datatonFach — Wildfire Risk from Satellite Imagery

Pipeline that turns multi-band satellite rasters (elevation + NDVI statistics) into dense DinoV2 patch embeddings, classifies fire-risk severity with a FAISS KNN, and reconstructs a pixel-level risk map for the full raster. Built for the FACh (Fuerza Aérea de Chile) datathon on wildfire analytics over Valparaíso and US regions.

Stack: Python 3.10 · PyTorch · DinoV2 (ViT) · FAISS · scikit-learn · rasterio · Streamlit · uv · ruff · pytest · GitHub Actions

What it does

1. Compose a false-RGB GeoTIFF from three physical signals — elevation, NDVI mean, NDVI std — so a vision model can exploit them as image channels.
2. Featurize each 224×224 tile with a frozen DinoV2 ViT-B/14, keeping dense patch embeddings (no fine-tuning).
3. Train a FAISS IVF KNN over the embeddings with stratified k-fold CV against bbox-annotated fire-risk classes (high, very_high, moderate, low, very_low, non-burnable, water).
4. Predict over arbitrarily large rasters via sliding-window tiling and stitch the output back into a georeferenced class map.
5. Quantify burn severity post-fire from dNBR rasters (Landsat pre/post).
6. Serve an interactive Streamlit demo that renders predictions and burn quantification on a Leaflet map.

Architecture

flowchart LR
    subgraph INPUT["Raw rasters (.tif)"]
        E[Elevation]
        N1[NDVI mean]
        N2[NDVI std]
    end

    subgraph FEAT["features/"]
        C["compose<br/>false-RGB GeoTIFF"]
        SW1["sliding window<br/>224x224 tiles"]
        D["DinoV2 ViT-B/14<br/>patch embeddings"]
    end

    subgraph MODEL["models/"]
        T["Stratified K-Fold<br/>FAISS IVF KNN"]
        M[("models/<br/>*.faiss + *.pkl")]
        P["sliding-window<br/>prediction"]
    end

    subgraph OUT["Outputs"]
        R["risk map<br/>GeoTIFF"]
        V["classification<br/>report"]
        Q["dNBR severity<br/>quantification"]
        S["Streamlit demo<br/>leafmap + folium"]
    end

    E --> C
    N1 --> C
    N2 --> C
    C --> SW1 --> D
    D -->|features.npy + labels_cls.npy| T
    T --> M
    M --> P
    D --> P
    P --> R
    R --> V
    R --> S
    Q --> S

Loading

Project layout

src/datatonfach/
  config.py              # ROOT / DATA_DIR / MODELS_DIR paths
  io/geo.py              # rasterio load/save + axis helpers
  features/
    input_images.py      # false-RGB composition
    featurizer.py        # DeepFeaturizer (DinoV2, CPU/GPU-safe)
    dataset.py           # (features.npy, labels_cls.npy) builder
  models/
    knn.py               # FaissKNeighbors, FaissKMeans
    sliding_window.py    # patchify / unpatchify
    train.py             # stratified k-fold training + metrics
    predict.py           # full-raster prediction pipeline
  eval/
    validation.py        # bbox-based classification report
    pixels_count.py      # dNBR severity quantification
  cli.py                 # typer entry point
interface/               # Streamlit demo
tests/                   # pytest smoke tests
.github/workflows/ci.yml # ruff + pytest on every push

Install

git clone git@github.com:sebastianbreguel/datatonFach.git
cd datatonFach
uv sync --extra dev
uv run pre-commit install

Data + pretrained weights: Google Drive (place under ./data/ and ./models/). For GPU DinoV2 inference install the matching PyTorch CUDA build from pytorch.org.

Usage

uv run datatonfach compose                       # build composed.tif per folder
uv run datatonfach featurize --backbone base     # DinoV2 embeddings + labels
uv run datatonfach train --k 11 --n-splits 5     # stratified k-fold KNN
uv run datatonfach predict                       # risk map over valpo/composed.tif
uv run datatonfach validate                      # classification report vs bbox GT
uv run streamlit run interface/app.py            # interactive demo

Trained artifacts → ./models/, predictions → ./data/predictions/.

Dev workflow

uv run ruff check .          # lint
uv run ruff format .         # format (line length 140, double quotes)
uv run pytest                # tests
uv run pre-commit run --all-files

CI runs ruff + pytest on every push/PR (.github/workflows/ci.yml).

Team

Luis Aros Illanes · Andrés Sebastián de la Fuente · Lucas Carrasco Estay · Benjamín Henríquez Soto · Sebastián Breguel González · Martín Bravo Díaz

License

MIT — see LICENSE.