A comprehensive, end-to-end geospatial pipeline for wildfire damage assessment.
This repository bridges the gap between raw orbital data and deep-learning-based classification. It provides a complete workflow from Google Earth Engine (GEE) data harvesting to Prithvi EO 2.0 (Vision Transformer) fine-tuning and rigorous Metric Comparison.
- Earth Observation: Google Earth Engine (GEE API), Sentinel-2 Multispectral Data.
- Deep Learning: PyTorch Lightning, Vision Transformers (ViT), UperNet Architecture.
- Geospatial Engineering: Delta-Channel Algorithm (Change Detection), Multi-Temporal Image Tiling.
- Foundation Models: NASA-IBM Prithvi EO 2.0 adaptation.
- Evaluation: Side-by-Side Model Benchmarking (+50.58% F1 improvement).
The core of this repository, including data generation, model specialization, and comparative analysis, was developed by Tushar Thokdar.
- End-to-End Pipeline: Developed the complete workflow from raw data export to model evaluation.
- Delta Channel Algorithm: Implemented a unique temporal stacking strategy that explicitly calculates the spectral difference between pre and post-fire states.
- Advanced Fine-Tuning: Configured a "Freeze-then-Unfreeze" training strategy, allowing the model to adapt its ViT backbone to burn scars without losing global spatial knowledge.
- Evaluation Framework: Built a comprehensive benchmarking suite that demonstrated a +50.58% improvement in Macro F1 score compared to the base foundation model.
The system follows a rigorous four-stage pipeline from raw orbital data to validated performance metrics.
graph TD
subgraph GEE ["1. Data Acquisition (GEE)"]
A[Sentinel-2 L2A] --> B[Spectral Band Selection]
B --> C[DNBR Calculation & Export]
end
subgraph Engineering ["2. Data Engineering"]
C --> D[Reflectance Normalization]
D --> E[Delta Channel Algorithm]
E --> F[Multi-Temporal 13-Band Tiling]
end
subgraph Model ["3. Model Specialization"]
F --> G[Prithvi EO 2.0 Backbone]
G --> H[UperNet Decoder]
H --> I[Two-Stage Training]
end
subgraph Eval ["4. Evaluation & Comparison"]
I --> J[Fine-Tuned Model]
K[Baseline Foundation Model] --> L[Comparison Framework]
J --> L
L --> M[Performance Benchmarks]
end
- GEE Export: Pulls 6 spectral bands (B2, B3, B4, B8A, B11, B12).
- Dataset Generator: Creates 224x224 chips with a 3D temporal stack (Pre, Post, Delta).
- Transformer Backbone: Leverages pre-trained Vision Transformer weights for spatial-spectral understanding.
- UperNet Decoder: Processes multi-scale features for precise pixel-level segmentation.
- Comparison Suite: Validates the model against ground truth and provides a quantitative "Verdict" on improvement.
- Backbone: Prithvi EO 2.0 (Vision Transformer based Masked Autoencoder).
- Decoder: UperNet (Unified Perceptual Parsing Network) for high-resolution segmentation.
- Input: 3-frame temporal stack (Pre-fire, Post-fire, and the explicit Delta channel).
- The Delta Strategy:
Delta = Clip(Post-fire - Pre-fire, -1.0, 1.0). This provides the model with direct information about the magnitude of spectral change. - Hybrid Loss: Uses a combination of Weighted Cross-Entropy (to handle class imbalance) and Dice Loss (weight: 0.8) to optimize for precise burn scar geometries.
- Training Pipeline:
- Stage 1: Backbone frozen for 5 epochs to train the UperNet decoder.
- Stage 2: Full unfreezing for joint optimization.
Fine-tuning a geospatial Foundation Model (Prithvi EO 2.0)/prithivi_finetune_with_delta.ipynb: Main fine-tuning notebook utilizing 3D temporal stacking.
Google Earth Engine to TIF Export/Google_Earth_Engine_to_TIF_Export.ipynb: GEE scripts for harvesting Sentinel-2 spectral data.
Prithvi data generation/Prithvi_data_generation.ipynb: Author: Tushar Thokdar. Handles 13-band chip generation and quality control.
Prithvi model comparison/Prithvi_model_comparison.ipynb: Author: Tushar Thokdar. Detailed metric analysis including F1-scores and confusion matrices.
The specialized fine-tuning process yields a massive performance boost over the base foundation model, particularly in identifying high-severity fire damage.
| Metric | Baseline Model | Fine-Tuned Model | Improvement |
|---|---|---|---|
| Accuracy | 35.86% | 69.93% | +34.07% |
| Macro F1 | 0.1160 | 0.6218 | +50.58% |
| Weighted F1 | 0.2035 | 0.7015 | +49.80% |
| Burned F1 | 0.0133 | 0.5553 | +54.20% |
The chart below illustrates the F1-score improvement across all severity classes. The fine-tuned model (green) consistently outperforms the foundation model (blue), which often fails to recognize specialized burn classes.
Side-by-side analysis shows that while the foundation model is biased towards the "Unburned" class, the fine-tuned model accurately distributes predictions across the severity gradient.
A direct comparison between the actual satellite imagery, ground truth labels, and model predictions. Notice how the fine-tuned model captures the intricate geometry of the burn scar that the foundation model misses.
Beyond simple classification, the model demonstrates high spatial coherence and spectral sensitivity to scorched earth patterns.
The following capture shows the model's performance on unseen validation data, demonstrating its ability to generalize across different terrain types.
Detailed breakdown of pixel-level classification accuracy for the specialized model.
- Model: Prithvi EO 2.0 (IBM/NASA Geospatial ViT)
- Libraries: PyTorch, PyTorch Lightning, Terratorch, Rasterio, SMP.
- Data: Sentinel-2 (6 Bands: Red, Green, Blue, NIR, SWIR1, SWIR2).
- Environment Setup:
pip install -r requirements.txt
- Data Export: Use the GEE notebook to pull pre/post-fire imagery.
- Dataset Preparation: Run
Prithvi_data_generation.ipynb. - Fine-Tuning: Execute
prithivi_finetune_with_delta.ipynb. - Comparison: Use
Prithvi_model_comparison.ipynbto verify results.