Deep Learning U-Net for Radar Imagery Flood Mapping
Problem Context
During heavy rainfall and floods, cloud cover often blocks optical satellites. Sentinel-1 Synthetic Aperture Radar (SAR) can pass through clouds and measure surface roughness. Because calm floodwater acts like a flat mirror, it reflects the radar signal away and appears darker than surrounding vegetation or buildings.
Model & Chip Pipeline
The pipeline splits radar scenes into 128 × 128 pixel chips paired with binary flood masks from the Kaggle dataset (dimskeee/unet-segmentation-using-radar-imagery).
- Encoder-Decoder Network: A compact U-Net using
16,32,48, and64filter blocks with skip connections to keep water boundaries sharp. - Memory-Friendly Training: Patch size and filter counts are kept light enough to train smoothly on a 6 GB VRAM GPU (RTX 2060 Super) without losing detail along narrow channels.
- Visual Output Check: Stitched
512×512cosine-blended predictions (IoU 0.74–0.83) can be compared pixel-by-pixel against SAR backscatter and ground truth in Plate 01 (/atlas/radar-flood-unet).
Engineering Notes
Keeping preprocessing and inference lightweight makes it easy to export predicted masks into standard GIS formats, such as GeoTIFF rasters or vector polygons, for flood response workflows.