Deep learning based semantic analysis of SAR imagery: From images to maps

Abstract

Semantic segmentation plays a central role in extracting detailed information regarding land cover, infrastructure, and natural disasters from SAR imagery. Semantic segmentation assigns a class label to every pixel in the image, enabling dense and context-aware mapping of semantic classes. This chapter explores how deep learning has transformed semantic segmentation from SAR images, tracing the evolution from traditional pixel-based classification to fully convolutional models and end-to-end semantic labeling pipelines. It highlights the specific challenges posed by SAR data and how these necessitate tailored preprocessing and model adaptations. The chapter also reviews state-of-the-art approaches, training strategies, and evaluation protocols, before presenting a practical case study demonstrating how U-Net-style architectures can be effectively applied to SAR imagery for flood mapping. As new SAR constellations launch and machine learning methods evolve, semantic segmentation will continue to be a key enabler for large-scale, reliable analysis of SAR data in both scientific and operational Earth observation.