Cross model sensitivity analysis of the 9 January 2018 debris-flow inundation event in Montecito, CA

Abstract:

Probabilistic forecasts of post-fire debris-flow inundation in urban areas have the potential to reduce harm to people and property. A first step towards developing such a product is understanding the inherent properties of different numerical models to assess the model’s potential utility. For example, how does uncertainty in model inputs—such as topography, volume of mobilized material, or parameters controlling rheology—propagate into uncertainty in model outputs—such as inundation area, depth, or flow velocity? We present a cross-model sensitivity analysis in which we document the properties of three numerical models in the context of the well-constrained 9 January 2018 debris-flow event in Montecito, CA. We consider three alternative models which vary in their representation of debris-flow physics: FLO2D and RAMMS are single phase models with different friction relationships, while D-CLAW represents both the granular and liquid phases.

We constrained total sediment volume and final inundation extent for multiple source basins using field data. No direct measurements of alluvial fan apex hydrographs exist, but we estimate them using rain gauge and seismic station data. We evaluate model sensitivity to topography (pre-event vs. post-event, bare-Earth topography vs. topography including infrastructure), sediment concentration, and values for the model input parameters that control the physical properties of the debris flow. We compared model outputs with the known inundation extent. As a first step, we assess variance in the four elements of the confusion matrix (true positive, true negative, false positive, and false negative), as well as other standard binary classification metrics (e.g., precision, recall, and accuracy). Our numerical experiment constrains each model’s internal sensitivities as well as variation between models which derives from physical representation and numerical implementation.

Documenting these sensitivities permits an assessment of how pre-event uncertainties propagate into uncertainty in inundation extent. Such results can support pre-event hazard assessment by identifying which observations reduce uncertainty in inundation the most, and where uncertainty is irreducible.