Road Segmentation in High-Resolution Images Using Deep Residual Networks

Automatic road detection from remote sensing images is a vital application for traffic management, urban planning, and disaster management. The presence of occlusions like shadows of buildings, trees, and flyovers in high-resolution images and miss-classifications in databases create obstacles in the road detection task. Therefore, an automatic road detection system is required to detect roads in the presence of occlusions. This paper presents a deep convolutional neural network to address the problem of road detection, consisting of an encoder-decoder architecture. The architecture contains a U-Network with residual blocks. U-Network allows the transfer of low-level features to the high-level, helping the network to learn low-level details. Residual blocks help maintain the network's training performance, which may deteriorate due to a deep network. The encoder and decoder structures generate a feature map and classify pixels into road and non-road classes, respectively. Experimentation was performed on the Massachusetts road dataset. The results showed that the proposed model gave better accuracy than current state-of-the-art methods.