| dc.description.abstract | Depth to Bedrock (DTB) is a critical parameter in several fields of study, including geology, hydrology, soil sciences, and civil engineering. However, obtaining this parameter through near-surface geophysical methods can be challenging and expensive, particularly in difficult terrain. Fortunately, high-quality borehole data from previous geotechnical investigations can be used to estimate the DTB in areas where no boreholes have yet been created.
This thesis presents a machine learning framework for estimating the DTB value in areas of interest using Gaussian Process models. The performance of different kernel functions, including Radial Basis Function (RBF), Matérn 3/2 kernels, and combined linear and RBF kernels, is evaluated, along with the impact of implementing anisotropy in the models.
The results show that the Matérn 3/2 kernel with anisotropic implementation performs the best in estimating DTB. However, challenges in hyperparameter optimization, non-Gaussian target variables, and model selection are highlighted, and further investigation into these areas is recommended. The framework presented here provides practical implications for geotechnical engineering. Further, it provides a basis for future research in this area, where the incorporation of additional geological and remotely sensed data could potentially improve the quality of DTB estimation. | |
