Coupled landslide analyses through dynamic susceptibility and forecastable hazard analysis

Landslides, specifically those triggered through an increase of soil moisture either during or after a rainfall event, pose severe threats to surrounding infrastructure. Herein, the term "landslide" refers primarily to translational movements of shallow colluvial soil upon a hillslope. These landslides are assumed to adhere to infinite slope approximations. Potential landslide occurrences are monitored through identification of areas susceptible to occurrence, through susceptibility analyses, or areas likely to experience a landslide at a given time, through hazard analyses. Traditional landslide susceptibility systems are created as a function of static geomorphologic data. This is to say that, while spatially differing, susceptibility via this system does not change with time. Landslide hazard analyses consider dynamic data, such as that of precipitation, and provide warnings of when landslide occurrences are likely. However, these hazard analysis systems typically only provide warnings in near real time (i.e., over the next few days). Therefore, dynamic susceptibility (susceptibility that is seen to change with time rather than remain static) as well as the ability to forecast landslide hazard analyses beyond real time is desired. The study herein presents a novel workflow for the creation of dynamic landslide susceptibility and forecastable hazard analyses over a domain within Eastern Kentucky. Dynamic susceptibility was developed through inclusion of static geomorphic parameters and dynamic vegetation levels over sites of interest. These susceptibility data were used in the development of a logistic regression classification machine learning approach which yielded susceptibility classifications with an accuracy of 89%. Forecastable hazard analyses were developed as a function of forecasted soil moisture, assumed to be a controlling factor in landslide occurrence, over a site. Forecasting of soil moisture was conducted through development of a Long Short-Term Memory (LSTM) forecasting machine learning system. Forecasts of soil moisture were then assimilated into an infinite slope stability equation to provide forecasts of hazard analyses. These forecasted hazard analyses were investigated over known landslides with satisfactory results obtained. Therefore, this study presents a novel workflow for both dynamic and forecastable hazard analyses that will undoubtedly provide greater warning and preparation periods to those within landslide prone regions.