Comparative Analysis of Local and Transferred ANN Models in Landslide Susceptibility Prediction in a Tropical Region

Abstract

Landslides are a common form of natural disaster in the tropics due to heavy rainfall in the wet season. Due to the hazards that come with landslides, determining the susceptibility of an area is of utmost importance. Currently, this is done through an ML-based approach. However, some areas may lack the required data. Thus, this study focused on comparing the impact of a transferred ML model from a comprehensive data region to a localized model. This was done by developing an ANN model trained on data from Western Sarawak and comparing it to the localized model in the west coast of Sabah and Selangor. The transferred ANN model results were acceptable, with recall scores of 0.89 and 0.86 for the west coast of Sabah and Selangor, respectively, while the localized models both achieved a recall score of 1. AUC scores were also comparable, at 0.988 and 0.995 for the west coast of Sabah and Selangor, respectively, while the localized models both achieved an AUC of 1. For the LSMs, in both target areas, the transferred ANN model predictions were heavily skewed in comparison to the localised model. It is recommended that future studies test the transferability in other tropical regions beyond Southeast Asia.