Data Augmentation Method for Extracting Partially Occluded Roads From High Spatial Resolution Remote Sensing Images

Abstract

Road extraction from high spatial resolution remote sensing images (HSRRSI) is valuable for thematic mapping, autonomous driving, and natural disaster assessment. Deep neural networks have become a powerful tool for road extraction from HSRRSI. However, these methods cannot effectively capture occlusion-independent features with insufficient occlusion samples, and usually perform poorly at extracting partially occluded roads. Existing data augmentation methods for occlusion may generate occluded backgrounds rather than occluded roads. To improve the robustness and accuracy of existing algorithms, we propose a novel data augmentation method for extracting such partially occluded roads from images. In each training mini-batch, our method selects an image to simulate occlusion with a certain probability, then randomly selects rectangle regions covering the pixels of roads and replaces these regions with pixels of non-roads in the same image. This method generates various levels of occlusion, guides the network to learn from less discriminative parts, and imparts robustness to occlusions to the model. Extensive experiments conducted with the D-LinkNet network on the Massachusetts Road Dataset demonstrated the effectiveness of our proposed method. In terms of accuracy, our approach outperforms state-of-the-art methods that use information dropping. The results also showed that optimizing the occlusion probability of images in the mini-batch and occlusion percentage of roads in the input improved road extraction performance.

Keywords

• Data augmentation
• high spatial resolution remote sensing images
• occlusion
• road extraction