Landslide susceptibility mapping using artificial neural networks

Abstract

Landslides are one of the main phenomena responsible for natural disasters in Brazil. Mapping can assess the spatial planning of susceptible areas. Artificial neural networks (ANN) stand out the susceptibility modeling and mapping by their high accuracy, as well as capacity learning and generalizing their results. Thus, this study aimed to map susceptible areas to landslides, considering four different sample sets, from an ANN model. For this, an inventory of landslides was drawn, with terrain attributes extracted and their importance to models analyzed, sample sets were organized according to two sample areas and two resampling processes, training, validation and test of the predictive models, followed by reclassification and spatialization of the susceptible areas. The were identified 297 landslides scars, covering a total area of 1.06 km². The most important predictive variables were elevation, slope, LS factor and valley depth. It was observed that the restriction of area for random sampling of non-occurrence may affect the model generalization capacity, while the reduction of the training sample set decreases the processing time, without significantly interfering with the accuracy. The ANN were able to model the susceptible areas, with mapping accuracy near or greater than 0.9.