GL-MambaLite: A Lightweight Global–Local Feature Enhanced Network for Building Extraction

As a key element of human settlements and fundamental features in maps, buildings play a critical role in urban planning and environmental monitoring. Automatically extracting buildings from remote sensing images is of significant practical value. However, accurately extracting buildings is challenging due to their diverse scales, complex geometries, and spectral similarities to backgrounds. To effectively capture fine-grained structural details and long-range spatial dependencies, current deep learning methods mainly rely on complex hybrid architectures combining convolutional neural networks (CNNs) and Transformer-based models. Although these methods achieve high accuracy, their substantial computational burden limits their scalability, making them impractical for large-scale building mapping tasks. In addition, while sophisticated multiscale feature extraction modules and fusion strategies help accommodate scale variations, their redundant parameters and elaborate structures further exacerbate the computational demands. Consequently, balancing building extraction accuracy and operational efficiency remains a critical challenge for practical deployment, particularly in resource-constrained scenarios. To address these limitations, we propose a lightweight global–local feature enhancement (GLFE) network, GL-MambaLite, for building extraction. Specifically, we construct a lightweight GLFE block that combines Mamba and CNN to simultaneously capture global dependencies and local details. Furthermore, to mitigate complex background disturbances, we design a lightweight parallel multiscale feature extraction block based on depth-wise convolution, enabling the efficient aggregation of multiscale features. Moreover, we introduce a holistic grouped aggregation strategy to facilitate direct interaction between semantic and detailed features, effectively reducing redundancy while preserving geometric integrity. Extensive experiments on the WHU, Inria, and custom UAV building datasets demonstrate the superiority of GL-MambaLite. Experimental results demonstrate that GL-MambaLite achieves the highest IoU scores across all three datasets, recording 90.38% on the WHU dataset, 79.71% on the Inria dataset, and 84.65% on the custom UAV dataset. With an inference speed of 593.79 FPS on 512 × 512 inputs, GL-MambaLite establishes an optimal equilibrium between accuracy and efficiency. Compared to the state-of-the-art Transformer-based model (SegFormer), GL-MambaLite reduces computational complexity (FLOPs) by 42.8% and improves inference speed by 52.2%, while also achieving a 7.5 times speedup over the recent Mamba-based baseline (RS3Mamba). Ablation studies confirm the architectural efficacy of our proposed components.