Integration of Physical and Statistical Knowledge in Landslide Susceptibility Characterization

Landslide susceptibility mapping (LSM) can provide valuable information for landslide characterization and mitigation. However, there are long-standing problems with LSM that have remained unaddressed or inadequately resolved, such as a lack of an effective framework to integrate the physical and statistical knowledge about landslide occurrence. To address the knowledge gap, this paper proposed to combine the physical model and machine learning techniques to produce a robust landslide hazard assessment. The hypothesis is that the physical modeling result itself can be a valuable model input to be incorporated in the training model of a machine learning framework. In this research, the tree augmented naïve (TAN) Bayes classifier has been selected as the machine learning model. In addition, we used PISA-m, a probabilistic landslide slope stability program, to estimate the probability of landslide occurrence in Magoffin County, Kentucky, United States. Results of physically assessed landslide susceptibility were used to be combined with other geomorphological variables in the TAN model to predict the landslide probability in the study area. We performed cross-validation to evaluate the model performance and reliability of the proposed methodology.