Multi-Scale Mixture-of-Experts With Lora for Building Extraction from Optical Remote-Sensing Images

Due to the high density of buildings, diverse illumination conditions, and varying scales in complex urban scenes, building extraction from high-resolution optical remotesensing images faces significant challenges in capturing multi-scale information and effectively utilizing spatial context. To address these issues, we propose MSLoRA-Net, a Multi-Scale LoRAEnhanced Building Extraction Network, designed to improve extraction accuracy and robustness. MSLoRA-Net introduces a novel hybrid expert LoRA (Low-Rank Adaptation) mechanism combined with a router network for dynamic parameter selection, enabling efficient and adaptive fine-tuning based on the Segment Anything Model (SAM). This enhances the model's ability to focus on building targets. To overcome the limitations of ViT in dense prediction tasks, an Auxiliary Multi-Scale Encoder (AME) is incorporated to capture global context while refining local detail modeling. In addition, a Feature Injection and Extraction Module is introduced to optimize interaction between dual encoders, ensuring effective integration of global and local features for better adaptation to complex building structures. Furthermore, a semantic-aligned progressive upsampling decoder is designed to restore spatial resolution while leveraging semantic alignment strategies, reducing boundary blurring in high-resolution feature reconstruction. Experimental results demonstrate that MSLoRANet achieves superior building extraction accuracy across multiple public datasets. This study also represents an innovative application of the SAM foundation model to building extraction tasks, verifying its great potential in remote-sensing image analysis and providing valuable insights for future research on efficient and precise building extraction.