Machine learning for natural hazard data analyses and data-driven geotechnical engineering applications

Abstract

Extensive exploration of novel machine learning (ML) technologies within the domain of geoscience has been actively pursued. This pursuit is driven by two primary factors: Firstly, there exists a wealth of concepts within geoscience that lend themselves to mathematical formulation, enabling more robust quantitative analyses. Secondly, the field of geoscience is inherently data-rich, offering fertile ground for the development of novel ML models tailored to geoscience. However, the integration of ML with geoscience is in its early stages and unevenly advancing. This Ph.D. thesis aims to contribute to this interdisciplinary research field by taking a modest step forward. The thesis concentrates on two main research subjects: (1) natural hazard data analyses and (2) data-driven geotechnical engineering applications. These two subjects encompass a total of five specific research topics: (1) classic ML for classification, (2) classic ML for regression, (3) supervised learning by convolutional neural network (CNN), (4) unsupervised learning by CNN, and (5) deep learning (DL) with 3D input data. These five research topics cover a wide range of contents, including (1) boulder fall volume range prediction, (2) seismic source localization, (3) rock type classification, (4) low-light rock image enhancement, and (5) 3D point cloud filtering. They collectively contribute to the integration of ML in geoscience. Based on the abovementioned research objectives, a series of experimental tests have been designed and carried out with the collection of massive geodata and the development of issue-specific ML models. The key findings and contributions regarding each research topic are as follows: Regarding classic ML for classification, the volume ranges of potential boulder falls in Hong Kong have been predicted based on given ML features. The suitability of eight classic ML algorithms is adequately demonstrated. Regarding classic ML for regression, the source locations of acoustic emission (AE) events are predicted using the support vector machine (SVM) and compared with those obtained by the conventional method. The negative impacts of anisotropy and water content on the accuracy of source localization are quantitatively investigated. Regarding supervised learning by CNN, a large-scale rock image dataset is established, comprising common rock types found in Hong Kong. An advanced CNN, namely HKUDES_Net, is developed for rock type classification, demonstrating state-of-the-art performance. Regarding unsupervised learning by CNN, a novel unsupervised DL model that caters to enhance low-light rock images is developed. The DL model utilizes the deep curve estimation (DCE) algorithm and a CNN architecture to perform automatic and pixel-wise enhancement. Regarding DL with 3D input data, a novel DL model for 3D point cloud filtering is developed for geotechnical applications. The DL model has demonstrated superior performance in three key aspects: (1) effective noise filtering, (2) point-wise adjustments for noise points, and (3) preservation of edge features. This thesis leverages domain knowledge of geoscience to enhance the development of ML models and utilizes advanced ML technologies to provide insights into traditional geoscience issues. The methodologies presented in this thesis extend beyond the illustrated applications and have broader applicability to various types of natural hazards and geotechnical engineering applications.