Quantifying livestock vulnerability to snow disasters in the Tibetan Plateau: Comparing different modeling techniques for prediction

Abstract

Quantitative vulnerability relationships describing the susceptibility of socioeconomic losses in response to climate change are critical for natural disaster loss modeling and risk assessment. Modeling such vulnerability requires methods capable of handling complicated multi-factor, non-linear, and interactive relationships. Here, we compared the performance of generalized additive models (GAM) and random forest (RF) and boosted regression trees (BRT) in quantifying livestock vulnerability to snow disasters in the Tibetan Plateau for both explanatory and predictive purposes. Our results indicated promising explanatory power of these three modeling methods, with deviance-based R2 up to 0.720. They consistently revealed geophysical and socioeconomic factors that contributed to higher mortality rates. Nevertheless, GAM model failed to identify the critical influence of snow depth, mainly due to its smoothing scheme when fitting models to data. They also differed in the selection of the most important socioeconomic variable to represent prevention capacity. From a predictive perspective, all three modeling methods also showed promising predictive power, yet RF had the smallest prediction error, with less number of predictors used. Therefore, the predictive version of RF may well be the best choice for use in future risk analyses, yet those of BRT and GAM can serve as an alternative if needed.