Galileo Pretrained Remote Sensing Model – Rwanda Crop Type Classification

A fine-tuning pipeline for the Galileo Pretrained Remote Sensing Model to classify crop types using pixel-wise segmentation with multi-temporal Sentinel-1 and Sentinel-2 data in Rwanda.

---

Quickstart

1. Clone the repository

git clone <repo-url>
cd galileo

2. Create Environment

With conda:

conda env create -f environment.yml
conda activate galileo-env

Dataset Preparation

Step 1: Export Sentinel-1/2 Data from Earth Engine

Use the script below to export Sentinel-1/ Sentinel-2 images over your area of interest for specific session (Season A or Season B):

python run_export_rwanda.py

Step 2: Download Exported Tiles from GEE

After GEE completes your export tasks, download the resulting GeoTIFF tiles::

python download_tiles_from_gee.py

Step 3: Extract Pixel-wise Patches from Tiles

Use preprocessed Sentinel-1/Sentinel-2 tiles and polygon labels to generate 8×8 pixel-wise training patches:

python extract_patches_from_tiles.py

This script will:

• Read raster tiles and label masks
• Extract valid 8×8 image and mask patches
• Save .npz patch datasets for train, val, and test splits

Folder Structure Example

data/
  patches/
    train/
      inputs/
      masks/
    val/
    test/

---

Fine-Tuning

Train the pixel-wise classifier using the Galileo encoder:

python finetune_rwanda.py \
  --data_dir data/patches/ \
  --encoder_ckpt data/models/nano/ \
  --save_dir checkpoints/ \
  --batch_size 8 \
  --epochs 50 \
  --lr 0.0001

Inference

Run inference on prepared raster tiles:

python inference.py

This script will:

• Load each tile (60-band: 5 timesteps × 12 bands)
• Apply a sliding window to generate 8×8 predictions
• Merge predictions and output a GeoTIFF (merged_prediction.tif)

Important:

• Your input tiles must have 60 bands (5 months × 12 bands).
• Invalid or low-confidence pixels are masked using a configurable confidence_threshold.

---