Evaluating the Performance of ANN and ANFIS Models for Spatial Precipitation Prediction in Complex Terrain: a Case Study in Central Anatolia

Abstract

Accurate spatial precipitation prediction is essential for hydrological modelling and climate risk management, particularly in regions with complex topography. This study evaluates and compares the performance of machine learning (ML) models, specifically Artificial Neural Networks (ANN) and the Adaptive Neuro-Fuzzy Inference System (ANFIS), with traditional geostatistical methods, including Kriging, Co-Kriging (COK), and Regression Kriging (RK), in Central Anatolia, T & uuml;rkiye. The models were trained and tested using data from 193 meteorological stations and 17 environmental and spatial predictors, such as elevation, humidity, pressure, and distance to the sea. The best-performing ANN model achieved a test Root Mean Square Error (RMSE) of 9.6 mm, Percent Bias (PBIAS) of - 1.45%, Nash-Sutcliffe Efficiency (NSE) of 0.577, and Willmott's Index of Agreement (WI) of 0.869. ANFIS also outperformed the geostatistical methods, with its best configuration yielding a test RMSE of 12.7 mm and an extrapolation error of 15.6 mm at the Erciyes Mountain Point (EMP). In contrast, Kriging-based methods exhibited large errors at high-altitude locations, particularly at EMP, where precipitation reached 85 mm but was underestimated by more than 60 mm. These findings demonstrate that ML models can effectively capture nonlinear spatial patterns and provide superior extrapolation performance in ungauged, mountainous regions. A ML-based framework is proposed, emphasizing local error evaluation and point-specific modelling to improve the reliability of spatial precipitation mapping.