Evolutionary Computing Theory And Its Applications In Remote Sensing Image Classification

Remote sensing image classification can be defined to label the pixels of remote sensing images in order to meet the needs of different application fields.Recently,the developed earth observation system generates huge amount of remote sensing images,providing the abundant data for remote sensing image classification.However,in some specific conditions,there are also some problems for remote sensing image classification,such as lack of samples,poor optimization capability,difficulty in considering the prior information,restricting the application of semi-supervised classification,unsupervised classification,subpixel mapping for remote sensing image.Evolutionary algorithm possesses the powerful capability in information representation,model optimization,information fusing,having great potential in resolving the above problems.In this thesis,based on the evolutionary algorithm,the semi-supervised classification,unsupervised classification,subpixel mapping for remote sensing image was studied.The contents are listed below.(1)Syetematically review the research status and summarize the problems in remote sensing image classification(i.e.semi-supervised classification,unsupervised classification,subpixel mapping for remote sensing image).With regard to these problems,analyze the advantages of evolutionary algorithm(i.e.artificial DNA computing,memetic algorithm,multi-objective optimization theory)and specify their great potential in remote sensing image classification;(2)An artificial DNA computing-based semi-supervised classification method for hyperspectral remote sensing image named SSDNA was proposed in order to disdinguish the similar objects when little samples are available for hyperspectral remote sensing.In this thesis,motivated by the similarities between the hyperspectral remote sensing image classification and natural DNA codes matching,the original hyperspectral remote sensing image classification was transformed into the optimization of DNA cube,in which the evolutionary algorithm was used to extract the key DNA codes of objects with the help of labeled and unlabeled samples.(3)Adaptive memetic algorithm-based clustering methods for remote sensing image were proposed.Due to the characteristics of stochastic and self-organizing for the task of clustering remote sensing image,the traditional remote sensing image clustering methods are easy to get stuck into local optima,or difficult in refine the solutions.This thesis introduced the memetic algorithm into the problem of remote sensing image clustering,Memetic algorithm can be thought of as the hybrid of global search methods and local search strategies,thus can improve the performance of remote sensing image clustering.In addition,several adaptive strategies were designed in order to automatic determine the spatial weight parameters and balance the global and local optimization capabilities adaptively.(4)An adaptive multi-objective subpixel mapping frameword for hyperspectral remote sensing image was proposed.In this framework,the problem of hyperspectral subpixel mapping was transformed into problem of multi-objective optimization,in which the reconstruction term and the prior term in the MAP-based objective functions can be optimized simultaneously using MOEA/D;Furthermore,the angle-based seletion method was utilized to select the best subpixel mapping solution from the pareto front;(5)An prototype of evolutionary algorithm-based remote sensing image classification system was built.Systematically analyze the application conditions of the proposed methods for semi-supervised classification,unsupervised classification,subpixel mapping.The thesis builds a set of remote sensing image classification method based on evolutionary algorithm,which can significantly enhance the potential applications of remote sensing image classification,therefore have important theoretical and applied significance.