FAMNet: Lightweight Road Extraction Network with Fused Attention and Multi-level Cascaded ASPP

Precision road extraction via remote imagery plays a vital role in smart transportation and urban digitalization. Current deep learning methods, however, confront three key limitations: high computational complexity from intricate architectures, road fragmentation caused by weak topological continuity, and environmental interference-induced errors in thin road detection. To address the aforementioned challenges, this paper proposes FAMNet, which establishes a collaborative edge surface perception mechanism to achieve high-precision road extraction in complex scenarios while maintaining topological continuity. Firstly, a lightweight MobileNetv3 is redesigned as the backbone network to reduce computational overhead and model complexity. Then, we develop an MCASPP combined with a Convolutional Block Attention Module to expand receptive fields while enhancing extraction capability and connectivity for slender road features. Thirdly, a composite loss function combining Focal loss and Dice loss is implemented to mitigate the impact of class imbalance between roads and background samples, suppress environmental interference, and emphasize road-specific information. Experimental results for DeepGlobe and CHN6-CUG datasets demonstrate that FAMNet significantly outperforms state-of-the-art methods in both the accuracy and structural completeness of road extraction, while maintaining advantages in parameter count and computational efficiency. Furthermore, the proposed method shows promising potential for advancing natural resource surveying and monitoring through enhanced geospatial environment perception modeling capabilities.