-in-10-yrs-multiple-studies-warn-of-flash-floods-emerging-as-major-killer-in-ukhand/articleshow/123206705.cms. Accessed 18 Aug 2025\u003c/li\u003e\n\u003cli\u003eThe Times of India(2025b) Cloudburst, heavy rains wreak havoc in Uttarakhand: Char Dham yatra halted, workers missing, landslides hit highways \u0026ndash; 10 key updates. https://timesofindia.indiatimes.com/india/cloudburst-heavy-rains-wreak-havoc-in-uttarakhand-char-dham-yatra-halted-workers-missing-landslides-hit-highways-10-key-updates/articleshow/122138901.cms. Accessed 17 Aug 2025\u003c/li\u003e\n\u003cli\u003eThe Times of India (2025c) Uttarkashi cloudburst havoc: At least 4 dead as flash floods swallow homes, hotels in Dharali; entire village washed away. https://timesofindia.indiatimes.com/city/dehradun/uttarakhand-cloudburst-flash-floods-hit-uttarkashi-several-villagers-washed-away/articleshow/123115132.cms. Accessed 17 Aug 2025\u003c/li\u003e\n\u003cli\u003eThe Times of India (2024) Natural disasters killed 358 across HP this year. https://timesofindia.indiatimes.com/city/shimla/himachal-pradesh-faces-catastrophic-year-with-358-lives-lost-to-natural-disasters/articleshow/115950309.cms. Accessed 18 Aug 2025\u003c/li\u003e\n\u003cli\u003eTripathi V, Mohanty MP, Singh H (2023) Fidelity of machine learning models in capturing flood inundation through geomorphic descriptors over Ganga sub‑basin, India. 12th World Con‑gress on Water Resources and Environment (EWRA 2023), Thessaloniki, Greece.\u003c/li\u003e\n\u003cli\u003eTripathi V, Mohanty MP (2024) Can geomorphic flood descriptors coupled with machine learning models enhance in quantifying flood risks over data-scarce catchments? Development of a hybrid framework for Ganga basin (India). Environ. Sci Pollut Res 1-20. https://doi.org/10.1007/s11356-024-33507-3 \u003c/li\u003e\n\u003cli\u003eUca, Lamada M, Nyompa S, Haris NA (2022) Morphometry and Topographic Wetness Index Analysis for flood inundation mapping in Mata Allo watershed (South Sulawesi, Indonesia). Forum Geogr XXI(2):133\u0026ndash;145. https://doi.org/10.5775/fg.2022.105.d\u003c/li\u003e\n\u003cli\u003eUddin K, Matin MA, Meyer FJ (2019) Operational Flood Mapping Using Multi-Temporal Sentinel-1 SAR Images: A Case Study from Bangladesh. Remote Sens 11(13):1581. https://doi.org/10.3390/rs11131581\u003c/li\u003e\n\u003cli\u003eUllah K, Wang Y, Fang Z, Wang L, Rahman M (2022) Multi-hazard susceptibility mapping based on Convolutional Neural Networks. Geosci Front 13(5):101425. https://doi.org/10.1016/j.gsf.2022.101425\u003c/li\u003e\n\u003cli\u003eUN Office for Disaster Risk Reduction (UNDRR), Centre for Research on the Epidemiology of Disasters (CRED)(2022) \u0026quot;Human cost of disasters - An overview of the last 20 years 2000-2019\u0026quot;. https://reliefweb.int/sites/reliefweb.int/files/resources/Human%20Cost%20of%20Disasters%202000-2019%20Report%20-%20UN%20Office%20for%20Disaster%20Risk%20Reduction.pdf. Accessed 18 Aug 2025\u003c/li\u003e\n\u003cli\u003eWang F, Feng X (2025) Flood change detection model based on an improved U-net network and multi-head attention mechanism. Sci Rep 15(1):3295. https://doi.org/10.1038/s41598-025-87851-6\u003c/li\u003e\n\u003cli\u003eWang Y, Fang Z, Hong H, Peng L (2020) Flood susceptibility mapping using convolutional neural network frameworks. J Hydrol 582:124482. https://doi.org/10.1016/j.jhydrol.2019.124482\u003c/li\u003e\n\u003cli\u003eZhao G, Pang B, Xu Z, Yue J, Tu T (2018) Mapping flood susceptibility in mountainous areas on a national scale in China. Sci Total Environ 615:1133\u0026ndash;1142. https://doi.org/10.1016/j.scitotenv.2017.10.037\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":true,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"natural-hazards","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"nhaz","sideBox":"Learn more about [Natural Hazards](https://www.springer.com/journal/11069)","snPcode":"11069","submissionUrl":"https://submission.nature.com/new-submission/11069/3","title":"Natural Hazards","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"em","reportingPortfolio":"Springer Hybrid","inReviewEnabled":true,"inReviewRevisionsEnabled":false},"keywords":"CNN, Deep learning, Flood, Indian Himalayas, Multi-Hazard, Susceptibility, U-net","lastPublishedDoi":"10.21203/rs.3.rs-7715126/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-7715126/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eThe Indian Himalayan Region (IHR) is increasingly threatened by hydro-meteorological hazards such as cloudbursts, flash floods, and landslides, driven by climatic extremes and rapid land-use changes. The June\u0026ndash;August 2025 monsoon floods in the states of Himachal Pradesh and Uttarakhand in India highlighted this vulnerability, causing severe loss of life and infrastructure damage. To address the urgent need for rapid and reliable information tools during the crucial initial hours of disaster occurrence in complex, data-scarce environments, this study develops a rapidly deployable deep learning-based flood susceptibility mapping framework tailored for the IHR. The framework employs a Convolutional Neural Network (CNN) with U-Net architecture, integrating 14 hydro-geomorphological predictors (e.g., altitude, slope, TWI, NDVI). A novel Local Convexity Factor (LCF), adaptively calibrated using curvature and slope, enhances micro-topographic characterisation, improving hazard delineation in rugged landscapes and reducing overestimation in depositional zones. The model achieved strong predictive skill (accuracy\u0026thinsp;=\u0026thinsp;97.12%, RMSE\u0026thinsp;\u0026asymp;\u0026thinsp;0.145, CSI\u0026thinsp;=\u0026thinsp;68.44%, AUC ROC\u0026thinsp;\u0026asymp;\u0026thinsp;0.99), demonstrating high predictive reliability despite limited flood observations. Compared to Sentinel-1 SAR and the JRC Global Flood Hazard Map, the CNN\u0026ndash;U-Net approach effectively captures both riverine and upland flood hotspots by understanding the hidden patterns in catchment physiography. Designed for rapid retraining and deployment within hours, the framework functions as an operational, stakeholder-ready information tool, supporting early warning, land-use planning, and climate-resilient infrastructure development. Beyond flood mapping, the proposed framework can be utilized for multi-hazard susceptibility mapping in other mountainous and data-scarce regions worldwide.\u003c/p\u003e","manuscriptTitle":"Rapid Flood Susceptibility Mapping in the Indian Himalayan Region using CNN-U-Net Segmentation: Insights from the 2025 Monsoon Events","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-10-29 07:26:00","doi":"10.21203/rs.3.rs-7715126/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Major revisions","date":"2026-02-22T13:00:51+00:00","index":"","fulltext":""},{"type":"reviewerAgreed","content":"","date":"2025-10-15T16:09:01+00:00","index":0,"fulltext":""},{"type":"reviewersInvited","content":"","date":"2025-10-15T01:22:28+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2025-09-29T09:11:54+00:00","index":"","fulltext":""},{"type":"submitted","content":"Natural Hazards","date":"2025-09-26T03:09:44+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"natural-hazards","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"nhaz","sideBox":"Learn more about [Natural Hazards](https://www.springer.com/journal/11069)","snPcode":"11069","submissionUrl":"https://submission.nature.com/new-submission/11069/3","title":"Natural Hazards","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"em","reportingPortfolio":"Springer Hybrid","inReviewEnabled":true,"inReviewRevisionsEnabled":false}}],"origin":"","ownerIdentity":"f5cb2d05-5491-499a-9a7a-4ca6c893b55f","owner":[],"postedDate":"October 29th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"under-review","subjectAreas":[],"tags":[],"updatedAt":"2026-04-20T15:47:38+00:00","versionOfRecord":[],"versionCreatedAt":"2025-10-29 07:26:00","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-7715126","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-7715126","identity":"rs-7715126","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}