DSMENet: A Road Segmentation Network Based on Dual-Branch Dynamic Snake Convolutional Encoding and Multi-modal Information Iterative Enhancement

Road segmentation from remote sensing images plays an important role in basic map data processing and services. However, roads in remote sensing images are characterized by long and narrow spans, intricate topological structures and being easily obscured, making road segmentation a challenging task in the field of re-mote sensing image object segmentation. To improve the accuracy and connectivity of road segmentation, this paper proposes a method based on Dual-branch dynamic Snake convolutional encoding and Multi-modal information iterative Enhancement (DSMENet). The multi-modal data are first encoded separately by dual-branch dynamic snake convolution encoders to adaptively focus on slender and winding local structures, accurately capturing the features of tube-like roads; next, attention driven feature fusion of multi-modal features are performed at different stages of the encoders, which are then input into the decoder for spatial resolution restoration. Finally, a multi-modal information iterative enhancement module is embedded at the end of the network to fully exploit spatial detail features of original multi-modal data and enhance the features at the end of the de-coder, thereby improving the connectivity of road segmentation. Experimental evaluations on the BJRoad dataset demonstrate that (1) The dynamic snake convolution enables the model to focus on tube-like roads effectively, resulting in a significant reduction in false alarms and an improvement in road segmentation accuracy. (2) The multi-modal information iterative enhancement module can provide supplementary spatial detail information to the road segmentation results, mitigating the effects of shadow occlusions and enhancing the connectivity of road segmentation.