Flood Potential Modeling in Zarineh Rood Watershed Using Artificial Intelligence Models

Among natural disasters, flood is undoubtedly the most catastrophic hazard in the world. One of the basic strategies for reducing the damage caused by floods is to prepare a flood sensitivity map. Spatial prediction of the flooding probability using models created from spatial and historical data, which ultimately leads to the preparation of flood sensitivity maps is an appropriate solution for land management planners in different areas to prevent the occurrence of this phenomenon. In this research, in order to determine flood-prone areas, the hybrid model of adaptive neural and fuzzy inference and the metaexploratory optimization algorithm of imperial competition (ANFIS-ICA) and the hybrid model of adaptive neural and fuzzy inference and the metaexploratory optimization algorithm of particle swarm (ANFIS-PSO) are used.

The Zarine River watershed has an area of 4485 km2 and is located in the northwest of Kurdistan province between the longitude of 45°48ʹ30ʺand 46°48ʹ20ʺ east and the latitude of 35°42ʹ20ʺ and 36°23ʹ15ʺ north. The climate of the region is humid and the average annual rainfall is 480 mm. Locations of flood events were randomly divided into two groups: training (70%) and validation (30%). Various environmental factors (height, direction, slope, surface curvature, land use, lithology, rainfall, flow power index, distance from river and topographic wetness index) were selected as independent variables in the modeling and their digital layers were prepared. The ANFIS-ICA and ANFIS-PSO models were used in this research and their prediction results were evaluated based on the criterion (AUC) and the true skill statistic (TSS).

On the basis of these findings, in the validation stage, the model (ANFIS-PSO) with an AUC of 0.98 and a true skill statistic (TSS) of 0.89 had the highest accuracy. The results also showed that the factor of distance from the stream was identified as the most important environmental factor. In addition, ground slope and TWI were ranked second and third in importance, respectively.

Based on the results, the hybridization approach, which combines machine learning models and meta-exploratory optimization algorithms, improves the learning power as well as the predictive power of the model. The results of this research showed that the distance from the stream and the slope of the land are the most important factors affecting flooding. Based on the results and analysis, it can be concluded that machine learning models have a high capability for predicting flood potential. The flood potential maps prepared in this research can be very useful for managers and experts and can be used in planning flood prevention measures. Directing flood control facilities and measures in situations with a high flood potential will improve flood management from an economic and technical point of view.