A Markov Switching Spatiotemporal GAM for Landslide Hazards in New Zealand

Landslides triggered by earthquakes evolve over time, leading to repeated damage in the affected areas. These slope movements are influenced by a range of factors, including climatic, seismic, and terrain conditions, which vary both temporally and spatially [1]. To predict the likelihood of landslides occurring across different times and locations, statistical models must account for these spatial and temporal dependencies. In this study, we employ the Markov Switching Spatiotemporal Generalized Additive Model (MSST-GAM), as introduced by Sridharan et al. [2]. Their research highlighted how this model effectively captures the spatial and temporal influences of various landslide-related factors, offering accurate susceptibility estimates for the Wenchuan area in China.

In this work, we further extend the model for hazard prediction. The model is used on a multitemporal dataset of landslides that occurred in New Zealand during and following the 2016 Kaikoura earthquake [3]. The years in which the landslides were mapped were used to separate the temporal units. Twelve covariates were used, including terrain (slope, aspect, curvature, distance from features like faults, etc.), climatic (rainfall and soil moisture), and seismic (when the year coincided with a major seismic event). We employ zero-inflated Poisson and Gaussian emission probabilities [4] for the dependent variables, which are the areas and counts of landslides in slope units. A Markov-switching GAM is used to predict the dependent variables from the covariables based on two hidden risk states (high risk and low risk). We introduce soil moisture as an additional dynamic variable to parametrize the transition probabilities between the hidden states. 

We tested the model using a five-fold spatiotemporal cross-validation. The results compare favourably to a number of cross-sectional models [5]. The model predictions indicate that MSST-GAM can capture the spatial and temporal dependence of the landslide occurrences in slope units when compared with other cross-sectional and spatiotemporal models in literature.

References

[1] Keefer, D., “Investigating landslides caused by earthquakes - A historical review,” Surv. Geophys., vol. 23, no. 6, pp. 473–510, 2002 

[2] Sridharan, A., Gutjahr, G., and Gopalan, S., “Markov–Switching Spatio–Temporal Generalized Additive Model for Landslide Susceptibility,” Environ. Model. Softw., vol. 173, no. August, p. 105892, Feb. 2024 

[3]  Bhuyan, K., Tanyaş, H., Nava, L. et al. “Generating multi-temporal landslide inventories through a general deep transfer learning strategy using HR EO data”. Sci Rep 13, 162, 2023

[4] Wagh, Y.S. and Kamalja, K.K., 2018. “Zero-inflated models and estimation in zero-inflated Poisson distribution”. Communications in Statistics-Simulation and Computation, 47(8), pp.2248-2265.

[5] Reichenbach, P., Rossi, M., Malamud, B., Mihir, M., Guzzetti, F. “A review of statistically-based landslide susceptibility models”. Earth-science reviews. 2018 May 1;180:60-91.