Comparing performance of random forest and adaptive neuro-fuzzy inference system data mining models for flood susceptibility mapping

Abstract

Flood is one of the important destructive natural disasters in the world. Therefore, preparing flood susceptibility map is necessary for flood management and mitigation in a region. This research was planned to compare the performance of frequency ratio (FR), adaptive neuro-fuzzy inference system (ANFIS), and random forest (RF) models for flood susceptibility mapping (FSM) in the Gilan Province, Iran. First, a geospatial database included 220 flood locations and eleven effective flood factors (slope angle, aspect, altitude, distance from rivers, drainage density, lithology, land use, topographic wetness index (TWI), and stream power index (SPI)) were produced. According to flood locations, 30–70% of them were used for training and validation of the models, respectively. Afterward, the mean of Gini reduction was used to determine the priority of effective flood factors. Finally, the receiver operating characteristic (ROC) curve, area under the curve (AUC), was used to evauate and compare the performance of the models. The validation results of the models show that FR, ANFIS, and RF models had 68.6, 63.9, and 71.3% accuracy, respectively. In addition, distance from rivers, altitude, and drainage density was the most important factor for FSM in the study area. The finding of the current research proved a reasonable prediction performance for the models. Therefore, these models can be proposed for preparing FSM in similar climatic and physiographic areas and flood susceptibility maps can be used to manage floodplains in the study area.