Comparison of Three Mixed-Effects Models for Mass Movement Susceptibility Mapping Based on Incomplete Inventory in China

Abstract

Generating an unbiased inventory of mass movements is challenging, particularly in a large region such as China. However, due to the enormous threat to human life and property caused by the increasing number of mass movements, it is imperative to develop a reliable nationwide mass movement susceptibility model to identify mass movement-prone regions and formulate appropriate disaster prevention strategies. In recent years, the mixed-effects models have shown their unique advantages in dealing with the biased mass movement inventory, yet there are no relevant studies to compare different mixed-effects models. This research compared three mixed-effects models to explore the most plausible and robust susceptibility mapping model, considering the inherently heterogeneously complete mass movement information. Based on a preliminary data analysis, eight critical factors influencing mass movements were selected as basis predictors: the slope, aspect, profile curvature, plan curvature, road density, river density, soil moisture, and lithology. Two additional factors, namely, the land use and geological environment division, representing the inventory bias were selected as random intercepts. Subsequently, three mixed-effects models—Statistical-based generalized linear mixed-effects model (GLMM), generalized additive mixed-effects model (GAMM), and machine learning-based tree-boosted mixed-effects model (TBMM)—were adopted. These models were used to evaluate the susceptibility of three distinct types of mass movements (i.e., 28,814 debris flows, 54,586 rockfalls and 108,432 landslides), respectively. The results were compared both from quantitative and qualitative perspectives. The results showed that TBMM performed best in all three cases with AUROCs (Area Under the Receiver Operating Characteristic curve) of cross-validation, spatial cross-validation, and predictions on simulated highly biased inventory, all exceeding 0.8. In addition, the spatial prediction patterns of TBMM were more in line with the natural geomorphological underlying process, indicating that TBMM can better reduce the impact of inventory bias than GLMM and GAMM. Finally, factor contribution analysis showed the key role of topographic factors in predicting the occurrence of mass movements, followed by road density and soil moisture. This study contributes to assessing China’s overall mass movement susceptibility situation and assisting policymakers in master planning for risk mitigation. Further, it demonstrates the tremendous potential of TBMM for mass movement susceptibility assessment, despite inherent biases in the inventory.