Road feature enhancement network for remote sensing images based on DeepLabV3Plus

Extracting road information from complex high-resolution remote sensing images to update road networks has become a focus research field in recent years. However, the scale of roads in remote sensing images varies greatly, often leading to obstructions caused by objects such as tree shadows and buildings. These factors contribute to incomplete and discontinuous extraction results of narrow and long roads. To solve the above problems, a road feature enhancement network based on DeepLabV3Plus network is proposed in this paper. The network introduces the dense atrous spatial pyramid pooling (DenseASPP) module incorporating a strip pooling branch and channel attention mechanism. The atrous spatial pyramid pooling (ASPP) module of the baseline network is replaced by the improved DenseASPP, which strengthens the network to aggregate multi-scale road features and improves the model’s ability to recognize long and narrow roads. In addition, the cascade feature fusion (CFF) unit is utilized to fuse the shallow features of different resolutions, enhance the context awareness of the network, and improve the generalization ability of the model. Subsequently, a road feature enhancement module (RFEM) is designed in the decoder part, which uses four strip convolutions in different directions to capture remote context information and avoid the interference of irrelevant regions. The experimental results on the open DeepGlobe dataset and CHN6-CUG dataset show a significant improvement in IoU and F1-Score compared to the baseline network. The proposed method outperforms other mainstream networks in road segmentation.