Research And Application Of Hyperspectral Remote Sensing Image Classification Algorithm Based On Ensemble Learning

Hyperspectral classification refers to a remote sensing technology that makes full use of the dense spectral information reflected from distant objects to discriminate and classify the objects,and has broad application prospects.Hyperspectral images have the characteristics of high spectral dimensions and few sample data,and are prone to produce similar phenomena such as spectral variation and homologous foreign objects,which bring great challenges to classification problems.The traditional classification model is based on the assumption that "similar spectra have the same category".Generally,no single classifier is omnipotent,and the same hyperspectral data will produce different classification results under different classifier models.For the learning method of neural network,first of all,the problem of shortage of hyperspectral data needs to be solved.Secondly,the more challenging problem is to design the network structure and make a reasonable interpretation of the classification results.For existing ensemble learning classification methods,multiple classifier models are often used alone to mechanically combine,or the base classifier is iteratively iterated through the Boost method.In order to solve this problem,the research in this paper proposes multi-stage ensemble learning by combining the integration ideas of cascade and parallel,thereby improving the overall classification effect of the model.First,from the perspective of the quantitative index measurement of classifier difference and complementarity,three machine learning classifiers with the highest matching index are screened for a specific data set.Then,the selected three classifiers are used as three-level base classifiers.The first stage of multi-stage ensemble learning is to obtain three second-level classifiers through multi-kernel learning,integration under different parameters or Boost linear combination based on the characteristics of each three-level base classifier.The second stage of multi-stage ensemble learning is to obtain the weight matrix through the feedback mechanism of the overall accuracy of the two-level classifiers,and fuse the first-level classifier.Selecting small training samples from three open source hyperspectral data sets for experiments,the results show that: the overall classification accuracy,average classification accuracy,and Kappa coefficient of the first-level classifier have been significantly improved.At the same time,the training sample set gradually increased from 5% to 80%,and the classification performance of the multi-stage ensemble classifier in this paper increased steadily with the increase of training samples.Finally,the above-mentioned multi-stage ensemble learning classification method is applied to the classification of actual crops in the Xiong'an New District with more complicated categories.The experimental results show that the above three classifier evaluation indicators have been greatly improved.Comparing the multi-stage ensemble method in this paper with the classification results obtained by convolutional neural network classification,the method in this paper has a great advantage in terms of time cost performance.