Description of Invention
This project developed deep learning models for flood detection using satellite imagery to enhance disaster monitoring capabilities. Two advanced neural networks (HRNet+OCR and Attention U-Net) were implemented to classify imagery into background, permanent water, and flood water categories. Models were trained on Sen1Floods11 dataset containing imagery from diverse flood events across multiple continents. Attention U-Net achieved superior performance with 81.3% IoU accuracy and processed images 1.8 times faster than alternatives. The system successfully integrates multi-source satellite data and distinguishes between permanent and flood water, providing critical information for emergency response while reducing false alarms.