Research On The Registration Of Vector Data And Remote Sensing Image Based On Deep Learning

Volunteered Geographic Information(VGI)has developed rapidly with the advancement of sensors and network technologies,making it a reality to obtain free vector data covering the world.It has the characteristics of fast update,free download,wide coverage,etc.,which provides a useful supplement for the data collection and update of basic geographic information.At the same time,with the vigorous development of remote sensing and photogrammetry technology,the efficient and rapid acquisition of multi-temporal,high-resolution,and multi-spectral digital images worldwide has become a reality.Domestic remote sensing satellites such as the Gao Fen series and the Zi Yuan series can achieve all-weather,high-precision,and efficient earth observation.my country's remote sensing satellite technology and information acquisition capabilities are at the world's advanced level,and the high-resolution remote sensing images provided by remote sensing satellites are also increasing rapidly.The continuous increase in the amount of spatial geographic data has caused the accumulation of geographic data update,change detection,and information extraction.Regardless of geographic data update,change detection,or remote sensing image information extraction,the prerequisite steps are the registration of spatial data.Therefore,studying the automatic registration of vector data and remote sensing images has important theoretical and practical significance.At present,the key to the registration of vector and remote sensing images lies in the extraction of remote sensing image features.However,the existing remote sensing image feature extraction methods have problems such as incomplete feature extraction and extraction errors,resulting in registration failure or low accuracy.This article mainly proposes a method of vector data and remote sensing image registration based on deep learning.The improved Faster R-CNN target detection model is used to extract the road intersection of the image as the image control point,and the vector data road intersection is selected as the vector control point according to the attribute information and geometric topological relationship.Then,analyze the offset distance between the vector and the image control point to establish a control point with the same name of the vector and the image.Finally,use the control points of the same name to construct the mapping relationship between vector data and remote sensing image to realize the automatic registration of vector and image.Around the research purpose of this article,the research of this article mainly includes the following aspects:(1)According to the vector and image data characteristics,select the road intersection as the vector and image matching feature with the same name.It is proposed to extract road intersections in vector data by constructing filter conditions such as attribute information,topological relationship,connectivity relationship,and intersection angle of vector data.In experimental areas such as Guangzhou,the screening conditions constructed in this paper can accurately extract the vector road intersections in the study area.(2)Propose an improved Faster R-CNN target detection algorithm to extract road intersections in the image.Among them,the feature pyramid network is used to construct the backbone extraction network for multi-scale road intersection feature extraction,the attention mechanism is introduced to enhance useful information,while suppressing useless information,and ROI Pooling is replaced with ROI Align to avoid the accuracy when the size of the candidate frame feature map is fixed.loss.The commonly used target detection algorithm is selected as the control group.The improved Faster R-CNN target detection model in the test set Chinese can extract road intersections more accurately than the commonly used target detection model.(3)The offset distance of the road intersection is extracted by analyzing the vector and the image,and the point with the same name of the vector and the image is constructed.At the same time,the clustering algorithm is used to clean the data of the points with the same name,the coordinate values of the points with the same name after the data cleaning are substituted into the affine transformation function,and the affine transformation parameters are solved by the least square method.Finally,perform affine transformation on the vector data to complete the automatic registration of the vector and the image.Select Nanchang,Xi'an and other places as the experimental area to carry out the vector and image automatic registration experiment.In the experimental area,the method can accurately realize the automatic registration of vector and image.(4)Considering the spatial difference of the experimental area,there is a problem of inconsistency in the spatial distribution of vector and image offset errors.A method for automatic registration of block vector and image is proposed.By analyzing the point error of the same name between the vector and the image,a Thiessen polygon is established to divide the study area into local areas where the vector and the image offset error are consistent.Perform vector and image registration in each local area,and stitch the local registration results to complete the vector and image registration work of the entire study area.Select cities such as Suzhou where the spatial distribution of vector and image migration errors are inconsistent as experimental areas,and use the block vector and image registration method to perform automatic registration experiments.The experimental results show that the block registration method can effectively solve the vector and image migration.The problem of inconsistent distance spatial distribution.