Street Block Classification Based on Urban Satellite Images

Due to its powerful representation capabilities, the convolutional neural network can naturally be applied to the scene classification of satellite images. In this paper, we consider the problem of the street block classification of urban satellite images. The street blocks are naturally formed by the urban road network of a city. However, because the roads in a city are usually not parallel, the street blocks have great variability in shape, scale and size. What is worse, if the street blocks in the satellite image are directly classified, these blocks will inevitably carry boundary information, which is also different from the traditional scene classification of satellite images. In order to address these problems of street block classification, we have designed a novel network structure based on the residual network. The experimental results show that our network architecture performs excellent and gets better classification results than others.