Deep Learning U-Net for Radar Imagery Flood Mapping

Semantic segmentation of Sentinel-1 polarimetric SAR imagery using a TensorFlow U-Net trained on 128×128 radar chips for rapid post-flood mapping.

ACTIVE · 2021-09-07
SPECIFICATIONS // GEOAI · FLOOD MAPPING 4 tags
Architecture U-Net (16 / 32 / 48 / 64 filters)
Sensor Input Sentinel-1 Polarimetric SAR (128×128 chips)
Training Stack TensorFlow / Keras · RTX 2060 Super (6 GB)
Research Origin PKM-RE · PIMNAS 34 (UGM, 2021)

Problem Context

During heavy rainfall and floods, cloud cover often blocks optical satellites. Sentinel-1 Synthetic Aperture Radar (SAR) penetrates clouds and responds to surface roughness and scattering behavior. Because calm floodwater acts like a specular reflector, it scatters radar energy away from the sensor and appears sharply distinct from surrounding vegetation and built-up land in polarimetric composites.

Research Origin

This pipeline originated from a 2021 PKM-RE1 research project at the Faculty of Geography, Universitas Gadjah Mada (UGM) (Analisis Pemetaan Cepat Pasca Bencana Banjir Menggunakan Dekomposisi Polarimetrik Radar Sentinel-1 Berbasis Model Deep Learning), representing UGM at PIMNAS 342.

  • Team: Iklila Rahmatika (Lead), Elyana Inggrid Widiastuti, Dimas Maulana Ichsan, and Made Dwiyana Surya Upadana
  • Advisor: Dr. Sc. Sanjiwana Arjasakusuma, S.Si., M.GIS.

Model & Chip Pipeline

Sentinel-1 scenes (including South Kalimantan flood scenes around 2.38°–2.85° S, 115.07°–115.15° E) are prepared as 128 × 128 pixel polarimetric false-color chips paired with binary water masks and published on Kaggle (dimskeee/unet-segmentation-using-radar-imagery).

  • Encoder-Decoder Network: A compact U-Net using 16, 32, 48, and 64 filter blocks with skip connections to preserve narrow river channels and flooded boundaries.
  • Memory-Friendly Training: Patch size and filter counts stay light enough to train smoothly on a 6 GB VRAM GPU (RTX 2060 Super) without losing boundary precision.
  • Visual Output Check: Stitched 512×512 cosine-blended predictions (IoU 0.74–0.83) can be inspected 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 straightforward to export predicted flood masks into standard GIS deliverables, such as GeoTIFF rasters or vector polygons, for rapid post-disaster response workflows.

Footnotes

  1. PKM-RE (Program Kreativitas Mahasiswa - Riset Eksakta): Indonesia’s national competitive undergraduate exact-sciences research grant program funded by the Ministry of Education, Culture, Research, and Technology (Kemendikbudristek). ↩

  2. PIMNAS 34 (Pekan Ilmiah Mahasiswa Nasional ke-34): The 34th National Student Scientific Week (2021), Indonesia’s annual national finals where top PKM-funded university research teams present and compete. ↩