Attention U-Net for Road Extraction in Remote Sensing Images

Abstract

The reliable road network plays a vital role in many applications. Owing to the development of remote sensing technology and the success of deep learning in computer vision, automatic road extraction from remote sensing images is a research hotspot in recent years. However, due to the complicated image background and special road structure, the results of automatic road extraction are still far from perfect. In this paper, we propose a road segmentation network that is designed based on improved U-Net, which contains an encoder and a decoder. First, the recurrent criss-cross attention module (CCA) is introduced into the encoder to obtain long-range contextual dependencies with a relatively small number of computations and parameters, which results in better understanding and expression of image information. Second, we propose the attention-based multi-scale feature fusion module (AMS) to resolve the problem of different shapes and widths of the roads, which is placed between the encoder and decoder and uses attention mechanisms to guide multi-scale information fusion. Experimental on the Massachusetts Roads Dataset show that the proposed method achieves better performance in road extraction than other methods in terms of precision, recall, F1-score, and accuracy.