Study On The Classification And Comparison Of Hyperspectral Data Based On Machine Learning Method

In today's world,the application field of machine learning technology is expanding rapidly,and the technology of data processing and information analysis has undergone tremendous changes.As a result,the classification effect of image data can be significantly improved.In the past,the classification and recognition of hyperspectral or multispectral images only focused on the characteristics of the spectral dimension of the pixels,and the statistical analysis of the features was only carried out on the spectral dimension.Because of the complexity of nature and the existence of mixed pixel problem,it is not sufficient to rely solely on the spectral characteristics of pixels.Adding spatial feature information to hyperspectral remote sensing image supervised classification can effectively improve the speed and accuracy of the classification results.In this study,two methods,extreme learning machine(ELM)and support vector machine(SVM)is combined to evaluate and compare the time and accuracy of the two classification methods before and after adding spatial feature information.Two hyperspectral datasets,the ROSIS data of Pavia university and the Hyperion data of Okavango Delta(Botswana),were selected to test the methods.In order to demonstrate the significance of the machine learning method in geological prospecting,this paper also selects the AVIRIS data,the hyperspectral remote sensing data of Cuprite in Nevada to test the methods.After image preprocessing,training sample selection and spectral feature analysis,two classification methods were used to classify the data set.Then,spectral features and spatial features were combined to classify the data.The major subjects and conclusions of this paper are indicated as below:(1)Experiments were carried out on hyperspectral data sets collected from two different sensors and different types of surface coverage using two classification methods of support vector machine and extreme learning machine.The spectral characteristics of the data were analyzed and compared with the classification results,which confirmed the spectral curves between different types.Higher similarity of features will increase the difficulty of classification of data sets and have a negative effect on classification accuracy.(2)The direct watershed algorithm based on the gradient image is easy to lead to over-segmentation of the image.The main reason for this phenomenon is that there are too many very small watershed basins in the input image,which results in that the segmented image cannot reflect the meaningful areas in the image.Therefore,similar regions of segmentation results should be merged..(3)The majority voting method is used to merge the spectral and spatial classification results,in which many discrete points are designated as neighborhood values,which improve the classification accuracy.Another effect is that the boundary between classes is smoothed,which solves the problem that the difference of classification accuracy often appears at these boundary points.(4)By comparing and analyzing the experimental results with the obfuscation matrix method,it can be seen that combining the classification results based on spectral features with spatial features can effectively improve the classification accuracy of the two algorithms.At the same time,in terms of classification time and accuracy,the utmost learning machine is superior to support vector machine in both research areas.