Research And Application Of Building Extraction Based On High Resolution Remote Sensing Image

With the rapid development of domestic remote sensing satellites,the research and application of remote sensing images are progressing,which plays a vital role in various fields.The characteristics of high spatial resolution and high temporal resolution of high-resolution remote sensing images make great achievements in the research of target detection and change detection.The detection results can be used in urban planning,dismantling,marine monitoring and other aspects,which are of great significance.However,most researchers now use foreign remote sensing image data to carry out target detection experiments,and due to the limitations of data sources,human and material resources,the amount of data and the image type of the data set need to be improved.Target detection based on deep learning enables the machine to analyze the image,extract the features of the target object,and get the region of interest,which is the main method for detecting instances in images in recent years.Application of remote sensing images in the actual production and living is important,because many enterprise personnel is not professional researchers,for the use of remote sensing image process and results of target detection approach is not understanding,caused the delinked research and practical application,the emergence of remote sensing image interpretation system can better link algorithm experiment and practical use.It is beneficial to the progress and development of remote sensing data.This paper based on the domestic high-resolution remote sensing satellite data,studies the annotation method,dataset,building detection method and remote sensing interpretation system.The main work contents are as following:(1)According to the characteristics of high-resolution remote sensing images,put forward a set of annotation method oriented multi-type instances of sample in remote sensing data.The annotation information can be corresponding to the panchromatic image,multispectral image and fusion image at the same time by using geographic information registration.This method makes the efficiency of annotation is two times higher than that of traditional patterns,and close to the edge of the target instance when adopted labelling.It reduces the background and irrelevant information in the annotation box,makes the annotation result more refined,and improves the annotation quality of individual instances.According to the proposed annotation method,we established a building large dataset based on the domestic remote sensing image data named 5M-Building,include five scene categories,house buildings,factory buildings,greenhouses,other Ⅰ and other Ⅱ.The number of instances of dataset reached five million,which is currently the largest known for building remote sensing datasets.(2)A building target detection method based on image feature complementarity is proposed.This method based on Faster R-CNN,taking the characteristics of panchromatic image and multispectral image as well as the complexity of remote sensing image information,the high-resolution panchromatic image and rich color information of multispectral image are combined through Gram-Schmidt orthogonalization.After feature extraction in the backbone network,the multi-scale feature extraction method is used to reduce the identification error caused by the large-scale difference of building samples.The experimental results show that the proposed dataset has achieved a good performance.(3)A set of remote sensing image interpretation system has been designed and developed,include the basic function such as implementation application,download and visual images at the same time.The system combines the target detection algorithm based on image feature complementarity and other target detection algorithms with the integration of system,through the system the operation can invoke server-side target detection algorithm model of training and testing.Users can select images and algorithms in the system for experiments and get visual detection results,so that research and practical applications are linked,making image management and application more convenient.