Comitês de redes neurais artificiais aplicados à avaliação de áreas suscetíveis a deslizamentos

Abstract

Mapping areas susceptible to landslides is regarded as fundamental since it allows preventive measures, strategies of action and evacuation to be carried out advance. This study proposes the application of an Artificial Neural Network Ensemble (ANNE) for the mapping of landslide susceptibility areas. To that end, the identification of susceptible areas was performed on the basis of landslide inventory databases containing the location information and seven geomorphologic parameters: lithology, elevation, aspect, slope, topographic moisture index, land use and vertical curvature. The training, validation and testing of the neural networks was performed with 14 different architectures (7-2-1 to 7-15-1) by resampling the training records and varying the initial conditions of the neural network parameters. The composition of the ensembles was made by selecting the individual networks that obtained best performances. The proposed methodology was applied to the generation of landslides susceptibility maps for two Brazilian capitals, Porto Alegre and Rio de Janeiro. The results were compared with those obtained by means of multiple linear regression and with landslides susceptibility maps published by the Geological Survey of Brazil and City Hall of Rio de Janeiro cities. The sensitivity assessment to the input parameters indicated that the parameters that exerted the greatest influence on the prediction of landslides were vertical curvature, topographic wetness index and slope. In addition, the ANNEs showed better results when compared to multiple regression, as well as a fair level of qualitative and quantitative agreement with the maps published by official government organs. These facts indicate that the essembles was able to identify the influence of model sensitivity to the input variables and their interaction in the occurrence of landslides. As the main result, we show that the susceptibility maps generated by the ANNE feature higher accuracy than those published by the Brazilian Geological Survey and City Hall Rio de Janeiro. This indicates that the proposed methodology can be an effective tool to assist the development of reliabe landslide susceptibility maps.