Wildfire Burn Scar Detection Pipeline

Overview

Built a comprehensive end-to-end geospatial ML pipeline achieving +50.58% Macro F1 improvement over foundation model baseline through novel algorithmic contributions.

Algorithmic Innovations

Delta Channel Algorithm (Original Research)

Invented a temporal stacking strategy calculating explicit spectral difference between pre/post-fire states: Delta = Clip(Post - Pre, -1.0, 1.0)

Multi-Temporal 13-Band Architecture

Designed 3D temporal stack (Pre, Post, Delta) with 6 spectral bands per frame, enabling the model to learn both absolute spectral signatures and temporal changes.

Two-Stage Fine-Tuning

Implemented freeze-then-unfreeze approach: 5 epochs frozen backbone then full joint optimization.

End-to-End Pipeline

1. Data Acquisition (GEE) — Automated Sentinel-2 L2A harvesting with spectral band selection
2. Data Engineering — Reflectance normalization, Delta channel computation, 224x224 chip tiling
3. Model Fine-Tuning — Prithvi EO-2.0 with UPerNet decoder for burn scar classification
4. Evaluation — Comprehensive benchmarking suite with confusion matrices

Performance