Deep Learning for Flood Detection

I have implemented a PyTorch-based U-Net model for water area segmentation, targeting both flooded and permanent water areas in Sentinel-1 satellite images. Trained on the Cloud to Street - Microsoft flood dataset, the model processes Sentinel-1 image chips labeled with water data to deliver accurate segmentation results.

The implementation includes scripts for preparing the data, such as tiling Sentinel-1 images and downloading additional datasets like SRTM DEM and JRC Permanent Water data from Google Earth Engine. Additionally, Jupyter notebooks are provided for training and inference. The dataset was split based on water percentage, with two U-Net models trained on their respective subsets. The combined predictions achieved a high Intersection over Union (IoU) score of 0.877 on the test set.

This work offers a complete pipeline for water area segmentation, from data processing to model evaluation, and serves as a valuable tool for similar satellite image segmentation tasks.