Landslide Geometry Reveals Its Trigger

Figure 1: The map of Japan shows the geographical locations of the seven landslide inventories used in this work.

Existing Landslide inventories rarely include the triggering mechanisms that make them unusable for landslide hazard modeling. Therefore, we developed a method for classifying the triggering mechanisms of landslides based on the geometric properties of the landslide’s polygon. First, employing a combination of feature selection methods, we choose seven geometric properties of landslide polygon that best classify the landslide into two categories: earthquakes and rainfall.

We used these geometric properties of landslide polygon as a feature space for the machine-learning-based classifier random forest. Random forest is a decision tree-based ensemble learning method that is highly robust to classification and regression problems. In this work, we use six landslide inventories spread over the Japanese archipelago, all having triggering information.

Using the known trigger of inventories, we used various combination of training and testing sets to achieve around 85 % classification accuracy of their triggers. Moreover, we trained the algorithm on five inventories in another approach and tested it on sixth inventory. Using this approach, we achieved mean classification accuracy ranging from 67% to 92%. Finally, we tested our approach with an additional 7th landslide inventory that has no triggering information as a real-world application scenario of the developed method.

Apart from classification, we analyzed the probability distributions of geometric attributes of the earthquake and rainfall polygons in-depth. We found dissimilarities between the probability distributions of geometric attributes of earthquake and rainfall triggered landslide polygons. The probability distributions of geometric attributes also support classification results for the 7th inventory that has no triggering information. The classification results show that the developed method is robust and provides high classification accuracy. Furthermore, the feature importance analysis indicates that landslides having identical trigger mechanisms exhibit similar geometric properties. A long-suspected fact that we anticipate landslide modelers will find helpful.


[Rana et al.(2021)Rana, Ozturk, and Malik] Kamal Rana, Ugur Ozturk, and Nishant Malik. Landslide geometry reveals its trigger. Geophysical Research

Letters, 48(4):e2020GL090848, 2021. doi: 10.1029/2020GL090848”. URL "https://doi.org/10.1029/2020GL090848"