An improved architecture for urban building extraction based on depthwise separable convolution

Accurate extraction of urban buildings is a key problem in urban remote sensing image processing. It can be applied to many kinds of urban problems, such as data statistics of urban management and smart cities. In recent years, the deep learning model based on convolutional neural network is widely used in the field of target recognition and semantic segmentation. In this paper, based on U-Net for urban building extraction from remote sensing image, we propose a neural network architecture for urban building extraction from remote sensing image. We use depth separable convolution to improve it and adjust the process of network super parametric optimization according to the characteristics of building. We call this new architecture XU-Net. We evaluate the performance of XU-Net through experiments with INRIA aerial image data set. The result shows that XU-Net is not only feasible but also efficient. Moreover, XU-Net reduces number of parameters 89%, from 18.8M to 2.13M, compared to classical architecture U-Net, at the same time, it guarantees the accuracy can reach 97.5%.