~/atlas/radar-flood-unet

Interactive 512×512 mosaiced Sentinel-1 SAR viewer comparing backscatter amplitude against U-Net flood masks, ground truth, and shoreline error overlays.

PLATE 01 · 512×512 SAR · U-Net
PLATE METADATA // FLOOD MAPPING PUBLISHED 2024-05-01
Sensor Sentinel-1 SAR (C-Band)
Resolution 512 × 512 mosaic (4×4 × 128px)
Dataset Kaggle Radar Flood Chips
Viewer Mode chip-inspector
VIEWER // CHIP INSPECTOR · 01 Sentinel-1 SAR (C-Band) · 512 × 512 mosaic (4×4 × 128px)
CHIP:
MODE:
LAYER:
512 × 512 mosaic (4×4 × 128px)
Active Sample Chip 01 · Main River Channel
Center Coordinates 2.38° S, 115.07° E
U-Net Flood Extent 34.1%
Ground Truth Extent 36.8%
IoU Score 0.74
False-Color Legend
SAR False-Color RGB (Base)
U-Net Pred
Water Boundary Edge
Real 512×512 Sentinel-1 SAR false-color RGB mosaics and cosine-blended U-Net inference (threshold 0.50) exported from unet_flood.h5.

Plate 01 · Sentinel-1 SAR Flood Prediction

Plate Interpretation

Each Plate in ~/atlas pairs a geospatial model output with a viewer built for that data type. This plate inspects contiguous 4×4 (512×512 px) Sentinel-1 scenes reconstructed using overlapping stride=32 cosine-blended inference to eliminate 128×128 tile seams:

  • SAR False-Color RGB Base: In the 3-band Sentinel-1 input composite, purple/magenta tones indicate smooth flooded or open-water surfaces while green tones indicate rough or dry land cover.
  • Overlay Mode: Blends the selected binary mask over the Sentinel-1 false-color base, marking water perimeter boundaries in NIR crimson.
  • Split Mode: Slides between raw SAR False-Color RGB on the left and the predicted mask, ground truth, or agreement diff on the right.
  • Mask Mode: Isolates the binary classification or pixel-wise agreement map on a clean canvas for shape and boundary inspection.
  • Layer Switch: Compare U-Net Pred directly against Ground Truth or inspect Pred vs GT Diff (True Positive, False Positive, and False Negative pixels).