Road network extraction in high-resolution SAR images based CNN features

Yue Li, Rong Zhang, Yunfei Wu

Abstract: In this paper, we proposed a convolutional neural network (CNN) features-based framework for road network extraction in high-resolution synthetic aperture radar (SAR) images. First, in consideration of rich structure information of road areas in high-resolution SAR images, a CNN model is proposed to extract road-area features and detect road candidates. The CNN model helps to improve the accuracy of the road candidates detection at the feature level. Then, an improved Radon transform and a Markov random field (MRF) are used to complete global road network extraction based on the detected road candidates. The experimental results on TerraSAR-X high-resolution SAR images over Beijing show that our approach outperforms the traditional methods.