Land use land cover segmentation using synthetic aperture radar data on Dhaka division, Bangladesh

Bangladesh’s agrarian economy relies heavily on agriculture, contributing 19.3% of its GDP.1 Rice, the staple crop, is crucial for national food security. Monitoring land use land cover (LULC) and paddy production are essential due to threats from climate change, natural disasters, and urbanization, ensuring efficient resource allocation and preventive measures. However, cloud cover from May to October coincides with major rice cultivation seasons, limiting optical remote sensing. This paper proposes deep learning approaches using Synthetic Aperture Radar (SAR) images that penetrate clouds, for crop monitoring in Bangladesh. We analyzed Dhaka division using SAR images from Copernicus Open Access Hub and an annotated BD-SAT high resolution ground truth. Training and testing datasets were input into UNET and DeepLabV3+ models, the stat of the art deep learning architectures for image segmentation. Despite cloud cover, our models achieved promising results. The analysis shows that Unet on SAR data achieves significantly improved Intersection over Union (IoU) 0.56 & F1 score of 0.72 for farmland; and 0.26 & 0.44 for meadow compared to the Sentinel2 RGB data demonstrating SAR’s effectiveness for agricultural monitoring.