Machine Learning-Based Paddy Field Mapping During Aman Season in Bangladesh Using Sentinel-1 SAR Data

Abstract

Bangladesh’s food security is closely tied to rice production. Conventional field-based surveys are labor-intensive, time-consuming, and spatially limited, whereas remote sensing provides a more efficient alternative. In Bangladesh, Aman season (July–December) representing the highest production period, particularly in the Barishal region, which is a dominant Aman rice producing region. Persistent cloud cover during this period restricts the use of optical satellite imagery. Therefore, this study utilizes Sentinel-1, C-band Synthetic Aperture Radar (SAR) data, which enables all-weather observation. Hence, this study aimed to map paddy fields during the Aman season in Bangladesh using Synthetic Aperture Radar (SAR) data and a Machine Learning (ML) method. For the Aman 2024 season, ground truth geolocation points were collected for four land cover classes: paddy, non-paddy vegetation, urban areas, and water bodies. Only 58 ground-truth samples for each land-cover class were collected due to limited field accessibility, fragmented land ownership patterns, and the relatively small extent of continuous paddy fields in the study area. Thirteen SAR-derived parameters were generated from whole-season feature sets. However, only four key parameters identified by the feature selection technique were used for machine learning model development, namely the average VH values from the early stage, late stage, and whole season, and the minimum VH value from the early stage. To enhance model reliability despite the limited sample size, 5-fold cross-validation was employed. Using these selected parameters, the Naive Bayes model with a Gaussian kernel achieved the highest accuracy (84.9%), followed by the SVM model with a Medium Gaussian kernel (83.3%) and the Tree model with both Fine and Medium kernel (78.5%). Overall, the selected parameters supported robust machine learning performance, with Naive Bayes yielding the best results, and demonstrated that the optimized parameter set can effectively map paddy fields during the Aman season.