Lithological classification from the ASTER data based on wavelet transform， SVM， and voting methods： A case study for the Weiya area in the eastern Tian Shan

In ASTER images, different lithologic units show obvious multiscale texture features, and wavelet transform has the advantage of extracting multiscale features. Support Vector Machine (SVM) is suitable for solving the classification problem of little training data and nonnormal data distribution. SVM is used to complete lithology classification. The classification results have high classification accuracy and low uncertainty. Using the voting method in selecting lithologic classification results can avoid the uncertainty of lithologic classification results caused by the extraction method of lithologic samples, thereby making the classification results statistically significant. An automatic classification method for ASTER image lithology integrating the wavelet texture, SVM, and voting method is proposed to improve the accuracy of ASTER imagery exploited for mapping assistance. First, the Haar wavelet is utilized for decomposing the ASTER image involving a multiscale wavelet, with the mean value of wavelet coefficients considered texture features. Moreover, the variance, homogeneity, and mean values of the gray-level co-occurrence matrix (GLCM) are extracted concurrently. Then, the feature vectors of the SVM classification are constructed with multiscale texture, GLCM texture, and spectral features. The classification is repeated 10 times. Finally, the lithologic unit is determined by the voting method, and the results are statistically evaluated. The lithologic classification involves 92.1934% accuracy, exceeding the accuracy of spectral classification by 13.3369%, with a kappa coefficient of 0.9202. The multiscale texture extracts detailed lithologic information. The voting method prevents the dynamic lithologic change caused by the spatial variability of samples. The SVM also demonstrates superiority over the maximum likelihood classifier for lithologic classification involving high-dimensional and nonnormal distribution data. The local optimal parameters of SVM are avoided using the artificial bee colony algorithm to search for optimal parameters. © 2023 Science Press. All rights reserved.