Rock lithology classification and parameter sensitivity analysis based on wavelet scattering transform and support vector machine

Abstract In geological exploration and tunnel/underground engineering, precise, rapid, and intelligent rock lithology identification is crucial. A WST-SVM rock image classification method is proposed that combines wavelet scattering transform (WST) with support vector machine (SVM) to address the limitations of conventional convolutional neural networks reliant on annotated samples. The method extracts multi-scale features from rock images using WST and trains an SVM classifier, achieving superior performance in test accuracy, macro-average precision, recall, and F1-score on a dataset of six rock types. Parameter analysis reveals that increasing invariant scale, decomposition transformations, and quality factor enhances feature matrix dimensionality and computational time. This approach reduces the need for extensive annotated samples and provides a practical solution for improving the accuracy and efficiency of rock lithology identification in geological exploration and tunnel engineering.