DoctorLaplace/Gaia

Predicting biodiversity in the South African Cape Region # Gaia: A Masked Vision Transformer for Species Richness Prediction from Hyperspectral Imagery Gaia is a masked spatial-spectral vision transformer fine-tuned on EnMAP foundation weights to predict animal species richness (birds, frogs, and insects) from AVIRIS-NG hyperspectral imagery across the Greater Cape Floristic Region (South Africa) using NASA BioSCape data. ## QuickStart ### 1. Prerequisites - **Python 3.10+** (Recommend a virtual env) - **CUDA GPU** with at least 8GB VRAM. - **NASA Earthdata Account**: Register here. ### 2. Environment Setup ```bash pip install -r requirements.txt ``` ### 3. Data Acquisition (Smart Sync) Downloads AVIRIS-NG granules, applies Dr. Clark's Bad Band List (BBL), and optionally downsamples. Data is automatically organized by resolution: `data/bioscape/30m/` and `data/bioscape/5m/`. #### Option A: Targeted Sync (Recommended, ~535 labeled granules) ```bash # 1. Generate the labeled inventory python src/generate_labeled_inventory.py # 2. Download 30m (default, ~35MB each, fast local training) python src/smart_sync.py --local # 3. Download 5m (original resolution, ~1.3GB each, for supercomputer) python src/smart_sync.py --local --res 5 ``` #### Option B: Full Dataset Sync (3,648 granules, 500GB+) ```bash python src/smart_sync.py --inventory full_inventory.txt --local --res 30 ``` #### Option C: S3 Upload (for NRP Nautilus) ```bash python src/smart_sync.py --res 30 ``` ### 4. Model Training Gaia uses a Masked Spatial-Spectral Transformer (SST) with EnMAP foundation weights. ```bash # Train on 30m data (default, from config) python src/train_production.py --epochs 250 # Train on 5m data (override directory) python src/train_production.py --nc_dir data/bioscape/5m --epochs 250 # Quick test run (2 granules, 1 epoch) python src/train_production.py --test-run ``` ### 5. Evaluation Evaluates the best checkpoint on a held-out validation set (10%, same split as training). ```bash python src/evaluate.py --nc_dir data/bioscape/ …