Investigating the effects of landslides inventory completeness on susceptibility mapping and frequency-area distributions: Case of Taounate province, Northern Morocco

In explanatory statistics, training data constitute the most significant input since they determine the reliability and accuracy of the output predictive models. In Landslide Susceptibility Mapping (LSM), such data is derived from landslides inventory maps (LIM) that feed the algorithms information regarding the location of past landslide occurrences. However, the impact of LIM size and completeness is rarely investigated. Therefore, we attempt to quantify in this research the amount by which LIM size variability influences the results of LSM modelling in two areas with contrasted geomorphological settings. The latter zones are located in the geo-morphologically heterogenous Intra and Mesorif domains (zone A) and the homogenous Prerif region (zone B). To prepare training data, we first extracted 33%, and 66% from the original dataset, which contained an excess of 10 000 landslides. Then, we investigated the effect of information loss on frequency area distribution analysis (FAD), which is a useful tool for quantifying the completeness of a given LIM. Then 18 LSMs were generated using Frequency Ratio (FR), Logistic Regression (LR) and Artificial Neural Networks (ANN) models. Our findings suggest that in heterogenous and complex geological settings such as zone A, the LSM and FAD distributions seem to be significantly impacted, while in a homogenous and monotonous study area (zone B), little to no variance is observed in the output models. However, the variability in susceptibility assessment results is also influenced by the algorithm used in LSM preparation, with LR being the most stable algorithm and Artificial Neural Networks presenting the most sensitive model with the most accurate results when using the full LIM.

Automated summary

This research quantifies the impact of landslide inventory map (LIM) size variability on Landslide Susceptibility Mapping (LSM) results. In heterogeneous areas, LSM and Frequency Area Distribution (FAD) distributions are significantly impacted, while in homogenous areas, little variance is observed. Additionally, the choice of algorithm used in LSM preparation also influences susceptibility assessment results, with Logistic Regression (LR) being the most stable algorithm and Artificial Neural Networks (ANN) presenting the most sensitive model.