一种利用Mask R-CNN的遥感影像与矢量数据配准方法

遥感影像数据与地理信息系统（geographic information system，GIS）矢量数据的配准是遥感与GIS集成的基础。目前遥感影像与矢量数据的配准关键在于遥感影像特征的提取，而现有遥感影像特征提取方法存在特征提取不完整、配准失败和精度不高等问题。由此提出了一种基于Mask R-CNN（region-based convolutional neural network）的遥感影像与矢量数据配准方法，首先，利用Mask R-CNN模型提取影像的道路交叉口作为影像控制点; 然后，依据几何拓扑关系筛选矢量数据道路交叉口作为矢量控制点，再根据遥感影像与矢量数据控制点的欧氏距离确定同名控制点；最后，以同名控制点为基础实现遥感影像与矢量数据的配准。选取上海市矢量数据和高分二号影像数据进行配准实验，实验结果表明, 所提方法鲁棒性强、精度高。

A Registration Method of Remote Sensing Image and Vector Data Using Mask R-CNN

  Objectives  The registration of remote sensing image data and GIS (geographic information system) vector data is the basis of the integration of remote sensing and GIS, which is widely used in the fields of map data update, city monitoring, map change detection and so on. At present, the key to the registration of remote sensing images and vector data is the extraction of remote sensing image features. However, the existing remote sensing image feature extraction has problems such as incomplete feature extraction, which leads to registration failure or low accuracy. This paper proposes a registration method for remote sensing images and vector data based on Mask region-based convolutional neural network (Mask R-CNN).  Methods  Firstly, we select the road intersection as the distinctive feature of the same name in the remote sensing image and vector data, and create a road intersection image data set to train Mask R-CNN model. Secondly, according to the geometric topological relationship, the vector data road intersections are selected as vector control points. And take the intersection control point of the vector data as the center, we use 400 × 400 pixels window to crop the remote sensing image data and input it into the Mask R-CNN model, extract the border of the road intersection in the image. The control points of the same name are determined according to the Euclidean distance between the remote sensing image and the vector data control points, and the control points of the same name are cleaned using the density-based spatial clustering of applications with noise algorithm. Finally, the affine transformation parameters are calculated according to the filtered control points of the same name to realize the registration of remote sensing image and vector data. The registration data of Shanghai vector data and Gaofen-2 image data were selected for registration experiment.  Results  Experimental results show that the average deviations of the experimental data before were 15.34 m and 1.44 m, before and after registration based on Mask R-CNN.This method can correctly register remote sensing images and vector data, and the proposed method has better application prospects in urban data registration, and has the characteristics of strong robustness and high accuracy.  Conclusions  The proposed method in this paper can automatically register remote sensing images and vector data from different sources in areas such as mountains, grasslands, and deserts. In the next step, the amount of sample data of remote sensing images will be increased, including types of remote sensing images and types of features.