Deforestation Embedder

Inspiration

RADD alerts provide perfect ground truth for deforestation events, but current monitoring relies on optical indices that struggle with clouds and miss early signals. Foundation models could learn deforestation patterns directly from SAR embeddings, revealing forest loss signatures invisible to traditional methods.

What it does

Trains a classifier on Copernicus Foundation Model embeddings to detect deforestation using SAR patterns. Processes RADD alerts to identify confirmed forest loss, extracts Sentinel-1 embeddings before/after events using CopernicusFM, and learns deforestation signatures in high-dimensional embedding space for cloud-independent monitoring.

How we built it

Data: RADD alert tiles from South America, Africa, Southeast Asia
Foundation Model: CopernicusFM (ViT-based) for SAR embedding extraction
Pipeline:
1. Download RADD tiles and extract deforestation events
2. Generate negative samples from stable forest areas
3. Load Sentinel-1 data temporally aligned with alerts
4. Extract embeddings via batch processing
5. Train MLP classifier with stratified splits
Tools: Python + PyTorch, TorchGeo, scikit-learn, xarray, cubo

Challenges we ran into

Temporal alignment between RADD alerts and Sentinel-1 observations
Negative sampling strategy ensuring spatial proximity but temporal integrity
Processing large-scale SAR data across multiple continents consistently
Handling variable observation frequencies in tropical regions
Batch processing optimization for embedding generation

Accomplishments that we're proud of

Built end-to-end pipeline from raw RADD alerts to trained classifiers
Robust negative sampling within same tiles avoiding location bias
Cross-continental processing across diverse forest ecosystems (3 continents)
Unified batch processing for efficient embedding generation
Stratified train/val/test splits with early stopping for production model

What we learned

Foundation model embeddings reveal deforestation patterns invisible to vegetation indices
SAR provides consistent cloud-independent monitoring for tropical regions
Temporal difference in embeddings shows deforestation as a process
Spatial matching crucial to prevent location-based rather than pattern-based learning