Hydro-morphological processes(HMP,any natural phenomenon contained within the spectrum defined between debris flows and flash floods)are globally occurring natural hazards which pose great threats to our society,leadi...Hydro-morphological processes(HMP,any natural phenomenon contained within the spectrum defined between debris flows and flash floods)are globally occurring natural hazards which pose great threats to our society,leading to fatalities and economical losses.For this reason,understanding the dynamics behind HMPs is needed to aid in hazard and risk assessment.In this work,we take advantage of an explainable deep learning model to extract global and local interpretations of the HMP occurrences across the whole Chinese territory.We use a deep neural network architecture and interpret the model results through the spatial pattern of SHAP values.In doing so,we can understand the model prediction on a hierarchical basis,looking at how the predictor set controls the overall susceptibility as well as doing the same at the level of the single mapping unit.Our model accurately predicts HMP occurrences with AUC values measured in a ten-fold cross-validation ranging between 0.83 and 0.86.This level of predictive performance attests for an excellent prediction skill.The main difference with respect to traditional statistical tools is that the latter usually lead to a clear interpretation at the expense of high performance,which is otherwise reached via machine/deep learning solutions,though at the expense of interpretation.The recent development of explainable Al is the key to combine both strengths.In this work,we explore this combination in the context of HMP susceptibility modeling.Specifically,we demonstrate the extent to which one can enter a new level of data-driven interpretation,supporting the decision-making process behind disaster risk mitigation and prevention actions.展开更多
The literature on landslide susceptibility is rich with examples that span a wide range of topics.However,the component that pertains to the extension of the susceptibility framework toward space–time modeling is lar...The literature on landslide susceptibility is rich with examples that span a wide range of topics.However,the component that pertains to the extension of the susceptibility framework toward space–time modeling is largely unexplored.This statement holds true,particularly in the context of landslide risk,where few scientific contributions investigate risk dynamics in space and time.This manuscript proposes a modeling protocol where a dynamic landslide susceptibility is obtained via a binomial Generalized Additive Model whose inventories span nine years(from 2013 to 2021).For the analyses,the data cube is organized with a mapping unit consisting of 26,333 slope units repeated over an annual temporal unit,resulting in a total of 236,997 units.This phase already includes several interesting modeling experiments that have rarely appeared in the landslide literature(e.g.,variable interaction plots).However,the main innovative effort is in the subsequent phase of the protocol we propose,as we used climate projections of the main trigger(rainfall)to obtain future estimates of yearly susceptibility patterns.These estimates are then combined with projections of urban settlements and associated populations to create a dynamic risk model,assuming vulnerability=1.Overall,this manuscript presents a unique example of such a modeling routine and offers a potential standard for administrations to make informed decisions regarding future urban development.展开更多
基金supported by the National Natural Science Foundation of China(grant no.42201452)the Fundamental Research Funds for the Central Universities(grant no.2412022QD003)the support from the China Institute of Water Resources and Hydropower Research(IWHR).
文摘Hydro-morphological processes(HMP,any natural phenomenon contained within the spectrum defined between debris flows and flash floods)are globally occurring natural hazards which pose great threats to our society,leading to fatalities and economical losses.For this reason,understanding the dynamics behind HMPs is needed to aid in hazard and risk assessment.In this work,we take advantage of an explainable deep learning model to extract global and local interpretations of the HMP occurrences across the whole Chinese territory.We use a deep neural network architecture and interpret the model results through the spatial pattern of SHAP values.In doing so,we can understand the model prediction on a hierarchical basis,looking at how the predictor set controls the overall susceptibility as well as doing the same at the level of the single mapping unit.Our model accurately predicts HMP occurrences with AUC values measured in a ten-fold cross-validation ranging between 0.83 and 0.86.This level of predictive performance attests for an excellent prediction skill.The main difference with respect to traditional statistical tools is that the latter usually lead to a clear interpretation at the expense of high performance,which is otherwise reached via machine/deep learning solutions,though at the expense of interpretation.The recent development of explainable Al is the key to combine both strengths.In this work,we explore this combination in the context of HMP susceptibility modeling.Specifically,we demonstrate the extent to which one can enter a new level of data-driven interpretation,supporting the decision-making process behind disaster risk mitigation and prevention actions.
基金This research was supported by the National Natural Science Foundation of China-Young Scientist Funds(No.42207174)。
文摘The literature on landslide susceptibility is rich with examples that span a wide range of topics.However,the component that pertains to the extension of the susceptibility framework toward space–time modeling is largely unexplored.This statement holds true,particularly in the context of landslide risk,where few scientific contributions investigate risk dynamics in space and time.This manuscript proposes a modeling protocol where a dynamic landslide susceptibility is obtained via a binomial Generalized Additive Model whose inventories span nine years(from 2013 to 2021).For the analyses,the data cube is organized with a mapping unit consisting of 26,333 slope units repeated over an annual temporal unit,resulting in a total of 236,997 units.This phase already includes several interesting modeling experiments that have rarely appeared in the landslide literature(e.g.,variable interaction plots).However,the main innovative effort is in the subsequent phase of the protocol we propose,as we used climate projections of the main trigger(rainfall)to obtain future estimates of yearly susceptibility patterns.These estimates are then combined with projections of urban settlements and associated populations to create a dynamic risk model,assuming vulnerability=1.Overall,this manuscript presents a unique example of such a modeling routine and offers a potential standard for administrations to make informed decisions regarding future urban development.
