Multi-scale Convolutional Neural Network for Road Extraction in Remote Sensing Imagery

Road is an important semantic region in remote sensing imagery and plays an important role in many applications. Deep learning has obtained a great success in image classification, since it can directly learn from labeled training samples and extract different level image features to encode the input image. In this paper, we propose a multi-scale convolutional neural network (MSCNN) for extracting road from high-resolution remote sensing image, in which road detection can be seen as a regional classification. This core trainable detection engine consists of an encoder-decoder network and a fusion model. Firstly, image was encoded as a feature representation with several stacked convolutional layers. And then guided by the different scale road training data, the pre-trained decoder networks output a series of classification maps. Finally, we investigate the fusion model utilizing different scale classification maps and obtain a final road decision map. To validate the performance of the proposed method, we test our MSCNN based method and other state-of-the-art approaches on two challenging datasets of high-resolution images. Experiments show our method gets the best results both in quantitative and qualitative evaluation.