Setup

Environment, repository, and the sample data

1. Clone the repository

Everything in this tutorial lives in the project repository, including the trained model, DI Klambu tertiary-block boundaries, and the reference index outputs you will need.

git clone https://github.com/firmanhadi21/s1-rice-irrigation-tutorial.git
cd s1-rice-irrigation-tutorial

2. Create the environment

The models run on CPU — no GPU needed. Use conda (recommended) or a plain virtualenv.

conda create -n s1rice python=3.10 -y
conda activate s1rice

pip install tensorflow rasterio geopandas rioxarray scikit-learn \
            imbalanced-learn numpy pandas matplotlib scipy joblib \
            earthengine-api requests

Tip

rasterio and geopandas pull in GDAL. If pip struggles, install them from conda-forge instead: conda install -c conda-forge rasterio geopandas rioxarray -y.

Force CPU-only execution (avoids GPU/XLA issues) — the training/prediction scripts already set this, but for interactive use:

export CUDA_VISIBLE_DEVICES=-1

3. What ships with the repo (your sample data)

You already have these after cloning — they make the workflow runnable immediately:

File	Used in
`model_files_paddy_vh/paddy_vh_model.keras` (+ `scaler.joblib`)	Paddy prediction
`model_files_paddy_vh/training_data_*.csv`	Paddy training (29 VH features + labels)
`data/training_points_paddy_binary.csv`	Training-point reference
`2026/irrigation_performance/klambu.gpkg`	DI Klambu tertiary blocks
`2026/irrigation_performance/*_block_results.csv`	Reference SI/CU/RI per block

The only thing missing is the Sentinel-1 VH stack for the AOI (too large to ship) — fetch it next.

4. Fetch the sample Sentinel-1 stack

Sentinel-1 GRD is free and open. We fetch a small VH 12-day composite stack over DI Klambu straight from Google Earth Engine.

Authenticate once (free account):

earthengine authenticate

On an HPC / headless server

Skip interactive login: drop a service-account key named ee-geodetic.json in the repo root (the script auto-detects it), or pass --service-account-key /path/to/key.json. No browser needed.

Then run the fetch script (shipped in tutorial/scripts/):

cd tutorial/scripts
python fetch_sample_aoi.py --year 2024 --start-period 7 --end-period 24 \
    --out ../../data/sample/klambu_vh_2024.tif

This writes data/sample/klambu_vh_2024.tif — a multi-band GeoTIFF with one band per 12-day period, named 2024_Period_7 … 2024_Period_24, at 50 m. Verify it:

cd ../..
gdalinfo data/sample/klambu_vh_2024.tif | head -20

No Earth Engine account?

The whole workflow accepts any VH GeoTIFF with 12-day period bands. If you preprocess Sentinel-1 yourself (e.g. with ESA SNAP), just point the scripts at your own stack. The band naming convention is <year>_Period_<n>.

5. Make an AOI mask (optional but recommended)

Most scripts accept a --mask so processing is limited to land/AOI. For the sample you can derive one from the DI Klambu boundary:

# rasterize the DI boundary onto the VH grid as a 0/1 mask
gdal_rasterize -burn 1 -tr 0.000449 0.000449 -a_nodata 0 \
    -te 110.72 -7.10 110.94 -6.94 -ot Byte \
    2026/irrigation_performance/klambu.gpkg data/sample/klambu_mask.tif

You are ready. Continue to the Paddy-field map.