Integrating flux footprint, random forest, and SHAP for interpretable hourly carbon flux upscaling: Development and application in the Qilian mountains watershed

Accurate estimation of regional carbon fluxes is essential for understanding the carbon cycle. However, conventional eddy covariance (EC) upscaling methods are often limited to daily or monthly resolutions, overlooking critical sub-daily dynamics. Furthermore, many machine learning-based frameworks lack interpretability, which obscures the environmental drivers of carbon flux variability. To address these gaps, this study introduces a novel integrated framework combining flux footprint modeling, Random Forest (RF), and SHapley Additive exPlanations (SHAP) algorithms. This approach upscale gross primary productivity (GPP) and ecosystem respiration (Reco) during the growing season using data from nine EC sites within the Qilian Mountains (QLM). Compared to estimating net ecosystem exchange (NEE) directly, the decoupling approach (separating NEE into GPP and Reco) significantly improved accuracy. SHAP analysis attributed this to their divergent drivers: GPP is predominantly regulated by radiation, while Reco is mainly controlled by air temperature and leaf area index relative fraction. Leveraging this model and Third Pole Meteorological Forcing Dataset (TPMFD) data, an hourly dataset of GPP and Reco with 1/30° spatial resolution spanning 1982–2020 over the QLM watershed is generated. Based on this dataset, we found that regional GPP, Reco, and NEE all exhibited significant increasing trends during the growing season, with growth rates of 0.05, 0.41 and 0.36 g C m −2 month −1 , respectively. Moreover, counties such as Tianjun (TJ) and Haiyan (HaiY) should therefore be designated as key areas for future ecological management. Our framework provides critical data support for improving regional carbon cycle models and enhancing predictions in complex terrain ecosystems.