SABNet: a scale-aware building extraction network for high-resolution remote sensing imagery

Accurate building extraction from high-resolution remote sensing imagery is fundamental for urbanization monitoring, disaster assessment, and sustainable urban development. Although deep learning–based approaches, particularly convolutional encoder–decoder architectures, have achieved remarkable progress, they still struggle to handle the diversity of building styles, variations in building scales, and the presence of visually similar background objects such as roads. These challenges often lead to missed detections and false alarms in complex urban environments. To address these issues, we propose a novel Scale-Aware Building Extraction Network (SABNet) that fully leverages multi-level encoder features for precise building extraction. Specifically, a Scale-Aware Module (SAM) is designed to employ different convolutions with different dilation rates, enabling the model to capture buildings across multiple scales. Subsequently, a Multi-Scale Self-Attention Module (MSSAM) is introduced to model local dependencies within overlapped windows, effectively mitigating the interference of diverse building styles and similar backgrounds. Finally, a Multi-Level Joint Interpretation Module (MJIM) is developed to adaptively integrate hierarchical features, producing structurally consistent and detail-preserving building maps. Extensive experiments conducted on three benchmark datasets demonstrate that the proposed method achieves superior performance compared with state-of-the-art models, particularly in scenes with complex urban structures and diverse building distributions.