Lightweight Multimodal Road Extraction With Trajectory-Mask-Guided and Multidirectional Cross-Layer Aggregation

Multimodal methods that integrate GPS trajectory information help improve the accuracy of road extraction from remote sensing images. However, existing methods often overlook the quality of trajectory data and directly feed it into dual-stream networks, which tends not only to introduce noise and affect the stability of the model, but also to significantly increase computational cost.Moreover, traditional skip connections are limited in modeling deep semantic relationships across contextual features. To address these issues, this paper proposes LMGANet, a lightweight multimodal road extraction network based on trajectory-mask guidance and multi-directional cross-layer aggregation. We design a Trajectory-Mask Guided Module (TMGM), which enhances the reliability of trajectory information by leveraging stage-wise prediction maps within a single backbone network, and guides the model to focus on key regions across both channel and spatial dimensions. TMGM enables efficient multimodal collaboration under a single-stream architecture, effectively balancing performance improvement and lightweight model design. Additionally, we propose a Multi-directional Cross-layer Aggregation Pyramid (MCAP), which leverages hybrid dilated strip convolutions to effectively fuse multi-scale directional information between encoder and decoder features. Experimental results on two public datasets, BJRoad and Porto, demonstrate that LMGANet outperforms existing methods in terms of accuracy. Compared to the latest dual-stream multimodal method DEFNet, our LMGANet employs a single-stream architecture, reducing parameters and FLOPs by 65.45% and 57.04%, respectively, while delivering superior performance with substantially lower resource consumption. The source code will be made publicly available at https://github.com/lgg0606/LMGANet.