AI Embeddings-based Flood Susceptibility Mapping in Kerala: A U-Net Deep Learning Approach

Kerala experienced catastrophic flooding in 2018, and reliable flood susceptibility maps are still needed for the region. The Proposed research presents a deep learning approach using lightweight U-Net model that generates 30-meter resolution susceptibility maps from 46 remote sensing features: 10 environmental variable bands, 7 SAR bands, and 29 Google Earth Engine deep learning embeddings, covering 156.5 million pixels across Kerala. Only 0.13% of pixels in the dataset were labelled as flood, so we used focal loss optimization and spatial sampling to handle the imbalance. The model reached 96.55% overall accuracy and 41.52% precision, and detected flood-prone areas more reliably than traditional approaches. Spatial validation confirmed that the predictions aligned with known flood corridors along major river systems and coastal lowlands. The generated susceptibility maps support flood risk assessment, early warning systems, and urban planning. The proposed research demonstrates a scalable AIdriven approach to disaster management in data-scarce, topographically complex tropical regions.