Research On Hyperspectral Image Classification Technique Based On Gravitational Search Algorithm And Ensemble Learning

With the development of remote sensing technology,hyperspectral remote sensing has been widely used in many fields as a novel remote sensing data acquisition pattern.Nowadays,hyperspectral remote sensing has been a hot spot of research in application area of remote sensing.However,the hyperspectral image(HSI)includes a great deal of band information;the storage space is relatively large for a single image,and needs a lot of time to make image processing,which brings about a new challenge for HSI interpretation,and how to devise the classifier model with good robustness to improve the classification accuracy and efficiency is one of the pop research project in this field.In recent years,as a large amount of evolutionary computing algorithms have been proposed,which provide a new idea to solve the combinatorial optimization problem.Some studies have shown that the performance of remote sensing image processing is satisfactory by using evolutionary computing algorithms.However,it is easy to trap into the local optimal,and hard to converge to the global optimal solution.So,the research work is carried out by using a newly proposed improved gravitational search algorithm in the paper,which makes a balance between the process of global and local search,and a stable and robust HSI classifier model is designed to simultaneously enhance the classification accuracy and efficiency.The main research contents are as follows:1.The fundamental principle of gravitational search algorithm(GSA)is briefly introduced,and the algorithm is improved by combining the chaotic mapping and Levy flights to make randomly perturbation for the superior particle.It is revealed that the improved gravitational search algorithm(IGSA)has better optimization ability comparing with other common used evolutionary computing algorithms based on the experiments with some public test functions according to the fitness value.2.A band selection method based on IGSA is proposed,and the optimal band combination is obtained by the search process of IGSA.The high spectral resolution of HSI provides a lot of spectral information,and brings some challenges at the same time.In order to satisfy the real time requirement of HSI classification,the data dimensionality must be reduced.Band selection is to select part of bands and removes the bands with low correlation for the HSI data,which most responses the basic information of original data.Then,the band feature subset is trained by the same classify model,and choose band combination with large differences from it to achieve the purpose of data reduction.3.An integration optimization method based on IGSA is proposed to make band selection and parameter optimization simultaneously.As a SVM mode,the kernel function parameter ? and the penalty factor C have great effect on classification results.The optimal parameter could make a preferable balance between the classification accuracy and the generalization ability of SVM.In fact,both parameter optimization and band selection could be regarded as the combinatorial optimization problem,and there is a certain connection between each other.So,the integration optimization method is considered to obtain the optimal band combination and classifier parameter by using IGSA at the same time.Moreover,because of the high dimensionality of HSI datasets,the wavelet function is utilized as the kernel function of SVM to enhance the robustness of the classifier model and the adaptability of the proposed classification approach.4.A multiple classifiers ensemble mode is built based on IGSA.In the process of HSI classification,it is difficult to make distinction for the object with similar digital number value by using single classifier model.As the cheapness of computation and storage,the multiple classifiers ensemble model by combined with a few of learning strategies has been widely used to solve the classification problem.Multiple classifiers ensemble is to fuse several of classifiers as a model by using the theory of ensemble learning,and the base classifier models are trained by the same assignment.The optimal band feature subset and classifier combination are selected by IGSA,and the regular band information and excellent performance classifiers are utilized to make adaptive classification for the images on large areas,synthesize the advantage of different classifier models,and the ensemble model keeps favorable generalization ability for each categories.In all,a HSI classification approach based on IGSA is presented in the paper,the dimensionality of HSI datasets has been effectively reduced,and the optimal classifier parameter is obtained by the operation of band selection and classifier parameter integration optimization for the original images;the robustness and generalization ability is enhanced by the multiple wavelet SVM classifier model.Experimental results demonstrate that IGSA has a good performance to solve the HSI classification problem,and the built classifier model fatherly improves the classification accuracy and efficiency,and has become a broad application prospect in this field.