Strengthening National Food Security: An Operational Machine Learning Framework for Accurate Multi-Temporal Rice Mapping in Central Bangladesh

Safeguarding Bangladesh's foodgrain self-sufficiency requires dynamic and precise agricultural monitoring, an objective historically constrained by the inefficiencies of conventional field surveys. To address these limitations, this study develops a robust machine learning framework to map Boro rice in the Manikganj district by integrating multi-temporal Sentinel-1 Synthetic Aperture Radar (SAR) and Sentinel-2 optical imagery. Implemented within the Google Earth Engine (GEE) platform, the methodology utilizes Sentinel-1 VH polarization to detect unique backscatter reductions caused by specular reflection during the flooded transplantation phase (late February to early March). This effectively overcomes the persistent challenge of seasonal cloud cover that hampers traditional optical monitoring. The SAR data is then combined with high-resolution spectral indices, including the Normalized Difference Vegetation Index (NDVI), Enhanced Vegetation Index (EVI), and Bare Soil Index (BSI), derived from Sentinel-2 during peak vegetative growth to accurately capture the distinct phenological signatures of paddy rice.

To isolate active croplands and reduce spectral noise from non-agricultural features, an ESA Worldcover 10 m agricultural mask was applied. A pixel-based Random Forest algorithm was trained and validated using 200 field-verified data points per sub-district, which were collected via the Geofairy mobile application and partitioned into a 70:30 split. The multi-sensor framework demonstrated high classification performance, achieving overall accuracies ranging from 88.2% to 94.8% across the seven sub-districts. At the regional scale, the total satellite-estimated area showed a minimal variance of only 2.34% when compared with official government statistics. To facilitate practical agricultural planning, the framework was operationalized through "RiceMapEngine," a customized web-based GIS application. This study demonstrates that the integrated multi-sensor machine learning framework offers a highly accurate, objective, and operational alternative to conventional eye-estimation methods, providing policymakers with a science-guided tool for evidence-based land-use planning and foodgrain balancing.