Mapping Rice Area Using Sentinel-1 SAR Data and Deep Learning

Rice is the primary staple food for the majority of the world’s population, and accurate monitoring of rice planting area plays a crucial role in the implementation of SDG 2 "Zero Hunger". To map rice area in 856 Farm, time series Sentinel-1 SAR data from 2019, 2020 and 2021 were collected and preprocessed to create a 10 m dataset. Subsequently, a U-Net based model, named ResUNet-CBAM, was proposed by integrating with the residual module and the embedded attention mechanism module. Compared to the U-Net and ResUNet models, the ResUNet-CBAM model exhibited a slight improvement while still achieving sufficiently high mapping accuracy. To transfer the trained ResUNet-CBAM model to the year 2020 and 2021, a group–mean mechanism was proposed. The Sentinel-1 SAR images in each year were grouped into 15 groups, taking into account the revisit cycle. Then the trained ResUNet-CBAM model was transferred to map rice and still exhibits good classification results. Based on the classification results of 2019, 2020 and 2021, the rice plant area in 856 farms remains stable and plays a significant role in the region’s granary.