Spatial Modelling of Landslide Occurrence; Landslide Susceptibility Assessment

Organised by Laboratory of Geo-information Science and Remote Sensing

Mon 30 October 2017 14:00 to 14:30

Venue Gaia, gebouwnummer 101
Room 1

By Tim Weerman (the Netherlands)


For the Collazzone study area, in central Umbria, ltaly, landslide susceptibility modelling was performed using multi-variate logistic regression. We assessed the effect of two different sampling strategies from the multi-temporal landslide inventory on landslide susceptibility modelling. The first sampling strategy was conjunctive, which selects the older time slices for calibration and newer time slices for validation. The second sampling strategy was disjunctive where the temporal range between calibration and validation datasets are relatively equal.

An interview has been performed to discover the preferred landslide susceptibility modeIs by landslide susceptibility modelling experts. The consensus from the landslide susceptibility experts was that there is not yet a best working model. Landslide susceptibility modelling is a context specific problem; it depends on the data you have and the history of the data. The choice for landslide susceptibility model is therefore based on which model served as a good reference model. We calculated our results with logistic regression modelling. We compared the results of landslide susceptibility modelling in conjunctive and disjunctive separation of time slices of the multi-temporal landslide inventory. The AUC values of ROC curves for the calibration and validation were compared.

The AUC values for calibration were 0.856 in conjunctive approach, and 0.842 in disjunctive approach. The AUC values for validation were 0.745, and 0.744 in conjunctive approach, and 0.793 for the disjunctive approach. The resulting AUC values show that the quality of the landslide susceptibility model runs were all either within 0.7 and 0.8 which is an acceptable validation score or above 0.8 which is an excellent validation score. Compared to the conjunctive approach, the AUC value of the disjunctive separation of time slices improved by 0.048-0.049 for validation while it had to give in 0.012 on calibration. The disjunctive separation of time slices of the multi-temporal landslide inventory is therefore a viable option to use in future research.

Keywords: Landslide susceptibility; multi-temporal landslide inventory; Logistic Regression; ROC; LAND-SE