Estimation of monthly rainfall missing data in Southwestern Colombia: comparing different methods

Abstract

ABSTRACT Historical rainfall records are relevant in hydrometeorological studies because they provide information on the spatial features, frequency, and amount of precipitated water in a specific place, therefore, it is essential to make an adequate estimation of missing data. This study evaluated four methods for estimating missing monthly rainfall data at 46-gauge stations in southwestern Colombia covering 1983-2019. The performance of the Normal Ratio (NR), Principal Components Regression (PCR), Principal Least Square Regression (PLSR), and Artificial Neural Networks (ANN) methods were compared using three standardized error metrics: Root Mean Square Error (RMSE), Percent BIAS (PBIAS), and Mean Absolute Error (MAE). The results generally showed a better performance of the nonlinear ANN method. Regarding the linear methods, the best performance was registered by the PLSR, followed by the PCR. The results suggest the applicability of the ANN method in regions with a low density of stations and a high percentage of missing data, such as southwestern Colombia.