Performance Comparison of Interpolation Methods for Precipitation Data Reconstruction: Kriging, Polynomial Interpolation, Cubic Splines, and Neural Networks

Abstract

Reliable and complete precipitation data are essential for hydrological modeling and water resource management. However, gaps in records due to sensor failures, human error, or limited station coverage can compromise analysis quality. This study evaluates the performance of four interpolation techniques: artificial neural networks, mul-tiple linear regression, ordinary kriging, and cubic splines for estimating missing daily precipitation values in the Cedar River basin, a mountainous region in Washington State, USA. Prior to interpolation, data quality control was applied using double mass curve analysis and Pettitt’s test. Performance was assessed using RMSE, MAE, Pearson correlation, and Nash–Sutcliffe efficiency. Results indicate that artificial neural net-works provided the most accurate estimates (RMSE = 2.64 mm; Correlation coefficient: = 0.98 and Nash–Sutcliffe efficiency= 0.96), followed by cubic splines, kriging, and multiple regression. Neural networks effectively captured nonlinear patterns in precipitation but required specialized knowledge and more computational resources. Kriging offered a robust and simpler alternative when spatial structure was well-defined, while regression worked well when station correlations were high. Cubic splines performed poorly under high temporal variability. The findings suggest that neural networks are best suited for complex conditions, but traditional methods re-main valid in operational settings with limited resources.