HSRoadNet: Hard-Swish Activation Function and Improved Squeeze–Excitation Module Network for Road Extraction Using Satellite Remote Sensing Imagery

Abstract

Road information plays an essential role in many fields. To prevent failed extraction of heterogeneous regions, fracture of extracted roads and others resulted from vehicles and trees when using very high resolution remote sensing images; a remote sensing image road extraction method based on Hard-Swish Squeeze–Excitation RoadNet is proposed in this article. First, road extraction task is divided into three correlated subtasks to reduce the impact of vehicles and trees in road extracting. Second, a normalization layer is adopted to prevent gradient levels from vanishing and exploring and avoid fracture of the extracted road. Then, adopting Hard-Swish activation function to improve the accuracy of road extracting, and then finally, using the improved squeeze–excitation module to make the trained net a full use of the characteristic information of the road that do not increase excessive capacity. Comparison experimental results indicate that, in various indicators, the proposed method performs serviceably, it, respectively, increased by 16.8%, 2.2%, 1.5%, and 8.5% over the suboptimal in F-score, global accuracy, class average accuracy, and recall ratio. The mean intersection over union (MIoU) value of the proposed method was the suboptimum with a disparity of 0.2% from the optimal. Ablation experiments show that the proposed method performs best in various indices, and the global accuracy, MIoU, class average accuracy, and recall rate are improved by 0.5%, 0.1%, 0.5%, and 0.2%, respectively, compared with the suboptimal method. The F-score is suboptimal, with a 0.3% difference from the best.