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

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.