Research On Mahalanobis Distance Matric Learning Based Hyperspectral Images Classification

Hyperspectral remote sensing images,with their nano-scale spectral resolution and spectral information of up to several hundred bands,reveal many unobservable features of the features of the traditional full-color detection,and have a stronger ability to distinguish features,and have been widely used in the field of remote sensing.However,the rich information contained in hyperspectral images also leads to the typical highdimensional nonlinear characteristics of its data,making it more difficult to accurately classify it,which makes the classification technology based on hyperspectral image become the research hotspots of the field of hyperspectral remote sensing.Although the widely used kernel learning and distance measurement learning methods have obvious advantages for solving the above problems and have been widely recognized by the academic community,they still have shortcomings in terms of computational efficiency and robustness.Therefore,based on the theory of kernel learning and distance measurement learning,this paper studies the existing efficiency and generalization ability of the algorithm,and further improves the classification accuracy of hyperspectral data on the existing basis.Most commonly used kernel learning classification algorithms use Gaussian kernel functions and Euclidean distance as a measure of similarity,but the defects of Euclidean distance susceptible to dimensionality lead to their limited ability to distinguish typical high-dimensional nonlinear data such as hyperspectral images.Although the abovementioned defects can be overcome by using Mahalanobis distance instead of Euclidean distance,most of the current kernel learning methods based on Mahalanobis distance learning are based on Gaussian kernels with limited ability to deal with hyperspectral classification problems.The main bottleneck of classification performance is that the parameters of the appropriate Gaussian kernel function cannot be selected accurately.So the performance of this algorithm is obviously insufficient.Aiming at the above problems,this paper proposes a Mahalanobis distance kernel learning algorithm based on polynomial kernel function transformation.In order to avoid the problem of classification accuracy caused by the difficulty of selecting kernel parameters caused by Gaussian kernel,this paper adopts a polynomial kernel function which is more efficient and more suitable for classification of hyperspectral image as the basic kernel function of kernel classifier.In addition,in order to improve the classification accuracy and the computational efficiency,a high-efficiency low-rank Mahalanobis distance learning algorithm is used to combine the distance measurement matrix to form a new type of Mahalanobis distance metric kernel learning algorithm.The experimental results show that compared with the existing algorithm of the same level algorithm,the algorithm has better classification accuracy and higher computational efficiency,which reflects its value in practical applications.Compared with other enhanced kernel functions constructed using Euclidean distance as the distance metric,although the Mahalanobis distance kernel function can rely on the principle of widening the distance between heterogeneous samples,the discrimination between samples is improved from a more essential level,but currently kernel algorithm based on Mahalanobis distance metric learning can only learn all the prior knowledge without indiscriminate learning,and can not carry out targeted learning on the category samples in the specific task whose distribution distance in the feature space is too close.This kind of learning strategy causes the features of the category groups whose distances between the classes are too close to be neglected,and the generated Mahalanobis distance matrix does not improve the discriminant degree of the individual easily misclassified features,resulting in an unbalanced overall classification effect.Aiming at the above problems,based on the current Mahalanobis distance kernel learning,this paper proposes a dual-layer Mahalanobis distance kernel learning algorithm for hyperspectral image classification.The algorithm can be monitored in the traditional Mahalanobis distance learning process according to the specific hyperspectral image dataset.When it is detected that some categories of samples are more likely to cause misclassification due to too close distance,the data will be targeted for these categories with the secondary learning,and then fuse the results of the two learnings to add more targeted feature information to the specific hyperspectral dataset based on the undifferentiated learning,improve the performance of the kernel function,and achieve a higher overall improvement.The purpose of spectral image classification accuracy.Introducing the spatial information of hyperspectral image into the classification process is beneficial to improve the classification accuracy.However,most of the spatialspectral joint learning frameworks used in the current academic circles cannot fully exploit and effectively combine the spatial spectrum information.The joint learning framework including vector superposition of spatial spectrum information can not effectively suppress noise,and the excessive fineness caused by over-reliance on the performance of unsupervised image segmentation algorithm,and even excessive corrosion caused by ignoring the real feature.Aiming at this problem,this paper proposes a new dual-kernel learning space-spectrum combined hyperspectral image classification algorithm based on Mahalanobis distance learning.This paper adopts the strategy of “pre-classification and post-segmentation” which is the most fully utilized in the space-spectrum joint learning framework,and introduces the kernel learning method in both the classification and segmentation processes,and finds the intermediate feature quantity that combines the distance measure learning algorithm,the kernel learning based classification algorithm and the kernel learning based image segmentation algorithm,so that the advantages of the kernel learning algorithm can be fully utilized in the two processes of classification and segmentation.In this paper,the proposed algorithm is tested in multiple real data sets.The results show that compared with similar framework algorithms and other aerial spectrum joint learning framework algorithms,the proposed algorithm not only improves the image classification accuracy,but also can effectively guarantee the sharpness of the boundary of the object,and have more excellent robustness.