Applied machine learning for rockfall source area prediction and a meteorological trigger analysis in Vestland County

Abstract

One of the main challenges in the rockfall hazard assessment is to determine where and when rockfall occurs. The existing national susceptibility map with potential rockfall source areas in Norway based on slope angle thresholds and whether the terrain is on "bare rock" do not consider any other factors influencing rockfall probability. This study investigates the potentials and limitations of applying four machine learning (ML) algorithms for the spatial prediction of rockfall source areas. The "ensemble" Random Forest (RF), the "ensemble" Gradient Boosted Regression Tree (GBRT), the Multilayer Perceptron neural network (MLP), and the Logistic Regression (LR) model were introduced for this purpose. Machine learning models, developed using different combinations of input features, were trained and cross-validated with data from the municipalities of Lærdal and Aurland. The final models were tested, without being recalibrated, for two other areas in Vestland County to investigate the models' regional performance. All four machine learning algorithms were capable of predicting rockfall source areas, with the GBRT model yielding the most promising results. However, the modeled rockfall source areas need further evaluation due to uncertainties related to spatial dependency in the data. This limitation is a fundamental challenge with machine learning in Geosciences. Maps displaying rockfall source areas categorized into four different susceptibility classes were created from predicted probabilities compared to the existing national map and other maps obtained from statistical methods. Two data sets were collected in order to investigate how temperature and precipitation affect rockfall release probability. One represents the "normal climate" in Vestland and the other the climate when rockfall occurs. Differences between the data sets were described using statistical methods. Freeze-thaw was the most significant weather type to trigger rockfalls in the current climate in Vestland County, with most rockfall events occurring on a day of thawing. The local and seasonal effects were explored by fitting individual Logistic Regression models for 16 locations and each month. Results show that the effect of temperature and precipitation varies with both location and month, leading to the recommendation of adopting Hierarchical Bayesian models for temporal prediction in the future.