IoT-Integrated Edge AI for Night-Time Rural Flood Detection Using Thermal and Low-Light Image Classification

Existing flood early warning systems fail rural communities, particularly at night and during extreme weather when visibility is poor and communication infrastructure is limited. We present an Edge AI flood detection solution integrating IoT technology for resource-constrained rural environments. The system uses inexpensive edge nodes equipped with thermal and low-light-capable visible cameras. A custom two-stream fusion CNN runs directly on edge devices: one stream analyzes thermal imagery for water detection via emissivity contrast, while the other processes low-light-enhanced visible images. Cross-attention fusion classifies conditions into Normal, Warning, and Flood states. We pre-train on the SEN12-FLOOD satellite dataset and fine-tune on field imagery from rural riverine and agricultural regions. The system achieves real-time notifications under 2-second latency and operates offline through local processing. Field testing under simulated nighttime flooding achieved 99.2% detection accuracy, far exceeding single-modality visible-light systems. This scalable, resilient solution complements satellite flood products, enabling timely decision-making for vulnerable rural households.