HydroChangeNet

Compact multi-temporal flood change segmentation with terrain-aware gates and a pixel-level log-variance head.

Important

Artifact	Meaning
This Hub demo checkpoint	Trained on synthetic sample tiles so anyone can predict / smoke-test without Kaggle data. Not a Kuro Siwo benchmark result.
Local 240-tile Kuro Siwo pilot (in the GitHub README)	Flood IoU 0.6287, F1 0.7433 vs post-event U-Net 0.4751 / 0.5910. Tile-level split, not event-disjoint, not comparable to published BlackBench numbers.

Quick inference

pip install torch huggingface_hub
# from the HydroChangeNet repo:
PYTHONPATH=src python -m floodchange_uq.cli predict \
  --hub pancakesnstrawberries/HydroChangeNet \
  --tile sample_data/sample_000.pt \
  --output artifacts/prediction

Architecture (short)

Shared temporal encoder (pre/post) → independent skip & bottleneck change gates → terrain encoder (HAND, slope, DEM, flow direction, flow accumulation) → decoder → 3-class logits + log-variance map.

Real-world comparison targets (not claimed by this demo)

For a publishable comparison, evaluate on the official Kuro Siwo / BlackBench event-disjoint protocol against published methods such as:

• U-Net + ResNet backbones (BlackBench GRD best setting reports flood F1 ≈ 80.12% / mIoU ≈ 76.20% for UNet-ResNet50 — see Bountos et al., NeurIPS 2024)
• DeepLabv3, UPerNet (Swin / ConvNeXt)
• Change-detection models: FC-EF, SNUNet-CD, ChangeFormer
• Temporal: ConvLSTM; SSL: FloodViT

Do not compare this Hub demo or the non-disjoint 240-tile pilot to those numbers.

Code: VarunikaN/HydroChangeNet

Citation

If you use Kuro Siwo / BlackBench:

@inproceedings{bountos2024kurosiwo,
  title={Kuro Siwo: 33 billion m\^{}2 under the water. A global multi-temporal satellite dataset for rapid flood mapping},
  author={Bountos, Nikolaos Ioannis and others},
  booktitle={NeurIPS},
  year={2024}
}