Artificial neural networks applied to landslide susceptibility: The effect of sampling areas on model capacity for generalization and extrapolation

Abstract

Artificial neural networks (ANNs) have been used to identify areas susceptible to landslides and constitute one of the most widely used methods for this purpose. Several factors can interfere in the performance of the models and their resulting maps (especially sampling). This research evaluated the influence of sampling areas on landslide susceptibility modelling and the capacity for generalization and spatial extrapolation of data. Based on an inventory of landslide scars, distributed in five areas of southern Brazil, non-occurrence samples were defined by means of different buffers (2–40 km) in relation to the landslides in order to test the effect of the spatial distribution of the non-occurrence samples on the modeling results. A total of 16 morphometric attributes of the terrain (extracted from a digital elevation model) were used as input variables of the model. Multilayered network training was carried out using a backpropagation algorithm and accuracy was calculated by means of the Area Under the Receiver Operating Characteristic Curve (AUROC). Model accuracy was between 0.739 and 0.931. This variation was explained mainly by the buffer used. The susceptibility map resulting from the model of greater accuracy was obtained with a 40-km buffer in order to collect non-occurrence samples. The great distance between the occurrence and non-occurrence samples facilitates the modelling, since it increases the morphometric differences between the sampling groups. When we used samples from only one of the sample areas, the spatial extrapolation of the susceptibility map to the other areas showed high performance. We conclude that the ANN model for landslides susceptibility mapping can be extrapolated spatially, considering the limits of the geomorphological unit or numerical domain of the data.