Multi-Modal Flood Detection and Validation: Integrating Sentinel-1 SAR, Deep Learning, and Named Entity Recognition for Enhanced Situational Awareness

Urban flooding is a recurring problem in densely populated cities, particularly during intense monsoon periods, leading to severe disruptions in infrastructure and daily life. To address this, the study presents a multi-modal flood detection framework that integrates remote sensing, machine learning, and natural language processing (NLP). Sentinel-1 SAR data was processed using VH-polarized backscatter differences and Otsu’s thresholding to identify inundated regions. False positives were reduced through the application of Digital Elevation Models (DEM) and Global Surface Water (GSW) masks. Classification was performed using Vision Transformers (ViTs), Convolutional Neural Networks (CNNs), XGBoost, and Random Forest, with ViTs achieving the highest accuracy of 96%. An NLP module further validated results using geotagged news and social media reports. The framework was tested on pre and post-flood Sentinel-1 images from June 29 and July 11, 2024, in Mumbai. The results confirm improved detection reliability, offering a scalable and real-time solution for urban flood monitoring and emergency response.