Machine learning gap-filling of hydrometeorological time series in south-central Chile: Streamflow, temperature and precipitation

Study Region This study was conducted in south-central Chile, between 34°S and 41°S, encompassing a hydroclimatically diverse region extending from the Pacific coast to the Andes Mountains. Study Focus We evaluated the non-parametric MissForest algorithm for gap-filling daily hydrometeorological time series from 1980 to 2022 across 57 streamflow, 19 temperature, and 97 precipitation stations with < 20% missing data. Performance was assessed under point and continuous-gap validation schemes using the modified Kling–Gupta Efficiency (KGE), root mean square error (RMSE), and, for precipitation, wet-day occurrence diagnostics. For meteorological variables, ERA5-Land was also evaluated both as an auxiliary predictor within the MissForest framework and as input to a linear regression baseline. New Hydrological Insights for the Region MissForest improved moderately incomplete hydrometeorological records; however, performance depended on variable type and gap structure. Point gaps were reconstructed more accurately than continuous gaps, with median observation-based KGE decreasing from 0.89 ± 0.11–0.66 ± 0.43. Temperature exhibited the most stable performance, whereas precipitation remained sensitive to wet-day occurrence and event timing. Streamflow showed the greatest variability under continuous validation, particularly during high-flow conditions. ERA5-Land enhanced meteorological gap-filling primarily under point validation, yielding near-optimal KGE values for temperature and precipitation while reducing false wet days. Overall, MissForest is most suitable for short, isolated gaps and general daily variability, whereas caution is needed for extremes events and long continuous gaps.