Combination of geostatistical and artificial intelligence methods for predicting spatiotemporal water level in the Hadishahr plain

The Hadishahr plain, with 56 km2 area, is located in the northwest of the East Azarbaijan province. Due to the intensive withdrawal of groundwater from this area in the recent years, the water level has been declined significantly. In order to find the best method for managing the groundwater resources of the study area efficiently, artificial neural networks and fuzzy methods were utilized to model and predict the temporal and spatial variations of the groundwater level. Firstly, the central piezometer was used for modeling artificial neural network and determining the best algorithm structure. The results show that the forward neural network with the LevenbergÜMarkvrat (LM) algorithm with 1, 2 and 3 order structure is the best method in this plain, respectively. Afterward, the selected piezometers of the plain were classified with the hierarchical clustering (HCA) methods and each piezometers batch was modeled with the Sugeno fuzzy technique. The results were compared using the statistical parameters such as RMSE and R2. In this study, monthly data of rainfall, evaporation, and groundwater level were used as inputs to the model. The results show that the fuzzy and neural network techniques using feed forward neural network with the Levenberg-Markvrat (LM) algorithm achieves the best answer. Thus the neural kriging and neural cokriging method were used, for predicting the temporal and spatial variations of groundwater level.