Semantic segmentation of aerial image using fully convolutional network

Dense semantic segmentation is an important task for remote sensing image analyzing and understanding. Recently deep learning has been applied to pixel-level labeling tasks in computer vision and produces state-of-the-art results. In this work, a fully convolutional network (FCN), which is a variant of convolutional neural network (CNN), is employed to address the semantic segmentation of high resolution aerial images. We design a skip-layer architecture that combines different layers of features in aerial images. This structure integrates the semantic information from deep layer and appearance information from shallow layer to make better use of the aerial image features. Moreover, the FCN can be trained end-to-end and produce segmentation output correspondingly-sized as the input image. Our model is trained on the extended GE-4 aerial image dataset to adapt FCN to the aerial image segmentation task. A full-resolution semantic segmentation is produced for each testing aerial image. Experiments show that our method obtains improvement in accuracy compared with several other methods.