Abstract. The increasing availability of long-term observational data can lead to the development of innovative modelling approaches to determine landslide triggering conditions at a regional scale, opening new avenues for landslide prediction and early warning. This research blends the strengths of existing approaches with the capabilities of generalized additive mixed models (GAMMs) to develop an interpretable approach that identifies seasonally dynamic precipitation conditions for shallow landslides. The model builds upon a 21-year record of landslides in South Tyrol (Italy) and separates precipitation that induced landslides from precipitation that did not. The model accounts for effects acting at four temporal scales: short-term “triggering” precipitation, medium-term “preparatory” precipitation, seasonal effects, and across-year data variability. It provides relative landslide probability scores that were used to establish seasonally dynamic thresholds with optimal performance in terms of hit and false-alarm rates, as well as additional thresholds related to user-defined performance scores. The GAMM shows a high predictive performance and indicates that more precipitation is required to induce a landslide in summer than in winter/spring, which can presumably be attributed mainly to vegetation and temperature effects. The discussion illustrates why the quality of input data, study design, and model transparency are crucial for landslide prediction using advanced data-driven techniques.
Abstract. The increasing availability of long-term observational data can lead to the development of innovative modelling approaches to determine landslide triggering conditions at regional scale, opening new avenues for landslide prediction and early warning. This research blends the strengths of existing approaches with the capabilities of generalized additive mixed models (GAMMs) to develop an interpretable approach that identifies seasonally dynamic precipitation conditions for shallow landslides. The model builds upon a 21-year record of landslides in South Tyrol (Italy) and separates precipitation that induced landslides from precipitation that did not. The model accounts for effects acting at four temporal scales: short-term “triggering” precipitation, medium-term “preparatory” precipitation, seasonal effects and across-year data variability. It provides relative landslide probability scores that were used to establish seasonally dynamic thresholds with optimal performance in terms of hit and false alarm rates, as well as additional thresholds related to user-defined performance scores. The GAMM shows a high predictive performance and indicates that more precipitation is required to induce a landslide in summer than in winter/spring, which can presumably be attributed mainly to vegetation and temperature effects. The discussion illustrates why the quality of input data, study design and model transparency are crucial for landslide prediction using advanced data-driven techniques.
The last three decades have witnessed a substantial methodical development of data-driven models for landslide prediction. However, this improvement has been dedicated almost exclusively to models designed to recognize locations where landslides may likely occur in the future. This notion is referred to as landslide susceptibility. However, the susceptibility is just one, albeit fundamental, information required to assess landslide hazard and to mitigate the threat that landslides may pose to human lives and infrastructure. Another complementary and equally important information is how large landslides may evolve into, once they initiate in a given slope. Only three scientific contributions have currently addressed the geographic estimation of how large co-seismic landslides may be. In the first one, the authors tested a model solely at the global scale, whereas the remaining two involved specific regional scale settings. The low number of previous research on the topic as well as specificities related to the associated study areas do not yet allow to fully support a standardized use of such models. In turn, this has repercussions on the operational feasibility and adoption potential of data-driven models capable of estimating landslide size in site-specific conditions. This manuscript addresses this gap in the literature, by further exploring the use of a Generalized Additive Model whose target variable is the topographically-corrected landslide extent aggregated at the slope unit level. In our case, the underlying assumption is that the variability of the landslide sizes across the geographic space behaves according to a Log-Normal probability distribution. We test this framework by going beyond the conventional non-spatial validation scheme in order to take a particularly critical look at the estimated model performance.The study focuses on co-seismic landslides mapped as a result of the ground motion generated by the Kaikōura earthquake (11:02 UTC, on November 13th 2016). The experiment led to further insights into the applicability of such approaches and produced more than satisfying performance scores, which we stress here in the prospect of stimulating further research towards spatially-explicit landslide size prediction. In line with the same idea, we share data and codes in a github repository (https://github.com/Mateo3195/GAM_LandslideSize/tree/main) to promote repeatability and reproducibility of this research.
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.
hi@scite.ai
10624 S. Eastern Ave., Ste. A-614
Henderson, NV 89052, USA
Copyright © 2024 scite LLC. All rights reserved.
Made with 💙 for researchers
Part of the Research Solutions Family.