Predição de áreas de risco de deslizamento usando a rede neural artificial como critério de decisão

Abstract

Landslides all over the world have been causing big losses both economically and in lives. With the use of geotechnologies associated to the potential of the artificial intelligence is believed that it is possible to prevent or mitigate the impacts caused by such event. The objective is to test the ANN technique to verify if it is able to learn the existent patterns in the risk areas, and if it is possible to apply this learning to verify areas with landslide potential, both in the same area as in another one with some different characteristics. For this, was researched with the competent organs of the estate for the occurrence of landslides and its spatial location, information about the use and occupancy of the land, spectral image, land class, slope, hydrography, direction and length of the flow in the same place of the events occurrence. The information was introduced in the input layer of the ANN. To maximize the ANN learning rate, different architectures were verified empirically. The best result for the learning corresponded to 91.66% of success that was passed to the simulation. The simulation was applied in a part of the area that was used for training and for the city of Dois Vizinhos, that corresponds to an area with different characteristics from the training area. The ANN found a model for the patterns of the risk areas presented and was verified that these patterns also happens in different regions from those where the ANN architecture was trained because they have some similarities in their spectral or spatial data. Therefore, these results can help the managing organs in the decision making regarding to the prevention in areas with potential landslide risk.