Abstract
Rice (
Oryza sativa L.
) is Bangladesh’s staple crop, cultivated across three distinct seasons: Rabi (Boro), Kharif-1 (Aus) and Kharif-2 (Aman). This study develops an integrated, cloud-independent framework for mapping Aus and Aman rice areas using freely available Sentinel-1 C-band Synthetic Aperture Radar (SAR) imagery; to accomplish this, supervised machine-learning (ML) algorithms are used in the Google Earth Engine (GEE) platform. Time-series dual-polarization (VV and VH) backscatter data were analyzed for June–November 2021 (Aman) and March–August 2022 (Aus). Four ML models such as Classification and Regression Tree (CART), Random Forest (RF), k-Nearest Neighbour (k-NN), and Support Vector Machine (SVM) were evaluated to identify the most accurate mapping approach. Among all ML models, k-NN achieved the highest accuracy of 94.5% and 99.8% for Aman and Aus, respectively. Despite the higher accuracy, RF for Aman (92.7% accuracy) and CART for Aus (99.6%) were finally selected based on their lower deviation from area estimates reported by Bangladesh Bureau of Statistics (BBS). The estimated Aman area (5.72 Mha) was closely matched with the official figure from the Bangladesh Bureau of Statistics (BBS), while the Aus estimate (0.76 Mha) showed a 29% underestimation compared to BBS data (1.06 Mha). However, the results demonstrate that Sentinel-1 SAR combined with ML classifiers provides a reliable, scalable, and weather-resilient approach for multi-season rice mapping. The proposed methodology establishes a strong foundation for timely production forecasting and climate-resilient agricultural planning, particularly in cloud-prone tropical regions.