The world’s second-largest, recorded landslide event: lessons learnt from the landslides triggered during and after the 2018 Mw 7.5 Papua New Guinea earthquake

Abstract

Widespread landslide events provide rare but valuable opportunities to investigate the spatial and size distributions of landslides in relation to seismic, climatic, geological and morphological factors. This study presents a unique event inventory for the co-seismic landslides induced by the February 25, 2018 Mw 7.5 Papua New Guinea earthquake as well as its post-seismic counterparts including the landslides triggered by either aftershocks or succeeding rainfall events that occurred between February 26 and March 19. We mapped approximately 11,500 landslides of which more than 10,000 were triggered by the mainshock with a total failed planimetric area of about 145 km2. Such a large area makes this inventory the world’s second-largest recorded landslide event after the 2008 Wenchuan earthquake. Large landslides are abundant throughout the study area located within the remote Papua New Guinea Highlands. Specifically, more than half of the landslide population is larger than 50,000 m2 and overall, post-seismic landslides are even larger than their co-seismic counterparts. Our analyses indicate that large and widespread landslides were triggered as a result of the compound effects of the strong seismicity, complex geology, steep topography and high rainfall. We statistically show that the 15-day antecedent precipitation, as a predisposing factor, contributes to the spatial distribution of co-seismic landslides. Also, we statistically demonstrate that the cumulative effect of aftershocks is the main factor disturbing steep hillslopes and causing the initiation of very large landslides up to the size of ~5 km2. Taking aside the role of the intense seismic swarm and antecedent precipitation, these inventories also provide evidence for landslide events where the active tectonics contribute to weaken hillslopes and the fatigue damage. Overall, the dataset and the findings provided by this paper is a step forward in seismic landslide hazard assessment of the entire Papua New Guinea mainland.

Space-time landslide hazard modeling via Ensemble Neural Networks

“…In fact, our output could bring additional information on this topic supporting the scientific debate on landslide recovery (the time required for a given landscape to go back to pre-earthquake susceptibility conditions) by observing the predicted susceptibility change over time. Overall, multi-temporal landslide inventories and various associated parameters (e.g., number, size, area or volume of landslides) have already been used to explore landslide recovery in post-seismic periods (eg., Tanyas et al, 2021).…”

“…Slopes are also the same unit geotechnical solutions aim to address. Thus, an improvement to our ENN could involve moving away from a gridded spatial partition and towards more geomorphological-oriented mapping units such as slope units (Alvioli et al, 2016;Tanyaş et al, 2022b), sub-catchments or catchments (Shou and Lin, 2020;Wang et al, 2022).…”

Explainable artificial intelligence in geoscience: a glimpse into the future of landslide susceptibility modeling

“…For this reason, response plots like Figure 7 add another level of understanding for they allow to monitor variations in SHAP with respect to each predictors' domain. This is a capability which is typical of statistical models (Lima et al, 2021;Tanyaş et al, 2022) and has found very few applications in machine learning (Park, 2015;Vorpahl et al, 2012). However, even in this case, the level of information provided is very generic and corresponds to the overall behaviour of each predictor with respect to the entire map it contributes to define.…”

“…The way they work is to assume a vector of landslide presence/absence data to behave across the geographic space according to a Bernoulli probability distribution, whose relation to the landslide is linearly related to a set of covariates. The latter are usually referred to as predisposing or triggering factors (Das et al, 2012;Tanyaş et al, 2022). However, the linearity assumption these models are based on, limited the performance one could obtain.…”

A data-driven framework for landslide size space-time modelling

“…In fact, it is common that a landslide area distribution is quite heavily tailed. In other words, the vast majority of inventories includes a predominant number of small landslides and only few extremely large ones, which is common in response to major triggering events, such as rainfall (Jones et al, 2021;Emberson et al, 2022) or earthquakes (Zhang et al, 2019;Tanyaş et al, 2022). This is the reason that has led Malamud et al (2004) to propose the Inverse Gamma distribution as a universal empirical size model, which lead to a series of studies on landslide Frequency Area Distribution (FAD; .…”