Monitoring rice paddy fields is critical for food security, water management, and methane emission tracking. However, traditional optical satellite imagery (like Sentinel-2) is often unusable in tropical and subtropical regions due to persistent cloud cover during the monsoon cropping season.

This project implements an end-to-end Deep Learning pipeline to automatically segment paddy fields using Sentinel-1 SAR (Synthetic Aperture Radar) Ground Range Detected (GRD) data. By leveraging the unique backscatter temporal signature of rice during its flooding and growth stages, the model achieves high-precision mapping regardless of weather conditions.

• Multi-Temporal Fusion: Processes a 60-band data stack representing a 20-date cropping season. Each date includes VV, VH, and VV/VH ratio polarizations to capture the "double-bounce" scattering effect characteristic of rice stems in water.
• Geospatial Preprocessing: Integration of ESA SNAP (pyroSAR) for specialized SAR calibration, thermal noise removal, and terrain correction (orthorectification).
• MLOps Architecture:
  • DVC (Data Version Control): Manages heavy 60-band GeoTIFF stacks and model weights without bloating the Git repository.
  • MLflow: Tracks hyperparameter experiments, loss curves, and evaluation metrics (IoU, F1-Score).
• High Performance: Optimized a U-Net + ResNet34 architecture to achieve a 0.85 IoU, successfully filtering "speckle noise" inherent in SAR data.

• Sentinel-1 SAR: Multi-temporal C-band data (20 dates).
• JAXA LULC Map: Used for automated ground-truth label generation (Rice vs. Non-Rice).
• ROI: Niigata, Japan (High-intensity rice production region).

• Python 3.10 or 3.11
• NVIDIA GPU with CUDA support (Recommended)

Construct a virtual environment and install the required dependencies:

python -m venv venv
source venv/bin/activate  # On Windows: venv\Scripts\activate
pip install -r requirements.txt

On Windows systems, you might encounter OSError: [WinError 1114] when importing torch. This is resolved by manually loading c10.dll. The following snippet is included in the project scripts:

import os
import platform
import ctypes
from importlib.util import find_spec

if platform.system() == "Windows":
    try:
        if (spec := find_spec("torch")) and spec.origin:
            dll_path = os.path.join(os.path.dirname(spec.origin), "lib", "c10.dll")
            if os.path.exists(dll_path):
                ctypes.CDLL(os.path.normpath(dll_path))
    except Exception:
        pass

We use specific stable versions to ensure compatibility on Windows:

• torch==2.5.1+cu121
• torchvision==0.20.1+cu121
• torchgeo==0.6.2

• src/: Source code for the pipeline stages.
  • training.py: Handles tiling and model training.
  • testing.py: Runs inference on test images and generates visualizations.
  • labeling.py: Prepares binary labels from external data.
• data/: Data directory (standardized structure).
  • raw/: Unprocessed input data.
  • processed/: Features, stacks, and labels.
  • external/: External shapefiles and validation data.
• config.yaml: Centralized configuration for paths and training parameters.
• dvc.yaml: DVC pipeline definition.

The project uses DVC to manage the workflow:

1. Prepare Labels: python src/labeling.py
   • Merges source TIFs and creates binary masks/geojson.
2. Train Model: python src/training.py
   • Generates tiles (patches) from input rasters and labels.
   • Trains the segmentation model.
3. Inference (Experimental): python src/testing.py
   • Performs semantic segmentation on test images.
   • Orthogonalizes results and generates split-map visualizations.

Modify config.yaml to adjust training parameters:

• tile_size: Size of the input patches (default: 512).
• epochs: Number of training iterations.
• batch_size: Number of samples per training step.

By default, large data files and models are stored in:

• Project root (data/, models/)
• External drive (if configured in training.py, e.g., G:\data)

To run the full pipeline through DVC:

dvc repro

To run individual steps:

python src/training.py
python src/testing.py