Landslides pose a risks to the communities settled in their vicinity. They pose a threat to human lives and significant economical damage to the affected communities. As the scientific community agrees that the climate change will further increase the risks associated with landslides, it is important to develop a reliable, cost effective and widely applicable technique to assess the stability of the potentially unstable slopes. Such a technique could allow to create an early warning systems meant to help evacuate the communities at risk before the landslide event happens. There have been many studies applying different methods of achieving this goal. This study will apply electrical resistivity tomography to investigate the subsurface and use machine learning methods in the form of supervised and unsupervised learning to interpret the measurements. The electrical resistivity tomography is a widely applied technique to investigate the stability of slopes. It can determine the water table in the subsurface, which plays the crucial role in the stability of the slopes. However, the manual interpretation of the electrical resistivity tomography profiles is a time-consuming and cumbersome task. In order to develop a reliable early warning system both the measurements and their interpretation have to be automatized. The first one has been already achieved. There are systems available that can perform electrical resistivity tomography automatically and continuously. In order to automatize the second part, the following research investigated the capability of machine learning to interpret resistivity profiles alongside with various methods of improving the accuracy of the results. The results show that machine learning is capable of performing this task. Some limitations were identified and guidelines for the preprocessing of the data were proposed. Several areas that require further investigation were identified.

The active layer thickness has become an important indicator in climate change research as permafrost degradation has long been documented. The thawing of permafrost causes the release of greenhouse gases accelerating Arctic warming. Monitoring and quantifying spatial and temporal changes of the active layer are challenging but crucial for reliable climate projections. Geophysical methods offer a non-invasive investigation of electrical properties and their distribution in permafrost areas, revealing phase transitions from water to ice. Subsurface electrical resistivity images can be obtained through inversion of electromagnetic data, yet are inherently ambiguous because of the ill-posed nature of the inverse problem. Since regularization methods offer the possibility to stabilize the inversion, lateral and spatial constraints are incorporated in the inversion algorithm to produce quasi-2D and quasi-3D subsurface models. The developed methodology is evaluated based on synthetic data sets to determine suitable inversion parameters, which are subsequently applied to a field example from the Seward Peninsula, Alaska. Laterally constrained inversion methods based on a few-layer starting model succeed in resolving sharp interfaces in quasi-layered environments. In more complex settings minimum-structure models can retrieve accurate subsurface representations leveraging on vertical and horizontal smoothness constraints. Enforcing lateral and spatial consistency between neighboring soundings thereby yields a similar degree of model smoothness. The inverted field data confirms the conclusions drawn from the synthetic study, as meaningful three-layered models with regard to electrical resistivities are recovered, indicating resistive snow overlying the conductive active layer and highly resistive permafrost. However, the inversion results imply that the snow layer has a significant effect on the predicted model. The implemented constraints help in reducing the ambiguity of the models, but uncertainties introduced by limited data availability cannot be overcome. The potential of adopting spatial and lateral constraints to the inversion is shown, although it becomes evident that additional a priori information needs to be integrated in the objective function in order to comprehensively image the active layer.