Filling Gaps in Daily Precipitation Series Using Regression and Machine Learning in Inter-Andean Watersheds

Abstract

As precipitation is a fundamental component of the global hydrological cycle that governs water resource distribution, the understanding of its temporal and spatial behavior is of great interest, and exact estimates of it are crucial in multiple lines of research. Meteorological data provide input for hydroclimatic models and predictions, which generally lack complete series. Many studies have addressed techniques to fill gaps in precipitation series at annual and monthly scales, but few have provided results at a daily scale due to the complexity of orographic characteristics and in some cases the non-linearity of precipitation. The objective of this study was to assess different methods of filling gaps in daily precipitation data using regression model (RM) and machine learning (ML) techniques. RM included linear regression (LRM) and multiple regression (MRM) algorithms, while ML included multiple regression algorithms (ML-MRM), K-nearest neighbors (ML-KNN), gradient boosting trees (ML-GBT), and random forest (ML-RF). This study covered the Malas, Omas, and Cañete River (MOC) watersheds, which are located on the Pacific Slope of central Peru, and a nineteen-year period of records (2001–2019). To assess model performance, different statistical metrics were applied. The results showed that the optimized machine learning (OML) models presented the least variability in estimation errors and the best approximation of the actual data from the study zone. In addition, this investigation shows that ML interprets and analyzes non-linear relationships between rain gauges at a daily scale and can be used as an efficient method of filling gaps in daily precipitation series.

Keywords

• precipitation gap filling
• regression
• machine learning
• standard normal homogeneity test
• K-nearest neighbors
• gradient boosting tree