Research On Construction Land Information Extraction From High Resolution Images With Deep Learning Technology

The rapid development of remote sensing technology greatly improves the quality and quantity of image data.High resolution remote sensing image can express rich surface information in small scales,they attract more attention and are being widely used in various industries of national economic construction due to their great application potential and development prospects.Image classification is the basis of remote sensing technology,and the traditional object-oriented image classification method can improve the utilization of image data and classification accuracy to a certain degree,however,different classification methods aim to solve different problems and require a wealth of experience and knowledge in the selection of features,resulting in the rich high-resolution image information not being fully utilized.Construction land information extraction is the foundation work of land use structure and land cover,urban and rural planning,land Change Monitoring,law enforcement inspection of land and investigation of illegal construction land.What's more,quickly and accurately extraction of construction land information is the basic work in these applications.In this context,this paper summarizes the relevant research theory and research results at home and abroad.Taking the part of Tongxiang City in Zhejiang Province as the research area and the multi-spectral remote sensing image data of GF-2 in July 2016 as experiments data,the deep learning technology,which has been successfully applied in the field of image and speech recognition,is used to study the best segmentation of construction land information extraction,and the effect and advantages of deep convolution neural network in automatic extraction of construction land information on the basis of object-oriented multi-scale segmentation.This paper includes the following contents:(1)The comparative analysis method is used to determine the multi-scale segmentation effect of GF-2 remote sensing image by segmentation scale of 30,50,100,150,200.(2)In order to convert the image objects to the format that can be input into deep convolution neural network,it is necessary to standardize the image object.In order to standardize the image object,this article uses three ways to assess the impact of different standardized methods on classification accuracy.At the same time,we explore the combination of object-oriented image segmentation technology and deep learning technology to form a complete information extraction process.(3)Automatic feature extraction and classification of construction land in high resolution remote sensing image are realized by deep convolution neural network,and the classification accuracy of number of objects,object area and land cover accuracy are evaluated based on the land cover data of whole study area,then the recall and precision of the extraction of construction land are calculated to evaluate the effect of the proposed method on the extraction of construction land information.The main achievements are as follows(1)Contrast analysis of the accuracy and effectiveness traditional object-oriented support vector machine,decision tree,nearest neighbor classifier and deep convolution neural network in the construction of land information extraction.The results show that the overall classification accuracy and the recall and precision of construction land of deep convolution neural network are all superior to the other three classification methods,thus it is a method and technology with high application potential.(2)The standardized method of image object has little effect on classification accuracy,it is shown that the depth convolution neural network has better fault tolerance and strong recognition ability for the shape,size and displacement of the image object.(3)The optimal segmentation scale of different objects is different.The internal structure,spectral characteristics and the shapes of the construction land is complicated,this paper finds that the optimal segmentation scale for multi-scale segmentation of GF-2 remote sensing image is between 50 and 100 by analyzing the effect of land use information extraction under different scale.