Annette I. Patton

Probabilistic models to inform landslide early warning systems often rely on rainfall totals observed during past events with landslides. However, these models are generally developed for broad regions using large catalogs, with dozens, hundreds, or even thousands of landslide occurrences. This study evaluates strategies for training landslide forecasting models with a scanty record of landslide-triggering events, which is a typical limitation in remote, sparsely populated regions. We evaluate 136 statistical models trained on a precipitation dataset with five landslide-triggering precipitation events recorded near Sitka, Alaska, USA, as well as > 6000 d of non-triggering rainfall (2002–2020). We also conduct extensive statistical evaluation for three primary purposes: (1) to select the best-fitting models, (2) to evaluate performance of the preferred models, and (3) to select and evaluate warning thresholds. We use Akaike, Bayesian, and leave-one-out information criteria to compare the 136 models, which are trained on different cumulative precipitation variables at time intervals ranging from 1 h to 2 weeks, using both frequentist and Bayesian methods to estimate the daily probability and intensity of potential landslide occurrence (logistic regression and Poisson regression). We evaluate the best-fit models using leave-one-out validation as well as by testing a subset of the data. Despite this sparse landslide inventory, we find that probabilistic models can effectively distinguish days with landslides from days without slide activity. Our statistical analyses show that 3 h precipitation totals are the best predictor of elevated landslide hazard, and adding antecedent precipitation (days to weeks) did not improve model performance. This relatively short timescale of precipitation combined with the limited role of antecedent conditions likely reflects the rapid draining of porous colluvial soils on the very steep hillslopes around Sitka. Although frequentist and Bayesian inferences produce similar estimates of landslide hazard, they do have different implications for use and interpretation: frequentist models are familiar and easy to implement, but Bayesian models capture the rare-events problem more explicitly and allow for better understanding of parameter uncertainty given the available data. We use the resulting estimates of daily landslide probability to establish two decision boundaries that define three levels of warning. With these decision boundaries, the frequentist logistic regression model incorporates National Weather Service quantitative precipitation forecasts into a real-time landslide early warning “dashboard” system (https://sitkalandslide.org/, last access: 9 October 2023). This dashboard provides accessible and data-driven situational awareness for community members and emergency managers.

Probabilistic models to inform landslide early warning systems often rely on rainfall totals observed during past events with landslides. However, these models are generally developed for broad regions using large catalogs, with dozens, hundreds, or even thousands of landslide occurrences. This study evaluates strategies for training landslide forecasting models with a scanty record of landslide-triggering events, which is a typical limitation in remote, sparsely populated regions. We train and evaluate 136 statistical models with a rainfall dataset with five landslide-triggering rainfall events recorded near Sitka, Alaska, USA, as well as >6,000 days of non-triggering rainfall (2002–2020). We use Akaike, Bayesian, and leave-one-out information criteria to compare models trained on cumulative precipitation at timescales ranging from 1 hour to 2 weeks, using both frequentist and Bayesian methods to estimate the daily probability and intensity of potential landslide occurrence (logistic regression and Poisson regression). We evaluate the best-fit models using leave-one-out validation as well as with testing a subset of the data. Despite this sparse landslide inventory, we find that probabilistic models can effectively distinguish days with landslides from days without. Although frequentist and Bayesian inference produce similar estimates of landslide hazard, they do have different implications for use and interpretation: frequentist models are familiar and easy to implement, but Bayesian models capture the rare-events problem more explicitly and allow for better understanding of parameter uncertainty given the available data. Three-hour precipitation totals are the best predictor of elevated landslide hazard, and adding antecedent precipitation (days to weeks) did not improve model performance. This relatively short timescale combined with the limited role of antecedent conditions reflects the rapid draining of porous colluvial soils on very steep hillslopes around Sitka. We use the resulting estimates of daily landslide probability to establish two decision boundaries for three levels of warning. With these decision boundaries, the frequentist logistic regression model incorporates National Weather Service quantitative precipitation forecasts into a real-time landslide early warning “dashboard” system (sitkalandslide.org). This dashboard provides accessible and data-driven situational awareness for community members and emergency managers.

Probabilistic models to inform landslide early warning systems often rely on rainfall totals observed during past events with landslides. However, these models are generally developed for broad regions using large catalogs, with dozens, hundreds, or even thousands of landslide occurrences. This study evaluates strategies for training landslide forecasting models with a scanty record of landslide-triggering events, which is a typical limitation in remote, sparsely populated regions. We train and evaluate 136 statistical models with a rainfall dataset with five landslide-triggering rainfall events recorded near Sitka, Alaska, USA, as well as >6,000 days of non-triggering rainfall (2002–2020). We use Akaike, Bayesian, and leave-one-out information criteria to compare models trained on cumulative precipitation at timescales ranging from 1 hour to 2 weeks, using both frequentist and Bayesian methods to estimate the daily probability and intensity of potential landslide occurrence (logistic regression and Poisson regression). We evaluate the best-fit models using leave-one-out validation as well as with testing a subset of the data. Despite this sparse landslide inventory, we find that probabilistic models can effectively distinguish days with landslides from days without. Although frequentist and Bayesian inference produce similar estimates of landslide hazard, they do have different implications for use and interpretation: frequentist models are familiar and easy to implement, but Bayesian models capture the rare-events problem more explicitly and allow for better understanding of parameter uncertainty given the available data. Three-hour precipitation totals are the best predictor of elevated landslide hazard, and adding antecedent precipitation (days to weeks) did not improve model performance. This relatively short timescale combined with the limited role of antecedent conditions reflects the rapid draining of porous colluvial soils on very steep hillslopes around Sitka. We use the resulting estimates of daily landslide probability to establish two decision boundaries for three levels of warning. With these decision boundaries, the frequentist logistic regression model incorporates National Weather Service quantitative precipitation forecasts into a real-time landslide early warning “dashboard” system (sitkalandslide.org). This dashboard provides accessible and data-driven situational awareness for community members and emergency managers.

Probabilistic models to inform landslide early warning systems often rely on rainfall totals observed during past events with landslides. However, these models are generally developed for broad regions using large catalogs, with dozens, hundreds, or even thousands of landslide occurrences. This study evaluates strategies for training landslide forecasting models with a scanty record of landslide-triggering events, which is a typical limitation in remote, sparsely populated regions. We train and evaluate 136 statistical models with a rainfall dataset with five landslide-triggering rainfall events recorded near Sitka, Alaska, USA, as well as >6,000 days of non-triggering rainfall (2002–2020). We use Akaike, Bayesian, and leave-one-out information criteria to compare models trained on cumulative precipitation at timescales ranging from 1 hour to 2 weeks, using both frequentist and Bayesian methods to estimate the daily probability and intensity of potential landslide occurrence (logistic regression and Poisson regression). We evaluate the best-fit models using leave-one-out validation as well as with testing a subset of the data. Despite this sparse landslide inventory, we find that probabilistic models can effectively distinguish days with landslides from days without. Although frequentist and Bayesian inference produce similar estimates of landslide hazard, they do have different implications for use and interpretation: frequentist models are familiar and easy to implement, but Bayesian models capture the rare-events problem more explicitly and allow for better understanding of parameter uncertainty given the available data. Three-hour precipitation totals are the best predictor of elevated landslide hazard, and adding antecedent precipitation (days to weeks) did not improve model performance. This relatively short timescale combined with the limited role of antecedent conditions reflects the rapid draining of porous colluvial soils on very steep hillslopes around Sitka. We use the resulting estimates of daily landslide probability to establish two decision boundaries for three levels of warning. With these decision boundaries, the frequentist logistic regression model incorporates National Weather Service quantitative precipitation forecasts into a real-time landslide early warning “dashboard” system (sitkalandslide.org). This dashboard provides accessible and data-driven situational awareness for community members and emergency managers.

Probabilistic models to inform landslide early warning systems often rely on rainfall totals observed during past events with landslides. However, these models are generally developed for broad regions using large catalogs, with dozens, hundreds, or even thousands of landslide occurrences. This study evaluates strategies for training landslide forecasting models with a scanty record of landslide-triggering events, which is a typical limitation in remote, sparsely populated regions. We train and evaluate 136 statistical models with a rainfall dataset with five landslide-triggering rainfall events recorded near Sitka, Alaska, USA, as well as >6,000 days of non-triggering rainfall (2002–2020). We use Akaike, Bayesian, and leave-one-out information criteria to compare models trained on cumulative precipitation at timescales ranging from 1 hour to 2 weeks, using both frequentist and Bayesian methods to estimate the daily probability and intensity of potential landslide occurrence (logistic regression and Poisson regression). We evaluate the best-fit models using leave-one-out validation as well as with testing a subset of the data. Despite this sparse landslide inventory, we find that probabilistic models can effectively distinguish days with landslides from days without. Although frequentist and Bayesian inference produce similar estimates of landslide hazard, they do have different implications for use and interpretation: frequentist models are familiar and easy to implement, but Bayesian models capture the rare-events problem more explicitly and allow for better understanding of parameter uncertainty given the available data. Three-hour precipitation totals are the best predictor of elevated landslide hazard, and adding antecedent precipitation (days to weeks) did not improve model performance. This relatively short timescale combined with the limited role of antecedent conditions reflects the rapid draining of porous colluvial soils on very steep hillslopes around Sitka. We use the resulting estimates of daily landslide probability to establish two decision boundaries for three levels of warning. With these decision boundaries, the frequentist logistic regression model incorporates National Weather Service quantitative precipitation forecasts into a real-time landslide early warning “dashboard” system (sitkalandslide.org). This dashboard provides accessible and data-driven situational awareness for community members and emergency managers.

Probabilistic models to inform landslide early warning systems often rely on rainfall totals observed during past events with landslides. However, these models are generally developed for broad regions using large catalogs, with dozens, hundreds, or even thousands of landslide occurrences. This study evaluates strategies for training landslide forecasting models with a scanty record of landslide-triggering events, which is a typical limitation in remote, sparsely populated regions. We train and evaluate 136 statistical models with a rainfall dataset with five landslide-triggering rainfall events recorded near Sitka, Alaska, USA, as well as >6,000 days of non-triggering rainfall (2002–2020). We use Akaike, Bayesian, and leave-one-out information criteria to compare models trained on cumulative precipitation at timescales ranging from 1 hour to 2 weeks, using both frequentist and Bayesian methods to estimate the daily probability and intensity of potential landslide occurrence (logistic regression and Poisson regression). We evaluate the best-fit models using leave-one-out validation as well as with testing a subset of the data. Despite this sparse landslide inventory, we find that probabilistic models can effectively distinguish days with landslides from days without. Although frequentist and Bayesian inference produce similar estimates of landslide hazard, they do have different implications for use and interpretation: frequentist models are familiar and easy to implement, but Bayesian models capture the rare-events problem more explicitly and allow for better understanding of parameter uncertainty given the available data. Three-hour precipitation totals are the best predictor of elevated landslide hazard, and adding antecedent precipitation (days to weeks) did not improve model performance. This relatively short timescale combined with the limited role of antecedent conditions reflects the rapid draining of porous colluvial soils on very steep hillslopes around Sitka. We use the resulting estimates of daily landslide probability to establish two decision boundaries for three levels of warning. With these decision boundaries, the frequentist logistic regression model incorporates National Weather Service quantitative precipitation forecasts into a real-time landslide early warning “dashboard” system (sitkalandslide.org). This dashboard provides accessible and data-driven situational awareness for community members and emergency managers.

Inclusive, participatory governance is a key enabler of effective responses to natural hazard risks exacerbated by climate change. This paper describes a community-level co-design process among academic, state, and federal scientists and the community of Sitka, Alaska to develop a novel landslide warning system for this small coastal town. The decentralized system features an online dashboard which displays current and forecast risk levels to help residents make their own risk management decisions. The system and associated risk communications are informed by new geoscience, social, and information science generated during the course of the project. This case study focuses on our project team’s activities and addresses questions including: what activities did the project team conduct, what did these activities intend to accomplish, and did these activities accomplish what they intended? The paper describes the co-design process, the associated changes in system design and research activities, and formal and informal evaluations of the system and process. Overall, the co-design process appears to have generated a warning system the Sitka community finds valuable, helped to align system design with local knowledge and community values, significantly modified the scientists’ research agendas, and helped navigate sensitivities such as the effect of landslide exposure maps on property values. Other communities in SE Alaska are now adopting this engagement approach. The paper concludes with broader implications for the role of community-level, participatory co-design and risk governance for climate services.