Data: 25/07/2018 à 27/07/2018
Local: Porto Alegre-RS
Mais informações: https://www.abrh.org.br/iend
COMITÊ DE REDES NEURAIS ARTIFICIAIS APLICADAS AO MAPEAMENTO DE RISCOS DE DESLIZAMENTOS
Lucimara Bragagnolo, ROBERTO VALMIR DA SILVA, Jose Mario Vicensi Grzybowski
3 - Previsão e mapeamento de movimentos de massa
Mapping landslide susceptibility areas is a fundamental task towards the adoption of preventive measures and adequate strategies of evacuation. Several actions have been taken by public agencies in Brazil toward the development of preventive measures aimed at minimizing risks to human life and infra-structure. However, much effort is still to be done in order to make landslide susceptibility mapping and management available countrywide. This occurs mainly due to two reasons. First, except for a very small number of its cities, Brazil still lacks the culture of maintaining detailed inventories of landslide events to compose databases of scientific value for landslide susceptibility mapping. Second, the deficit of technical and scientific expertise on landslides mapping at civil defense offices. As a means to assist and enhance the process of generating reliable landslide susceptibility maps, the scientific community has turned its attention to machine learning techniques, such as artificial neural networks. In this work, we test the performance of an Artificial Neural Network Ensemble (ANNE) in the task of generating landslide susceptibility maps. As it turns out, scientific studies have indicated that artificial neural network ensembles have the potential to outperform individual neural networks. The advantage of ANNEs over traditional methods is that they combine the cognitive capacity of several individual neural networks with different learning backgrounds, thus enhancing the potential to give accurate results. To that end, we train, validate and test about 42 thousand artificial neural networks using data from the landslide inventory of Porto Alegre, Rio Grande do Sul state, Brazil, and then select those with best performance to compose the ensemble. In order to evaluate the susceptibility maps generated by the ANNE, we use the landslide susceptibility map produced by the Geological Survey of Brasil (CPRM) agency. The results show that the landslide susceptibility maps generated by the ANNE feature high level of quantitative and qualitative agreement with the map by CPRM, thus indicating that ANNEs are able to capture the essential features of the phenomenon and thus may be an effective tool to assist civil defense professionals in the task of producing such maps.