| Hyperspectral remote sensing promotes the rapid development of earth observation technology towards a more complete,clearer and accurate view,and nano spectral resolution makes fine spectral interpretation gradually provide effective technical support for national major needs such as natural resource monitoring,ecological environment monitoring and emergency response.However,hyperspectral remote sensing image processing still faces its own problems,such as noise pollution,band information redundancy,lack of sample data and so on.Thus,these lead to the tasks of denoising restoration and de redundant band selection in the hyperspectral remote sensing image preprocessing process,as well as unsupervised classification(i.e.clustering)and supervised classification in the information extraction process.In addition,task solving can be summarized into two steps of “problem modeling optimization solution”.In order to facilitate the optimization solution,a variety of prior information or evaluation functions are often introduced into problem modeling,and the weight parameters are combined into a single objective function for optimization solution.However,in problem modeling,there are usually contradictions between multiple prior terms or evaluation functions,and the information expressed is sensitive,such as the amount of information and redundancy in band selection,the distance between classes and within classes in clustering.In addition,it is difficult to determine the sensitive weight parameters adaptively.In the optimization stage,the gradient descent method based on convex optimization faces the dilemma of falling into local optimal solution in the complex space of hyperspectral image interpretation and modeling.The multi-objective evolutionary optimization method has strong non-convex global optimization ability and accurate multi-element information trade-off ability,which has great potential to solve the above problems.Therefore,this paper focuses on the multi-objective evolutionary optimization method in hyperspectral remote sensing image processing.The main research contents are as follows:(1)The research status and existing problems of unsupervised denoising,band selection,clustering and deep supervised classification of hyperspectral remote sensing images were systematically summarized.Aiming at the common problems,the multi-objective evolutionary optimization theory and its application potential in hyperspectral remote sensing image processing were deeply analyzed;(2)A multi-objective evolutionary optimization based low-rank and sparse nonconvex optimization denoising method for hyperspectral images was proposed(AMOLRS).The hyperspectral image denoising problem was modeled as a multi-objective optimization problem of sparse noise term,low rank clean image term and data fidelity term combined with spatial spectrum total variation.The multi-objective evolutionary algorithm was used for progressive trade-off optimization,the sub-fitness ranking and updating strategies of solving variables with high-dimensional low rank terms and sparse terms were designed to obtain the best visual effect of noise removal and quantitative results of quality evaluation;(3)A hyperspectral image band selection method based on multi-objective discrete sine cosine optimization was proposed(MOSCA＿BS).The multi-objective functions were constructed to weigh the amount of information and redundancy of hyperspectral image band subsets,which are variance function and the ratio function of Jeffries–matusita distance and mutual information respectively,and expect to maximize simultaneously.A discrete sine cosine optimization algorithm that balances the global and local search capabilities was designed,which improves the discrete combination optimization capability of hyperspectral image feature selection,and conform to the multi-objective optimization theory to maximize the amount of information of the selected band subset while minimizing redundancy;(4)The multi-objective evolutionary optimization spatial-spectral clustering approaches for hyperspectral remote sensing image were proposed.Firstly,cluster the hypersectral images after dimensionality reduction,automatically determine the optimal number of clusters by using the evolutionary algorithm to jointly optimize the inter class distance(XB)and intra class distance(Jm),establish the intra class distance joint spatial information objective function(Jm＿S)and XB function for multi-objective global local evolutionary optimization,and a multi-objective spatial-spectral automatic clustering method(FAS2FC＿AMOMA)was proposed.In addition,in order to effectively adjust the global and local optimization in the multi-dimensional and multi peak target space,a multi-objective sine cosine algorithm based spatial-spectral clustering method(MOSCA＿SSC)was proposed.Secondly,for the direct clustering of non-dimensionally reduced hyperspectral image,a multi-objective sparse subspace clustering method(MOSSC)for hyperspectral images was proposed,the non-convex sparse coefficient matrix,data fidelity term and total variation term were modeled separately and were optimized simultaneously by multi-objective evolutionary optimization algorithm,and the sparse coefficient matrix performs spectral clustering and outputs the final clustering results.It effectively avoids the selection of sensitive weight parameters and improves the intelligent optimization and fusion degree of sensitive information of objective function;(5)A multi-objective evolutionary optimization based neural architecture search method for hyperspectral image deep learning classification was proposed(MONAS＿HSI).The lightweight search space of hyperspectral classification network architecture was constructed,the network structure coding and searching were realized by evolutionary algorithm,the flexible hierarchical extraction of image information was realized.In addition,the objective functions that weigh the interpretation accuracy and parameter quantity of classification network was modeled,and the network structure was gradually evolved by multi-objective evolutionary optimization method,which improves the accuracy of hyperspectral classification and the lightweight degree of network architecture;(6)A hyperspectral image processing system based on multi-objective evolutionary optimization was constructed.The performance and applicability of the proposed methods and comparison methods in a UAV hyperspectral remote sensing image of tailings pond were analyzed.In addition,for disaster emergency applications,the emergency response system framework of “UAV emergency multi-objective path planning,hyperspectral remote sensing image multi-objective evolutionary optimization fine processing and emergency change information extraction” was presented,which further promotes the practical application and social value of hyperspectral remote sensing images... |
