Unsupervised mapping of rice paddy fields and their inundation patterns using Sentinel-1 SAR images and GIS

The periodic flooding of rice paddies presents significant environmental challenges, including methane emissions, fertilizer pollution, and water resource stress. This study introduces a scalable approach using high-resolution (10 m) Sentinel-1 SAR imagery and Global Surface Water Extent rasters to map rice paddies and their flooding patterns. Temporal variations in SAR backscatter are condensed into a custom multi-band image for each planting season, enabling unsupervised classification to delineate rice paddies and flooding patches. This approach achieves high detection accuracy, with overall accuracy rates for paddy detection of 83.8%-89.0% in Japan, 71.5%-83.5% in the Philippines, and 72.3% for rice flooding patches in Bali, Indonesia. Unlike most existing methods, our approach does not rely on ancillary data or context-specific information for training or labeling, making it scalable and adaptable across diverse geographies. We explore this potential by examining the method's underlying assumptions and identifying areas where these assumptions may be challenged.