Advancing Paddy Rice Mapping in Cloud-Prone Regions Using SAR and Machine Learning: A Case Study in Gangetic West Bengal, India

Abstract

Accurate seasonal mapping of paddy rice is important for agricultural monitoring in cloud-prone tropical regions, where optical satellite observations are frequently obstructed during critical growth stages. This study developed a cloud-resilient workflow that integrates multitemporal Sentinel-1 Synthetic Aperture Radar imagery, a Random Forest classifier, and Google Earth Engine to map Boro, Aus, and Aman rice across Gangetic West Bengal, India, at 10 m spatial resolution. Season-specific classifiers were trained using reference areas derived from Sentinel-2 imagery, very-high-resolution Google Earth imagery, field photographs, and more than 5,000 field samples collected in 2023. Vertical transmit-horizontal receive backscatter time series were used because they captured temporal changes associated with flooding, transplanting, vegetation development, and post-harvest conditions. Independent validation showed overall accuracies of 94.6% for Boro, 95.3% for Aus, and 96.2% for Aman. The resulting maps indicated that Aman rice had the greatest spatial extent, followed by Boro and Aus, while central and southern parts of the region contained the highest concentrations of multi-season rice cultivation. Overlaying the seasonal outputs identified areas of single, double, and triple cropping, although direct field validation of cropping intensity was unavailable. The workflow supports the practical value of combining SAR time-series data, machine learning, and cloud computing for consistent seasonal paddy rice mapping in fragmented and persistently cloudy agricultural landscapes.