Machine learning applications for imputing missing hydroclimatic data: the case of Tormes catchment in Spain

Incomplete hydroclimatic records remain a recurrent limitation for hydrological modelling and water-resources assessment, particularly in basins with sparse station networks and extended observational gaps. This study benchmarks machine-learning (ML) approaches for imputing missing daily mean air temperature (Tmean) and precipitation at a ground-based station in the upper Tormes catchment (south-western Douro Basin, Spain), using auxiliary predictors from ERA5 grid cells, a CFSR point, and simple calendar features. For temperature, a compact comparison between a linear baseline (multiple linear regression) and a non-linear ensemble learner (eXtreme Gradient Boosting, XGBoost) shows that Tmean can be reconstructed with high skill, with only marginal gains from the more flexible model, consistent with the variable’s strong seasonal structure. For precipitation, which is intermittent and strongly skewed, six configurations are evaluated: direct regression benchmarks, a wet/dry two-stage strategy, lagged-input variants, a hybrid Seasonal AutoRegressive Integrated Moving Average (SARIMA)–XGBoost setup, and a log-transformation configuration. XGBoost with lagged auxiliary predictors provides the best generalisation performance, highlighting the value of short-term temporal dependence for daily precipitation reconstruction. SHapley Additive exPlanations (SHAP) further indicate greatest sensitivity to upstream ERA5 cells, supporting targeted feature engineering and interpretable links between regional reanalysis signals and local station infilling.