GLM-Net: A Mamba-Based Building Extraction Network With Global–Local Feature Collaboration for Remote Sensing Imagery

Abstract

The automatic extraction of buildings from remote sensing imagery is crucial for urban planning, change detection, and disaster assessment. Existing CNNs struggle to effectively capture global context, while Transformers are constrained by the quadratic computational complexity of self-attention. These limitations result in inadequate multi-scale feature utilization when extracting buildings with complex boundaries or irregular shapes. To address these issues, this paper proposes GLM-Net, a building extraction network based on the Mamba framework. The model uses VMamba as the backbone for efficient global semantic modeling, while introducing the Up-Mamba upsampling module for precise boundary detail reconstruction and the Pin-Mamba fusion module for enhanced multi-scale feature representation. On the Inria Aerial Image Labeling Dataset (Inria), GLM-Net achieves IoU, mIoU, and BIoU scores of 86.26%, 89.33%, and 77.55%, respectively; the corresponding scores on the WHU Building Dataset (WHU) are 91.16%, 93.10%, and 82.32%. Cross-domain experiments on the Massachusetts Buildings Dataset and the Khartoum subset of WHU-Mix further show competitive transfer performance, particularly in terms of IoU and BIoU, although the performance degradation on Khartoum indicates that challenging domain shifts remain unresolved.