Mishra et al. (2026) Lightweight Hybrid Deep Learning and Fuzzy-AHP Framework for Predictive Flood Susceptibility Mapping in the Ghaghara River Basin, India: A Data-Driven Approach for Enhanced Spatiotemporal Precision and Risk Prediction

Identification

• Journal: Earth Systems and Environment
• Year: 2026
• Date: 2026-01-13
• Authors: Priya Mishra, Sanjeev Kr. Prasad
• DOI: 10.1007/s41748-025-00970-y

Research Groups

School of Computer Science and Engineering, Galgotias University, Gr.Noida, U.P, India

Short Summary

This study proposes a novel hybrid deep learning framework (FAHP-CNN-LSTM) for predictive flood susceptibility mapping in the Ghaghara River Basin, India, demonstrating enhanced spatiotemporal precision and risk prediction compared to standalone models. The framework integrates expert-driven Fuzzy-AHP weighting with a CNN-LSTM deep learning architecture to identify and map flood-prone areas into five risk classes.

Objective

• To develop and validate a novel hybrid framework integrating Fuzzy-AHP, Pareto analysis, and a CNN–LSTM deep learning model for enhanced spatiotemporal precision and risk prediction in flood susceptibility mapping in the Ghaghara River Basin, India.

Study Configuration

• Spatial Scale: Ghaghara River Basin, India, specifically focusing on the districts of Gorakhpur, Maharaj Ganj, and Siddharth Nagar in Uttar Pradesh. The total mapped area for flood-prone zones is 9302 square kilometers.
• Temporal Scale: Historical flood information and data from the monsoon seasons between 2020 and 2025 were used for performance evaluation and validation. Average rainfall data spanned from 2012 to 2023.

Methodology and Data

• Models used:
  • Fuzzy Analytic Hierarchy Process (Fuzzy-AHP) for expert-informed factor weighting.
  • Pareto analysis for initial factor selection.
  • Hybrid Convolutional Neural Network (CNN) - Long Short-Term Memory (LSTM) deep learning model, with MobileNetV2 as the CNN architecture for spatial feature extraction.
  • Weighted Sum Model (WSM) for initial susceptibility map generation.
  • Variance Inflation Factor (VIF) analysis for multicollinearity reduction.
• Data sources:
  • Digital Elevation Model (DEM) and elevation datasets from NASA and the United States Geological Survey (USGS).
  • Land Use/Land Cover (LULC) information from Landsat-8 imagery.
  • Satellite-based precipitation products (e.g., CHIRPS, approximately 5.6 km resolution) and event-based rainfall data from NASA platforms and Google Earth Engine (GEE).
  • Runoff maps and other hydrological data prepared using Google Earth Engine.
  • Soil and ancillary environmental datasets from global satellite-derived databases via GEE.
  • Historical flood inventory data (2020–2025) from the Remote Sensing Applications Centre, Lucknow, Uttar Pradesh, India.
  • Expert opinion survey for Fuzzy-AHP and Pareto analysis.
  • Flood-conditioning factors (9 selected from 21): elevation, slope, LULC, soil type, Topographic Wetness Index (TWI), distance from river, drainage density, event-based rainfall, and runoff.

Main Results

• The proposed hybrid FAHP-CNN-LSTM model significantly outperformed standalone CNN-LSTM and Fuzzy-AHP models.
• Performance metrics for the proposed model: training accuracy of 94.56%, validation accuracy of 92.36%, ROC-AUC of 96.78%, and F1-score of 91.28%.
• Flood susceptibility maps were classified into five categories (very high, high, moderate, low, very low), covering a total area of 9302 square kilometers.
• The proposed model identified a larger proportion of high-risk areas (51.38%) compared to other models.
• Spatial validation using the Kappa Coefficient (κ) and k-fold cross-validation confirmed the robustness and reliability of the model.
• The most flood-prone regions were consistently identified as areas of low elevation and proximity to rivers, particularly in the southern zones of Gorakhpur district near the Rapti River and its tributaries.

Contributions

• Introduction of a novel lightweight hybrid deep learning framework (FAHP-CNN-LSTM) for flood susceptibility mapping, specifically tailored for the Ghaghara River Basin.
• Integration of expert-driven knowledge (Fuzzy-AHP) with data-driven deep learning (CNN-LSTM) to enhance model interpretability, accuracy, and robustness in capturing complex spatiotemporal flood dynamics.
• Application of Pareto analysis to systematically identify and select nine key flood-influencing factors from an initial set of 21, optimizing model complexity and preventing overfitting/underfitting.
• Utilization of both static (topographical, land-surface, pedological, geomorphological) and dynamic (meteorological, hydrological) variables, including event-based rainfall and runoff, for comprehensive flood assessment.
• Implementation of spatial validation with three folds and buffer zones to ensure model generalizability and prevent spatial leakage, alongside k-fold cross-validation for robust evaluation.
• Incorporation of MobileNetV2 as the CNN component for efficient and fast spatial feature extraction, making the model suitable for real-time or resource-constrained applications.
• Generation of detailed, five-class flood susceptibility maps that provide practical insights for early warning systems, risk-informed land-use planning, and disaster mitigation strategies in the Ghaghara Basin.

Funding

This work did not receive funding from any source.

Citation

@article{Mishra2026Lightweight,
  author = {Mishra, Priya and Prasad, Sanjeev Kr.},
  title = {Lightweight Hybrid Deep Learning and Fuzzy-AHP Framework for Predictive Flood Susceptibility Mapping in the Ghaghara River Basin, India: A Data-Driven Approach for Enhanced Spatiotemporal Precision and Risk Prediction},
  journal = {Earth Systems and Environment},
  year = {2026},
  doi = {10.1007/s41748-025-00970-y},
  url = {https://doi.org/10.1007/s41748-025-00970-y}
}