A Deep Learning Building Extraction Method Based On Geographical Condition Monitoring

Remote sensing,in a broad sense,is a means of information acquisition.It uses the non-contact sensor to obtain the spatiotemporal information of the target remotely,provides the geometric and physical characteristic information of the target object,and provides the data basis for the understanding and interpretation of the target object.Remote sensing image is an important data result,which can be divided into aerial image and satellite image according to the platform.Among them,the high-resolution satellite remote sensing image covers a wide geographical range and contains abundant ground feature information,so it can observe the actual situation of the earth's surface from a global perspective.Over the past few decades,the research of remote sensing science has made continuous progress.However,due to the complexity of remote sensing image content and the relatively backward processing means,the accuracy of current remote sensing image classification,target extraction and other technical methods has been difficult to meet the requirements of production practice,and there is still a considerable distance from actual application.Geographical situation monitoring is a process of dynamic statistics,analysis and synthesis of geographical data information.It combines a variety of modern surveying and mapping technologies to provide units at all levels with timely,accurate and reliable information on geographical conditions from a macro perspective.Land cover data is an important achievement of the monitoring of geographical conditions,which describes the actual situation of geographical conditions in vector form.It is one of the most critical tasks to timely and accurately discover and update the changes of ground cover.Research of remote sensing image classification and remote sensing image target monitoring is an important goal is to help find the change of the surface coverage,provide power for geographical monitoring the state of the union,but due to the automation techniques to meet the requirements,the current update work still mainly by artificial visual interpretation,to a large extent rely on the workers experience and subjective judgment.In recent years,deep learning technology has risen rapidly,and the image processing method based on convolutional neural network has been rapidly promoted,bringing new research ideas and vitality to many traditional fields.This paper mainly studies how to apply the two-stage target detection method based on deep learning to the extraction of buildings in high-resolution remote sensing images by combining the surface coverage data in the national conditions monitoring.The main contents of this paper are as follows:(1)expound the basic discipline theory related to the research content of this paper.In this paper,the development process of convolutional neural network and twostage target detection model is introduced in detail,and the basic principles and improvements of r-cnn,Fast r-cnn,Faster r-cnn and Mask r-cnn models are described.Then,it introduces the basic concept of geographical condition monitoring,the traditional building extraction method of high resolution remote sensing image and the general process of building extraction combining remote sensing image and land cover data.(2)design and implement a deep learning platform oriented to geographical conditions.This article explains the platform's development requirements,development environment,infrastructure,and capabilities.This paper designs a deep learning process for high-resolution remote sensing images and ground cover data,including automatic sample generation,model training and block prediction method.Finally,the Mask r-cnn model was used on the designed platform to complete the data training and prediction in nanjing area.(3)adjust the Mask r-cnn model based on the experimental situation in nanjing.In this paper,the Mask r-cnn model is decomposed into several parts,such as the main dry network,RPN layer and optimizer,etc.,and the influence of structural adjustment of each part on the accuracy of the final extraction result is analyzed.After all,the paper determin a Mask r-cnn model that is more suitable for the data used in nanjing region.Finally,several different building extraction methods are compared on the data of nanjing area.