Upscaling Net Ecosystem Exchange on the Qinghai‐Tibet Plateau Using Ensemble Machine Learning at High Spatial‐Temporal Resolution

Abstract Machine learning models are effective in predicting the spatiotemporal dynamics of net ecosystem exchange (NEE), a critical component in carbon cycling. However, existing machine learning‐based NEE data sets for the Qinghai‐Tibet Plateau (QTP), recognized as the Earth's Third Pole, have been constrained by coarse resolution or have focused solely on grassland ecosystems. This study addressed this gap by developing a high‐resolution NEE data set (500 m, 8‐day intervals) for the QTP, utilizing in situ measurements from 36 plateau sites and 80 forest sites outside it, combined with remote sensing and meteorological data. Various combinations of environmental variables and machine learning algorithms were applied to assess prediction uncertainties. The results indicated that the mean annual NEE for the QTP was −147.63 ± 15.03 Tg C yr −1 , with NEE becoming more negative at a rate of 0.96 Tg C yr −1 from 2002 to 2022. Meadows contributed most (38.1%) to total NEE, followed by shrubs (26.7%), forests (23%), steppes (11.8%), and wetlands (0.5%). Compared to previous studies, this research utilized more eddy covariance (EC) observations and covered all QTP ecosystem types except deserts. Additionally, this research revealed that the selection of environmental variables significantly impacted prediction results, and substantial differences among various meteorological data sets on the QTP introduced considerable prediction uncertainty. The updated NEE prediction data set for the QTP provided essential data for advancing regional carbon cycle research.