Logistic Regression and Neural Networks Applied for Probabilistic Forecasts of Flooding in Curitiba, Pr (PAP014828)

Código

PAP014828

Autores

Marciel Lohmann, Leonardo José Cordeiro Santos

Tema

Urban Floods

Resumo

This research aims to study through the use of logistic regression and neural network, the characteristics related to rainfall patterns in Curitiba, in order to establish the relationship between rainfall and flooding, using the integration of hydrometeorological data. To achieve the proposed aims, two models were built based on logistic regression and Kohonen neural network type (Self Organizing Map (SOM)), to predict the probability of flooding. The two methods were compared and evaluated their performance through the ROC (Receiver Operating Characteristic) curve as well as from diagrams of reliability, discrimination and refinement. For the construction of the models were used rainfall data estimated from the integration of meteorological radar, satellite and rain gauges data, using the analysis for statistical purposes (ANOBES) method. In adition, data records of flooding were used as a reference. These data were supplied by the Municipal Civil Defense and compiled by IPPUC (Institute for Urban Research and Planning of Curitiba).Rainfall estimates were used to calculate cumulative average rain of 6 hours in a basis of 4 days, using the method of Thiessen and Squared Inverse Distance. These the two methods were compared to see which has better results for data generation to be used as models' data input. Regarding the performance of the two methods used to construct the models, it was found that the SOM (Self Organizing Map) has superior performance when compared with the logistic regression, either for calibration and verification. Results generated by the SOM, indicated that it is possible to define the main rainfall patterns responsible for triggering flooding in Curitiba and also estimate the expected number of floods (NEA), for each watershed. From this perspective, this work has as a first innovation the use of artificial intelligence (AI) tools designed for the recognition of rainfall patterns that can cause flooding. Regarding the expected number of floods, the innovation refers to the spatial distribution of floods based on historical occurrences. As a proposal, it is suggested that the results generated in this work should be used in a flooding alert system for Curitiba, and that the information and data generated can be used by Civil Defense to increase the resilience of the population and mitigate potential impacts of flooding.