Land cover classification of Sentinel-1 images at subpixel level using deep learning

The operational use of Sentinel-1 Synthetic Aperture Radar (SAR) images for high-resolution land cover mapping is often limited by its spatial resolution, leading to mixed pixels in heterogeneous landscapes and constraining classification accuracy. This study addresses this challenge by proposing and evaluating a U-Net-based deep learning (DL) framework for subpixel mapping (SPM) and benchmarking its performance against the state-of-the-art DeepLabV3+ architecture to generate 5 m land cover maps from Sentinel-1 Single Look Complex (SLC) data. We systematically assess five different sets of data configurations including single and dual orbit (ascending/descending), feature indices and auxiliary information to deconstruct their contributions to the overall framework. The results demonstrate that dual-orbit geometry, targeted feature engineering, and auxiliary data each play a critical role in improving class separability and spatial detail. Overall, the study shows that combining multi-azimuth SAR acquisitions with complementary features and external data is essential for overcoming inherent sensor limitations in SPM based on Sentinel-1 SAR data.