A Review of Urban Building Extraction from Synthetic Aperture Radar Imagery Based on Deep Learning

Urban building monitoring is crucial for disaster prevention, mitigation, and sustainable urban development. Synthetic Aperture Radar (SAR) is an important and unique Earth observation sensor, offering all-weather and all-day monitoring capabilities that optical sensors cannot match. However, traditional methods for extracting urban buildings from SAR images face limitations such as heavy reliance on expert knowledge for feature selection and classifier design, as well as insufficient robustness and generalizability. Recently, deep learning approaches, which have achieved remarkable success in computer vision, are increasingly becoming a major trend in SAR image urban building extraction. However, there remains a lack of systematic reviews on urban building extraction from SAR image based on deep learning. This article aims to present a comprehensive review of methodologies in this field since 2015. It includes detailed analyses and discussions of approaches, datasets, and evaluation metrics. The fundamental imaging principles of urban buildings in SAR imagery are first explored to clarify existing challenges. The methods are categorized into urban building extraction and urban building height estimation based on application perspectives, while approaches are further classified as only-SAR image based methods and SAR-optical data based methods, depending on the data sources. Additionally, given the distinct feature representations across SAR resolutions, investigations are differentiated between building-level analysis for high-resolution images and built-up area level analysis for medium-resolution data. Benchmark datasets are summarized, and evaluation metrics are reviewed. Finally, the article conducts discussions, proposes potential solutions, and outlines future development directions. It is expected that this work will provide valuable references for researchers in this field.