Abstract: As an important part of image interpretation and analysis, segmentation of remote sensing images has been widely researched. However, traditional segmentation method based on hand-crafted features has its limitations on accuracy and generalization, state-of-the-art methods are mainly relied on deep learning in recent years. In this paper, we propose a new segmentation method based on multi-scale deep residual neural networks, which aims at improving segmentation accuracy, especially on small-scale objects. We frstly utilize Residual Network (ResNet) and transform it to fully convolution networks (FCN), in which, Atrous convolution is introduced during the up-sampling process to ensure the feld of view on each layer. Then we add multi-scale data augmentation to improve the robustness for small objects. The proposed approach is applied on ISPRS 2D Vaihingen semantic labeling contest dataset, and yields high accuracy at 89.7%, outperforming most state-of-the-art methods.

Key words: semantic segmentation of remote sensing image, deep residual network, Atrous convolution, multi-scale data augmentation