A tool for the automatic aggregation and validation of the results of physically based distributed slope stability models

Distributed physically based slope stability models usually provide outputs representing, on a pixel basis, the probability of failure of each cell. This kind of result, although scientifically sound, from an operational point of view has several limitations. First, the procedure of validation lacks standards. As instance, it is not straightforward to decide above which percentage of failure probability a pixel (or larger spatial units) should be considered unstable. Second, the validation procedure is a time-consuming task, usually requiring a long series of GIS operations to overlap landslide inventories and model outputs to extract statistically significant performance metrics. Finally, if model outputs are conceived to be used in the operational management of landslide hazard (e.g., early warning procedures), the pixeled probabilistic output is difficult to handle and a synthesis to characterize the hazard scenario over larger spatial units is usually required to issue warnings aimed at specific operational procedures. In this work, a tool is presented that automates the validation procedure for physically based distributed probabilistic slope stability models and translates the pixeled outputs in warnings released over larger spatial units like small watersheds. The tool is named DTVT (double-threshold validation tool) because it defines a warning criterion on the basis of two threshold values-the probability of failure above which a pixel should be considered stable (failure probability threshold, FPT) and the percentage of unstable pixels needed in each watershed to consider the hazard level widespread enough to justify the issuing of an alert (instability diffusion threshold, IDT). A series of GIS operations were organized in a model builder to reaggregate the raw instability maps from pixels to watershed; draw the warning maps; compare them with an existing landslide inventory; build a contingency matrix counting true positives, true negatives, false positive, and false negatives; and draw in a map the results of the validation. The DTVT tool was tested in an alert zone of the Aosta Valley (northern Italy) to investigate the high sensitivity of the results to the values selected for the two thresholds. Moreover, among 24 different configurations tested, we performed a quantitative comparison to identify which criterion (in the case of our study, there was an 85% or higher failure probability in 5% or more of the pixels of a watershed) produces the most reliable validation results, thus appearing as the most promising candidate to be used to issue alerts during civil protection warning activities.