Wildfire Detection (Sentinel-2)

Small, focused pipeline to build a 3-class dataset (fire / no_fire / burn_scar) from FIRMS + Sentinel-2, and train ResNet classifiers.

Quick Start

• Install: pip install -r requirements.txt
• Configure .env with Sentinel Hub creds (SH_CLIENT_ID, SH_CLIENT_SECRET)
• Pipeline:
  1. python scripts/build_candidates.py
  2. python scripts/download_dataset.py
  3. python scripts/prepare_splits.py
• Train:
  • RGB: python models/train_resnet.py
  • 6-band: set EXPERIMENT="all" in train_resnet.py

What you get

• Data in data/splits/ (train/val/test), bands: B02,B03,B04,B08,B11,B12
• Quality filters: cloud/water/valid pixels + thermal (B12) thresholds
• Geographic split to avoid spatial leakage

Key Files

• scripts/build_candidates.py — match FIRMS to S2 catalog
• scripts/download_dataset.py — download patches + metadata
• scripts/prepare_splits.py — filter, balance, and split
• models/dataset.py — PyTorch dataset + transforms
• models/train_resnet.py — RGB/6-band training (ReduceLROnPlateau, dropout, label smoothing)

Notes

• MPS/Metal and CUDA are auto-detected.
• Checkpoints saved to models/checkpoints/.
• For detailed steps, see DATA_PIPELINE.md.