On the prediction of landslide occurrences and sizes via Hierarchical Neural Networks

Abstract

For more than three decades, the part of the geoscientific community studying landslides through data-driven models has focused on estimating where landslides may occur across a given landscape. This concept is widely known as landslide susceptibility. And, it has seen a vast improvement from old bivariate statistical techniques to modern deep learning routines. Despite all these advancements, no spatially-explicit data-driven model is currently capable of also predicting how large landslides may be once they trigger in a specific study area. In this work, we exploit a model architecture that has already found a number of applications in landslide susceptibility. Specifically, we opt for the use of Neural Networks. But, instead of focusing exclusively on where landslides may occur, we extend this paradigm to also spatially predict classes of landslide sizes. As a result, we keep the traditional binary classification paradigm but we make use of it to complement the susceptibility estimates with a crucial information for landslide hazard assessment. We will refer to this model as Hierarchical Neural Network (HNN) throughout the manuscript. To test this analytical protocol, we use the Nepalese area where the Gorkha earthquake induced tens of thousands of landslides on the 25th of April 2015. The results we obtain are quite promising. The component of our HNN that estimates the susceptibility outperforms a binomial Generalized Linear Model (GLM) baseline we used as benchmark. We did this for a GLM represents the most common classifier in the landslide literature. Most importantly, our HNN also suitably performed across the entire procedure. As a result, the landslide-area-class prediction returned not just a single susceptibility map, as per tradition. But, it also produced several informative maps on the expected landslide size classes. Our vision is for administrations to consult these suite of model outputs and maps to better assess the risk to local communities and infrastructure. And, to promote the diffusion of our HNN, we are sharing the data and codes in a githubsec repository in the hope that we would stimulate others to replicate similar analyses.

Space-time landslide hazard modeling via Ensemble Neural Networks

“…So far, no spatially nor temporally explicit model exists for landslide area density. However, four recent articles have explored the capacity of predicting landslide areas (Lombardo et al, 2021;Aguilera et al, 2022;Bryce et al, 2022;Moreno et al, 2022). All of them have returned suitable predictive performance, but still far from the match seen in the second panel of Figure 7, between observed and predicted landslide density.…”

“…In their work, the authors exclusively estimated the potential landslide size at a given location, without informing whether the given location would have been susceptible in the first place. This limitation has been further addressed by Bryce et al (2022) and Aguilera et al (2022), implementing models that couple susceptibility and landslide area prediction together. Nevertheless, even in these cases, the absence of the temporal dimension in their work implies that no current datadriven model has even been capable to solve the landslide hazard definition (Guzzetti et al, 1999), jointly estimating where, when (or how frequently) and how large landslides may be in a given spatio-temporal domain.…”

Space-time landslide susceptibility modelling in Taiwan

“…The progress initially welcomed bivariate statistical models (e.g., Van Westen et al, 2003), and naturally evolved towards their multivariate counterparts mainly represented by generalized linear models (e.g., Atkinson and Massari, 1998). The multivariate context further differentiated over time, in the form of machine learning models (e.g., Marjanović et al, 2011) and their deep learning (Wang et al, 2019;Fang et al, 2021;Aguilera et al, 2022) extensions. In this plethora of available solutions, the way a potential user may navigate through them and understand their strength and weaknesses mainly depends on two elements.…”

Multi-hazard susceptibility mapping of cryospheric hazards in a high-Arctic environment: Svalbard Archipelago

“…Our modelling strategy relies on a Generalised Additive Model (Titti et al, 2021). This class of models ensures the same level of interpretability of the simpler and more common Generalised Additive Model (Atkinson et al, 1998;Titti et al, 2022) while providing much higher performance, close to more complex architectures belonging to machine/deep learning (Aguilera et al, 2022). GAM can be used to explain data distributed in a few exponential family distributions (Gamma, Gaussian, etc.).…”

“…Such direction has recently been explored for landslides occurring at lower latitudes (Lombardo et al, 2021;Moreno et al, 2022). And, an even better extension has already been tested where the expectation of locations prone to landslides are modelled together with the expectation of the resulting landslide size (Aguilera et al, 2022;Bryce et al, 2022).…”