Identification of torrential valleys using GIS and a novel hybrid integration of artificial intelligence, machine learning and bivariate statistics

Abstract

The detection of zones exposed to flash-flood and also the torrential valleys on which flash-floods are propagated, represents a crucial measure intended to eliminate the issues generated by these phenomena. In this paper, in order to locate the regions prone to runoff occurrence, a number of 4 hybrid models were employed: Naïve Bayes – Certainty Factor (NB-CF), Naïve Bayes-Evidential Belief Function (NB-EBF), Multilayer Perceptron – Certainty Factor (MLP - CF) and Multilayer Perceptron – Evidential Belief Function (MLP - EBF). The first step of the methodology consisted in the mapping of the territories with torrential relief microforms. These areas were split into training sample (70%) and validating sample (30%). By mean of Information Gain statistic method, 10 flash-flood causal variables were chosen to construct the models and to compute the Flash-Flood Potential Index values. In order to calculate the Flash Flood Potential Index (FFPI) values, the CF and EBF coefficients were determined and, subsequently, were incorporated into the NB and MLP models. The results of the four hybrid models were validated by using two methods: i) relative distribution of torrential pixels within FFPI classes; ii) Receiver Operating Characteristic (ROC Curve). Since the MLP-CF model achieved the best performance, its results have been further used in a Flow Accumulation procedure for identifying torrential valleys within the research territory. Valleys with a high and very high torrentiality degree have a total a length of 1304 km. These valleys were mainly developed in the North-Western zone of Prahova river basin.