Scattering-Based Land Cover Segmentation of ESA BIOMASS P-Band PolSAR Data Using Deep Learning

The European Space Agency's (ESA) BIOMASS mission introduces the world's first spaceborne P-band Polarimetric Synthetic Aperture Radar (PolSAR), enabling enhanced observation of vegetation structure, forest biomass, and subsurface scattering processes. Despite its potential for Earth observation, the application of P-band PolSAR data to semantic land-cover mapping remains relatively unexplored. This study presents a physics-guided deep learning framework that integrates polarimetric scattering information with advanced semantic segmentation architectures for automated land-cover mapping using ESA BIOMASS observations. Single Look Complex (SLC) P-band PolSAR data acquired over Dehradun, India, were processed through polarimetric calibration, speckle suppression, and terrain correction, followed by Yamaguchi four-component decomposition to derive physically interpretable surface (Ps), double-bounce (Pd), volume (Pv), and helix (Pc) scattering components. The decomposition features were examined across spatially aligned reference land-cover classes to assess their physical consistency and were subsequently used as input features for multiple semantic segmentation models, including U-Net ResNet34, DeepLabV3+ ResNet50, DeepLabV3+ ResNet101, DeepLabV3+ EfficientNet-B4, and SegFormer-B4. Experimental results demonstrate the capability of decomposition-derived PolSAR features to represent diverse scattering mechanisms and support discrimination among forest, agricultural land, water, urban areas, hill forest, and dry riverbed classes. Comparative evaluation further highlights the potential of transformer-based architectures, particularly SegFormer-B4, for capturing spatial and contextual relationships in P-band SAR imagery. By integrating polarimetric scattering physics with deep learning, the proposed framework provides a scalable methodology for semantic interpretation of P-band observations and supports potential applications in Earth observation, environmental mapping, biomass-related studies, and geospatial intelligence. The study provides an early investigation of deep learning-based semantic segmentation using ESA BIOMASS P-band PolSAR data and offers insights for future large-scale Earth system monitoring and land-cover analysis.