首页 > , Vol. , Issue () : 1993-2002
Road extraction from remote sensing image has important application value, and it is also a hot issue in the field of remote sensing. In recent years, the geometric texture of the target objects of aviation and satellite optical images is more refined in the wake of the rapid development of aviation technology, which provides a sufficient basis for automatic extraction of road information. However, it is still difficult to fully automate road extraction by existing methods. In view of this, this paper collects and synthesizes the existing methods based on a large number of related literatures in recent years. Ultimately, these road extraction methods are divided into four categories: template matching, knowledge-driven, object-oriented and deep learning. Template matching methods can be generally divided into rule template and variable template according to template type. The difference between the two types is whether the template can be drawn with regular graphics. Template matching mainly consists of the following three steps: (1) template design. The template can usually be set manually or by certain rules; (2) measure analysis. The target template is given. Furthermore, the region extreme value is found by the measure function within the defined area; (3) location updating, that is to say, location of road centerline is dynamically updated. Road, as an artificial ruled feature, has a large amount of relevant knowledge. Hence, knowledge-driven methods based on relevant knowledge are used to road extraction work. According to the relationship between knowledge and road, this paper divides the knowledge-driven methods in three kinds: geometric knowledge, context knowledge and auxiliary knowledge. Three methods are respectively described as follows: (1) Geometric knowledge. The model is mainly constructed based on the geometric features of road; (2) contextual knowledge. The method utilizes auxiliary knowledge (motor vehicles, trees, zebra crossings, etc.) related to road to assist in identifying the road; and (3) assisting knowledge. The road extraction is guided by using multi-source remote sensing data, vector data, GPS data, navigation data, and public source data. With the rapid increase of spatial resolution of optical remote sensing images, object-oriented methods have gradually become one of key methods in road extraction. Firstly, the method selects the spectral and geometric feature criteria to segment image. Secondly, it uses the geometric radiation similarity criterion to classify the segmentation region. Finally, the road extraction results are output by post-processing methods such as mathematical morphology, matching tracking and tensor voting. Road extraction research belongs to the problem of remote sensing image interpretation, while the deep learning supplies a new opportunity to satellite image interpretation in semantic direction. Deep learning extracts road information with high precision through the procedure of convolution, pooling, and training. Nevertheless, there still has problems of road breakage and mistaken extraction. At the end of this paper, we prospect the direction of development of road extraction and propose two main trends: (1) multi-method complementary road extraction system; (2) deep integration of deep learning and traditional methods. Thus, the detection area should be studied further.