A Multiscale and Multidirection Feature Fusion Network for Road Detection From Satellite Imagery

The completeness of road extraction is very important for road application. However, existing deep learning (DP) methods of extraction often generate fragmented results. The prime reason is that DP-based road extraction methods use square kernel convolution, which is challenging to learn long-range contextual relationships of roads. The road often produce fractures in the local interference area. Besides, the quality of extraction results will be subjected to the resolution of remote sensing (RS) image. Generally, an algorithm will produce worse fragmentation when the used data differ from the resolution of the training set. To address these issues, we propose a novel road extraction framework for RS images, named the multiscale and multidirection feature fusion network (MSMDFF-Net). This framework comprises three main components: the multidirectional feature fusion (MDFF) initial block, the multiscale residual (MSR) encoder, and the multidirectional combined fusion (MDCF) decoder. First, according to the road's morphological characteristics, we develop a strip convolution module with a direction parameter (SCM-D). Then, to make the extracted result more complete, four SCM-D with different directions are used to MDFF-initial block and multidirectional combined fusion decoder (MDCF-decoder). Finally, we incorporate an additional branch into the residual network (ResNet) encoding module to build multiscale residual encoder (MSR-encoder) for improving the generalization of the model on different resolution RS image. Extensive experiments on three popular datasets with different resolution (Massachusetts, DeepGlobe, and SpaceNet datasets) show that the proposed MSMDFF-Net achieves new state-of-the-art results. The code will be available at https://github.com/wycloveinfall/MSMDFF-NET .