Feature Selection And Feature Extraction For Hyperspectral Images Based On Evolutionary Optimization And Learning

Recently,with the development of hyperspectral remote sensing technology,it has been applied to environment monitoring,medical images,precise agriculture and urban moni-toring and so on.Meanwhile,the high-dimensional data structure brings new challenges to hyperspectral images(HSI)processing.Generally,much redundancy information exists in adjacent bands.High-dimensional data structure can cause "Hughes" phenomenon and bring more burden on storage space and computation complexity.Moreover,labeled data are very scarce in HSIs and labeling data is very difficult.This dissertation makes a deep analysis on the characteristics of HSI structure and proposes multiple unsupervised feature selection and feature extraction methods which are summarized as follows.(1)For the unsupervised band selection problem,information preserving and redundancy reducing are the two crucial aspects which should be considered.We design an objective function,which takes the two aspects into account simultaneously.The proposed objective function is a combinational optimization problem in nature,which is hard to obtain deriva-tive information because its solution space is discrete.To optimize the proposed objective function,a heuristic stochastic searching strategy is designed which is based on memetic computation.Experimental results indicate that the feature subsets obtained from the pro-posed method can improve the classification accuracy significantly.(2)Standard unsupervised FCM is very sensitive to initialization condition and noisy points,and it is also easy to be trapped in local optimum.To address these problems,a PSO-based optimization method is utilized to replace the lagrangian multiplier method in standard FCM.The combined FCM and PSO method is applied to HSI unsupervised band selection prob-lem.Experimental results show that the proposed method is robust to initialization condi-tions and noisy bands.Moreover,it can avoid to be trapped in local optimum effectively.In classification experiments,the proposed method has superior performance.(3)For unsupervised feature selection problem,how to determine the appropriate number of selected features is a challenging open issue.To address this issue,a multiobjective op-timization model is built to quantify the conflicting relationship which is between the num-ber of the selected bands and the amount of information preserved.Moreover,a heuristic stochastic searching strategy which is based on decomposition based evolutionary multi-objective optimization is designed to optimize the proposed model.The proposed method can handle solutions with different numbers of selected bands simultaneously.Experimen-tal results illustrate that the proposed method can obtain a series of solutions with different numbers of selected bands in a single run,which can offer decision makers more options.Moreover,these obtained band subsets have good and stable classification performance.(4)Some noisy bands exist in HSIs,which have little information and are different with other bands significantly.This facts lead to the phenomenon that information preserving and redundancy reducing are conflicting in nature which can not be optimized simultane-ously.However,different HSI data sets have different preferences on these two aspects.It has become a challenging issue that how to explore the optimal trade-offs between these two aspects according to different HSI data sets.To address this issue,the two aspects are quantified through two objective functions,and these two objective functions constitute a multiobjective optimization model.Under the designed multiobjective optimization model,the optimal trade-offs between these two aspects can be explored.To solve this multiob-jective optimization model effectively,a heuristic stochastic searching strategy is designed which is based on multiobjective artificial immune algorithms.Experimental results show that the proposed method can explore optimal trade-offs between the two aspects effective-ly and obtain a series of solutions which represent different optimal trade-offs.The results of classification experiments show that the band subsets obtained by the proposed method improve the classification performance significantly.(5)In general,the hyperparameters in networks of deep learning depend heavily on a large amount of labeled training samples for training.However,labeled samples are insufficient severely in HSI data sets.They are also very hard to be obtained because obtaining la-bel information is very time and labor consuming.Reducing the requirement of amounts of labeled samples in deep learning without significant performance decreasing becomes a challenging open issue for HSI feature extraction.To address this issue,we propose an unsupervised feature extractor which combines WGAN and CNN.In the proposed method,the ability of CNN to capture spatial and spectral information is used by designing a CNN-based feature extrator,and a WGAN-based framework is designed to train the designed CNN-based feature extractor without any labeled samples.The proposed method gets rid of the dependence of labeled samples thoroughly during feature extractor process,which is completely unsupervised.Classification experiments are implemented on three real data sets.Experimental results show that the proposed method can improve the classification performance significantly compared with some traditional unsupervised feature extraction methods,which validates the effectiveness of the proposed method.