Deep learning-based key-block classification framework for discontinuous rock slopes


Zhu H., Azarafza M., AKGÜN H.

Journal of Rock Mechanics and Geotechnical Engineering, vol.14, no.4, pp.1131-1139, 2022 (Peer-Reviewed Journal) identifier identifier

  • Publication Type: Article / Article
  • Volume: 14 Issue: 4
  • Publication Date: 2022
  • Doi Number: 10.1016/j.jrmge.2022.06.007
  • Journal Name: Journal of Rock Mechanics and Geotechnical Engineering
  • Journal Indexes: Science Citation Index Expanded, Scopus
  • Page Numbers: pp.1131-1139
  • Keywords: Block theory, Discontinuous rock slope, Deep learning, Convolutional neural network, Image-based classification, STABILITY ANALYSIS

Abstract

© 2022 Institute of Rock and Soil Mechanics, Chinese Academy of SciencesThe key-blocks are the main reason accounting for structural failure in discontinuous rock slopes, and automated identification of these block types is critical for evaluating the stability conditions. This paper presents a classification framework to categorize rock blocks based on the principles of block theory. The deep convolutional neural network (CNN) procedure was utilized to analyze a total of 1240 high-resolution images from 130 slope masses at the South Pars Special Zone, Assalouyeh, Southwest Iran. Based on Goodman's theory, a recognition system has been implemented to classify three types of rock blocks, namely, key blocks, trapped blocks, and stable blocks. The proposed prediction model has been validated with the loss function, root mean square error (RMSE), and mean square error (MSE). As a justification of the model, the support vector machine (SVM), random forest (RF), Gaussian naïve Bayes (GNB), multilayer perceptron (MLP), Bernoulli naïve Bayes (BNB), and decision tree (DT) classifiers have been used to evaluate the accuracy, precision, recall, F1-score, and confusion matrix. Accuracy and precision of the proposed model are 0.95 and 0.93, respectively, in comparison with SVM (accuracy = 0.85, precision = 0.85), RF (accuracy = 0.71, precision = 0.71), GNB (accuracy = 0.75, precision = 0.65), MLP (accuracy = 0.88, precision = 0.9), BNB (accuracy = 0.75, precision = 0.69), and DT (accuracy = 0.85, precision = 0.76). In addition, the proposed model reduced the loss function to less than 0.3 and the RMSE and MSE to less than 0.2, which demonstrated a low error rate during processing.