Evaluation of prediction capability of the artificial neural networks for mapping landslide susceptibility in the Turbolo River catchment (northern Calabria, Italy)

Landslides are one of the most widespread natural hazards that cause damage to both property and life every year, and therefore, the spatial distribution of the landslide susceptibility is necessary for planning future developmental activities. In this paper the artificial neural network (ANN) technique is tested for developing a landslide susceptibility map in Turbolo River catchment, North Calabria, South Italy. Landslides were mapped through air-photo interpretation and field surveys, by identifying both the landslide depletion zones (DZs) and accumulation zones (AZs); and relevant geo-environmental thematic layers pertaining to landslide predisposing factors were generated using air-photo interpretation, field surveys and Geographic Information System (GIS) tools. Ten predisposing factors were related to the occurrence of landslide: lithology, faults, land use, drainage network, and a series of topographic factors: elevation, slope, aspect, plan curvature, topographic wetness index (TWI) and stream power index (SPI). In order to evaluate and validate landslide susceptibility, the DZs were divided in two groups using a random partition strategy. The first group (training set) was used to prepare the susceptibility map, employing a back-propagation learning algorithm in the Idrisi Taiga software. The second group (testing set) was used to validate the landslide susceptibility model, using the confusion matrix and the receiver operating characteristic (ROC) curve. The susceptibility map was classified into five susceptibility classes: very low, low, moderate, high, and very high. About 46% of the study area falls in high to very high susceptible classes and most of the DZs mapped (87.3%) occur in the same classes. The validation results showed satisfactory agreement between the susceptibility map and the DZs locations; over 85% of the DZs of the validation set are correctly classified falling in high and very high susceptibility areas. Also, the ROC curve had shown an area under curve (AUC) value of 0.90 which demonstrates the robustness and good reliability of the landslide susceptibility model. According to these results, we conclude that the map produced by the artificial neural network is reliable and the methodology applied in the study produced high performance, and satisfactory results, which may be useful for land planning policy. (C) 2013 Elsevier B.V. All rights reserved.