Convolutional encoder-decoder network for road extraction from remote sensing images

In this paper, we propose a convolutional neural network, which is based on down sampling followed by up sampling architecture for the purpose of road extraction from aerial images. Our model consists of convolutional layers only. The proposed encoder-decoder structure allows our network to retain boundary information, which is a critical feature for road identification. This feature is usually lost when dealing with other CNN models. Our design is also less complex in terms of depth, number of parameters, and memory size. It, therefore, uses fewer computer resources in both training and during execution. Experimental results on Massachusetts roads dataset demonstrate that the proposed architecture, although less complex, competes with the state-of-the-art proposed approaches in terms of precision, recall, and accuracy.