Improving the Quality of the Surface Heat Flux Data over the Tibetan Plateau by Using an Optuna–CatBoost–Shapley Additive exPlanation Method

Accurate and reliable latent heat (LH) and sensible heat (SH) flux data are crucial for determining energy exchange on the Tibetan Plateau (TP) and for studying land–atmosphere interactions and their impacts on the climate. Improving the accuracy of heat flux datasets over the TP and comprehensively assessing their reliability and uncertainty, especially in data-sparse regions, remain insufficiently explored. This study utilized the Optuna–categorical boosting (CatBoost) model to fuse daily TP heat flux datasets from 2000 to 2022, incorporating five reanalysis datasets (CRA, ERA5, MERRA2, NCEP, and JRA55) and environmental variables. Shapley Additive exPlanation (SHAP) analysis and the generalized three-cornered-hat (TCH) method were applied for result interpretation and spatiotemporal uncertainty quantification, respectively. The results showed that the fusion method can significantly increase the accuracy of the data, i.e., the Pearson CC can be increased from 0.74 to 0.85 for the LH and from 0.52 to 0.70 for the SH. Moreover, it can significantly decrease the uncertainty of the fused LH data and decrease the uncertainty of the fused SH data in general (the average relative uncertainty (RU) was 14.84% for LH and 27.88% for SH). Dewpoint and LST are found to be the most important environmental factors influencing the LH and SH simulations (with average contribution rates of 12.9% and 15.6%, respectively). When the dewpoint exceeds 266.56 K and the LST exceeds 13.79 °C, their influence on LH and SH shifts from negative to positive. Wetlands and glaciers are the land cover types with the highest uncertainty in LH and SH data, respectively. The Qaidam Basin shows the highest mean uncertainties among all the climatic zones. The machine learning-based fusion method developed by this paper not only effectively enhances the quality of fused data but also provides valuable insights for future site expansion and further data quality improvement through its uncertainty analysis findings.