Predicting seasonal landslide activity with Bayesian inference

Improving landslide prediction in time is key to reducing damage and fatalities in areas susceptible to landsliding. While most landslide early warning research has focused on establishing hydro-meteorological landslide thresholds on hourly to daily timescales, few studies globally have attempted to model or predict landslide seasonality. We use probabilistic models based on two intuitive metrics — counts of landslides and presence or absence of landslides — to predict landslide activity at monthly resolution. Our focus area is the Pacific Northwest region of the United States, which has one of the highest densities of landsliding in the country, and where seasonal landslide activity has been recognized but hardly quantified. We use Bayesian inference to combine data from five landslide inventories from the region with varying spatial and temporal coverage, data density, and reporting protocols to learn the regional pattern of seasonal landslide activity. Results of logistic and negative binomial regression show that the landslide season in the Pacific Northwest begins in November and is marked by credible increases in the probability of landsliding, average landslide intensity, and inter-annual variability. Landslide activity is highest between November and February, decreases from March through May, and stays low between June and October. Inter-annual variability in landslide activity is higher in winter than in summer months. These flexible models could be easily adapted to learn diverse seasonal patterns from other regions of the world, such as the East Asian Summer Monsoon peak observed in Japan or the Atlantic hurricane season fall peak seen in the Caribbean. Our results also show that Bayesian multi-level models are a promising way to combine data from multiple, seemingly incompatible landslide inventories from a single region with potentially wide-ranging future applications.