MDNet: A Multi-modal Dual Branch Road Extraction Network Using Infrared Information

Ambient occlusion and confusing terrain are difficult problems in remote sensing road extraction. In order to improve the accuracy of road extraction, this paper introduces infrared images as the basis for road extraction and proposes a multi-modal road extraction model (MDNet) suitable for additional infrared channel remote sensing images. First, this method adds a unidirectional D-LinkNet branch that is exactly the same as the backbone based on the D-LinkNet semantic segmentation network to learn infrared images and construct multi-modal image data information. Second, the attention mechanism was introduced in the multi-modal image processing stage, and a weighted model based on the attention mechanism was proposed to improve the utilization of information about the road. Finally, a road detection dataset composed of RGB-IR four-channel remote sensing images was built, and an experiment is conducted compared with some relatively advanced deep learning methods. The experimental results show that compared with DLinkNet with RGB remote sensing image as input, the mean intersection-over-union (mIoU) of D-LinkNet with RGB-IR remote sensing image as input is improved by about 4.3%, while MDNet is improved by 2.6% compared with the RGB-IR version of D-LinkNet. It can be seen that the multi-modal road extraction model MDNet combined with infrared information proposed in this paper can achieve more accurate road extraction from remote sensing images.