Bayesian-optimized lithology identification via visible and near-infrared spectral data analysis

Bayesian-optimized lithology identification has important basic geological research significance and engineering application value, and this paper proposes a Bayesian-optimized lithology identification method based on machine learning of rock visible and near-infrared spectral data. First, the rock spectral data are preprocessed using Savitzky-Golay (SG) smoothing to remove the noise of the spectral data; then, the preprocessed rock spectral data are downscaled using Principal Component Analysis (PCA) to reduce the redundancy of the data, optimize the effective discriminative information, and obtain the rock spectral features; finally, a Bayesian-optimized lithology identification model is established based on rock spectral features, optimize the model hyperparameters using Bayesian optimization (BO) algorithm to avoid the combination of hyperparameters falling into the local optimal solution, and output the predicted type of rock, so as to realize the Bayesian-optimized lithology identification. In addition, this paper conducts comparative analysis on models based on ANN/Random Forest (RF), dimensionality reduction/full band, and optimization algorithms. It uses the confusion matrix, accuracy, Precision ( P ), Recall ( R ) and F 1 values ( F 1 ) as the evaluation indexes of model accuracy. The results indicate that the lithology identification model optimized by the BO – ANN after dimensionality reduction achieves an accuracy of up to 99.80%, P up to 99.79% and F 1 up to 99.79%. Compared with the BO-RF model, it has higher identification accuracy and better stability for each type of rock identification. The experiments and reliability analysis show that the Bayesian-optimized lithology identification method proposed in this paper has good robustness and generalization performance, which is of great significance for realizing fast, accurate and Bayesian-optimized lithology identification in tunnel site. • Bayesian-optimized lithology identification is proposed. • Non-destructive and fast identification of lithology is achieved through hyperspectral technology. • The identification model is developed using ANN and BO algorithm.