Abstract:  With the improvement of spatial resolution of optical remote sensing images and the enrichment of acquisition channels, optical remote sensing images has become an efficient technological method to achieve intelligent interpretation of land features. Due to the powerful feature extraction ability of convolutional neural networks (CNN) and the demand of road information in many fields, road extraction methods based on CNN have become a current research hotspot. In view of this, this paper summarizes the road extraction method based on CNN from four aspects: Improvement of shape features, improvement of connectivity, improvement of multiscale features and improvement of extraction strategy according to the relevant research literature in recent years. Then, we describe typical road occlusion cases and use classical CNNs to analyze and validate the current technical difficulties at the level of limitations of sample labels. Finally, the development trends of road extraction from remote sensing images are outlooked from four aspects, namely, multisource data synergy, sample library construction, weakly supervised modeling and domainadaptive learning.

Key words: convolutional neural network, optical, remote sensing image, road extraction, intelligent interpretation