Research On Removing Redundancy Method For Hyperspectral Remote Sensing

Hyperspectral remote sensing technology combines the imaging technology and the spectroscopy technology organically,and the data acquired covers both spatial and spectral information of features.Therefore,hyperspectral remote sensing data can be regarded as a three-dimensional cube composed of spectral dimensions and spatial dimensions.On the one hand,the fine division of the spectral dimension can obtain more bands,but it also leads to a strong correlation between the bands,resulting in redundancy of the spectral dimensions.The existence of a large number of redundant bands in the spectral dimensions is not conducive to the storage and transportation of data,and it is easy to cause the curse of dimensionality,resulting in a decrease in the identification rate of classification of features.On the other hand,when there are a large number of feature training samples in space dimensions,although these training samples can bring more possibilities for predictive analysis of unknown samples,they also lead to more complex and inefficient training models.The above problems can be solved by data redundancy elimination,in which the data redundancy elimination of the spectral dimensions can be realized by band selection,and the data redundancy elimination of the spatial dimensions can be realized by sample selection.Band selection is to select a subset of bands that play a major role from all bands of a hyperspectral image,which can not only remove the redundancy of the spectral dimensions,but also retain more complete and useful information.The sample selection method is different from the band selection method,which is to select a subset of samples with high contribution to the model training from all the training samples of the hyperspectral data,which can effectively remove the redundant information without affecting the generalization ability of the model.Therefore,based on the study of the previous research results.the dissertation has made further research on the band and sample selection of the hyperspectral data.The main research contents are as followsFirst of all,in order to improve the band search efficiency of suboptimal search algorithm,two kinds of band selection methods in subspace are proposed:1)Fast band search method based on subspace partition.This method is carried out in the band subspace,and performed by orderly selecting the band with the largest variance in each subspace as the initial band,setting the objective function,and then replacing the bands gradually in the subspaces to optimize the objective performance until there is no band to make the object better.It only needs to input the size of the subset of the selected band in this method without settings of other parameters,and can converge faster while ensuring the quality of the searched object compared with the classical band selection method;2)the band selection method based on the artificial bee colony algorithm.The method takes into account both the retention of important information and the removal of redundant information,and takes the weighted sum of the JM(Jeffreys-Matusita)distance and the optimal index factor as the objective function.At the same time,in order to solve the combination optimization problem quickly,the band subspace is used as the search range of the band combination,and the Artificial Bee Colony algorithm(ABC)which has almost no requirement for the objective function is introduced as the search algorithm.The proposed algorithm is experimentally compared with the typical intelligent band search method,and the results show that the proposed method is a band selection method with better search efficiency and solution efficiency.After that,in order to obtain both the information content and the correlation band combination,two methods based on the inter-band distance calculation are proposed:1)Band selection method based on distance weighting.The method measures the information content in the band by standard deviation,and iteratively calculates product of the inter-band JM distance and the standard deviation of the selected band,selects the current band with the largest product to enter the band subset,and removes the bands that are close to the selected band in each iteration" and obtains a subset of bands supported by separability and information content;2)Band selection method based on max-min distance.Firstly,according to the characteristic that the distribution of hyperspectral data in the band space is of convex geometric structure,the Euclidean distance is used as the distance standard between the bands to calculate the max-min distance from all brands to the selected band,and select the bands that are as far apart as possible in the high-dimensional space.Then,according to the clustering characteristics of the adjacent bands of the hyperspectral data,the subset of the bands selected by the max-min distance method is taken as the initial cluster center,and all bands are clustered by the K-medoid algorithm,and the Euclidean distance is also used as the calculation standard to classify by close principle until convergence,and the finally obtained cluster center is the final output band subset.Comparing the proposed two algorithms with the typical band selection method,the results show that the bands selected by the two methods in this chapter are more ideal in terms of classification accuracy,and the generated band combinations are more stable,which can better meet the actual needs.Finally,in order to reduce the complexity of the training model and improve the classification efficiency,aiming at the defect that the least square support vector machine(LS-SVM)participates in model training with all the training samples,a sample reduction strategy based on Coulomb Force is proposed,which introduces the gravity model between charged particles in physics into the calculation of the positional relationship of the training sample space,and judges if the sample is retained by calculating the contribution value of the sample point to the classification hyperplane.Compared with the case without sample selection,the proposed method can effectively improve the efficiency of the classifier under the condition that the classification accuracy is basically the same.The simulation experiments show that the sample reduction strategy can greatly reduce the classification time.