Examining the spatial estimation success of artificial intelligence methods for precipitation maps

The accuracy of spatial estimations is crucial for performing accurate analyses. In this study, the spatial prediction performance of artificial intelligence methods for the estimation of spatial distributions of precipitation has been examined. Using the 2016 average monthly precipitation magnitudes belonging to 193 meteorological stations located in the Central Anatolia Region of Turkey as an example, Artificial Neural Network (ANN) and the Adaptive Neuro-Fuzzy Inference System (ANFIS) methods have been utilized for spatial predictions. As part of the thesis study, a method called Geostatistical and Heuristic Spatial Prediction of Regionalized Variables (Geo-HUSREV), which combines geostatistical methods with heuristic algorithms, has been developed. Root mean square errors (RMSE) of test points, not used in applications, and prediction errors observed at 11 characteristic test points with different features have been examined to evaluate the prediction performances. 17 secondary variables associated with precipitation have been used as inputs in predictions. All input combinations have been used in numerous applications by using 1, 2, 3, 4, and 5 inputs together for the predictions. Different parameters of the ANN and ANFIS applications have been examined, and their impacts on the predictions have been evaluated based on the findings. The models of the best applications selected from ANN and ANFIS applications have been retrained using the Particle Swarm Optimization PSO and Genetic Algorithm (GA) methods with the same input variables, and the success of ANN-PSO, ANN-GA, ANFIS-PSO, ANFIS-GA applications, kriging, COK, RK, and Geo-HUSREV methods have been examined. As a result, the most accurate predictions for the study area have been obtained with the ANN method. After the ANN method, ANFIS and the Geo-HUSREV methods, in turn, provided a better prediction performance than the kriging, COK, and RK methods. Artificial intelligence methods have been very successful at the extrapolation test point where the prediction of maximum precipitation has been attempted. In the applications with the most accurate prediction values, the variables have been mostly humidity, pressure, distance to sea, temperature, distance to rivers. 16 different prediction maps have been produced, and significant differences have been observed in the textures of the maps. The thesis study concluded that artificial intelligence methods can be successfully applied for spatial predictions, and it is important to examine the findings of these methods in addition to geostatistical analysis studies in spatial prediction studies.