Predicción espacial de procesos de remoción en masa a partir de la construcción de un modelo híbrido de aprendizaje de máquina basado en estructuras de dependencia espacio - temporales

Abstract

Predicting areas susceptible to landslide is a complex task due to the high uncertainty in space and time. It is for this reason that in the present investigation a retropropagation red neural was constructed with space-time dependence structures (ANN-ST) made up of 11 static variables as triggers and 4 space-time clusters which are the specific spatial heterogeneity aa through the GeoSOM technique. The main objective of the work is to evaluate whether the precision in the prediction of mass removal phenomena from an ANN-ST is improved in place of the current techniques of prediction of mass removal phenomena. For the development of the research, there were a total of 1245 historical landslide records that have been presented in the city of Bogota in a period from 9/01/1996 to 4/4/2013, the landslide records were divided into two subsets of data: 70% as a training sample and the rest for the validation of the results. The ANN-ST model was validated by comparing its performance against a backward-propagating artificial red neuron without space-time dependency structures (ANN-BP). The results indicated that the ANN-ST model was 0.03 superior to the traditional RNA-BP model in terms of the area under the curve (AUC). In general terms, this research will confirm the improvement in computational precision and performance by incorporating spatio-temporal dependency structures in an artificial red neuronal, in this way this model manages to generate maps of susceptibility to higher quality mass removal processes than current techniques and helps develop policies that are focused on reducing his risk.