DTVT: a GIS tool for the automatic validation of Physically Based Landslide Models and the identification of the optimal warning criterium

Abstract

<p>Distributed physically based slope stability models represent the most advanced and scientifically sound method to forecast landslide triggering conditions. However, their operational application in regional warning systems is still hindered by some limitations. Among these, the problem of a robust validation (a task that is time consuming and not standardized) and the difficulty to manage a model output that (especially in the most advanced applications) is constituted by a raster of small pixels expressing the probability of landslide triggering: to activate an operational response an evaluation is usually performed on the overall conditions of larger spatial units and not on a pixel basis.</p><p>To overcome these shortcomings, we developed a GIS tool that can be fed with the results of slope stability models (raster maps representing the probability of landslide occurrence) and landslide inventory maps. The tool automatically performs a long series of operations traditionally performed by GIS operators to validate their models: the raw instability maps are reaggregated from pixels to watershed; warning maps are drawn; they are compared with the landslide inventory; a contingency matrix (with true positives, true negatives, false positive, and false negatives) is built; the validation results are drawn in a map. The warning criterium is defined based on two threshold values:  the probability of failure above which a pixel should be considered stable and the percentage of unstable pixels that a watershed needs to consider the hazard level widespread enough to justify the issuing of an alert. The tool was named Double Threshold Validation Tool (DTVT) and after some tests in three different test sites it was verified that: (i) DTVT can be used to carry out a standardized validation procedure in a very shorter time than traditional methods (ii) a reiterated application of the tool (by varying the values of the thresholds) can be used to identify the best warning criterion for each test site (e.g. which double threshold maximizes correct predictions while minimizing missed alarms). It is important to stress that DTVT does not improve the results obtained with the slope stability model; instead, this newly proposed tool that can be used to shift form a triggering model to a warning model, the latter being aimed at identifying when larger spatial units need the activation of operational procedures.</p>