Objects Extraction Based On Graph Cut And Deep Learning For High Spatial Resolution Remote Sensed Imagery

With the rapid improvement of spatial resolution,remote sensing imagery nowadays can clearly express the geometric structure and rich spatial information of the land covers,making it possible to extract more detailed elements on land.However,the land cover information is too detailed that the spectral of different land covers have somowhat overlapped,while the spectral distribution within a typical class is variable.So the relative separability between classes is reduced,thus bringing difficulities and challenges to the extraction of land cover information.One of the most popular method to solve such problems is to use object-oriented ways.This method can effectively incorporate spatial information,which is quite useful for high spatial resolution remote sensing imagery scene with complex land cover distribution.But there are still some problems of these methods:(1)The traditional object-oriented methods are based on the local homogeneity to produce no results which will consider global optimization results;(2)For the specific elements of the region,the traditional methods often need to design targeted features and extraction process,the lack of simple and effective modeling methods.When some typical kinds of land covers need to be exracted,the process is usually too complex to be design,which needs human interference.In order to solver those problems,this paper has carried out the research on the feature extraction method for high spatial resolution remote sensing imagery based on graph cut model and deep learning theory.The main idea is to make use of the modeling ability and the global optimal solution of nonconvex data sets to solve the common problem that traditial methods usually face,that is,the dependency on the similarity measurement of local adjacent pixels may produce oversegmentations.Deep learning,which is a theory based on the unique structure learning ability,can model the process of features exraction.This tool can help to avoid the difficulity in designing the complex procedure and the manual intervention.The research content and innovation of this thesis are as follows:(1)The basic theory and methods for feature extraction of high spatial resolution imagery.Based on the characteristics of high-resolution remote sensing images,this paper expounds the difficulties in the methods of object-oriented methods and the extraction of typical classes.And then summarizes the current research background and present situation of the topic.(2)In the aspect of object-oriented ways,a multi-scale and multi-feature-based normalization cut is proposed for high-resolution remote sensing imagery feature extraction.As to the problem that when the graph model is applied to the high-resolution remote sensing image segmentation process,the scale of the graph is too large,the scale is analysed and the similarity measurement is limited.This paper proposes a feature extraction framework based on the multi-scale and multi-feature normalization cut algorithm,which is used to enhance the expression of high-resolution remote sensing imagery.(3)In the aspect of building material and road extraction,high-resolution remote sensing image feature extraction based on multi-channel spatial offset convolution neural network is designed.In this paper,we consider the idea of block prediction,and one model to extract the combination of different categories,to solve problems such as lacking of local space context information and learning structure redundancy.At the same time,spatial offset is added for data enhancement.