Fusing Enhanced Flux Measurements and Multi-Source Satellite Observations to Improve GPP Estimation for the Qinghai–Tibet Plateau Based on AutoML Techniques

The Qinghai–Tibet Plateau (QTP) plays a crucial role in the terrestrial carbon cycle, but the gross primary productivity (GPP) estimates for the region remain highly uncertain due to limited flux observations and modeling challenges. Here, we integrated 65.2 site years of eddy covariance data from 19 flux sites with multi-source remote sensing observations to develop a data driven GPP model for the QTP. Eleven machine learning algorithms from two automated machine learning (AutoML) platforms, H2O AutoML and FLAML, were evaluated to construct an ensemble model named AutoML. The model showed strong performance at site-level across alpine meadow, steppe, wetland, and shrub ecosystems, achieving R2 up to 0.95 and RMSE as low as 0.42 g C m−2 d−1. By validating extracted site-level GPP values from the upscaling GPP datasets against with flux observations, AutoML-GPP demonstrates overall superior or equivalent performance over global GPP products (FLUXCOM X-base, GOSIF, and FluxSat). Regional upscaling estimated a mean annual total GPP of 374.20 Tg C yr−1 from 2002 to 2018, with a slight upward trend of 0.08 Tg C yr−1. Spatially, higher GPP occurred mainly in the eastern QTP, with anomalies linked to climate extremes in 2008, 2010, and 2015. AutoML-GPP effectively captures climate-induced interannual anomalies in the QTP’s GPP, coinciding with GOSIF-GPP and FluxSat GPP, and outperforming the recent released well-known global upscaling flux dataset FLUXCOM X-base. This study provides improved GPP estimation for the QTP, offering new insights into carbon cycling and climate–vegetation interactions.