Spatio-temporal variations in global ecosystem respiration and its drivers over the last two decades

• We used a machine learning-based method for temporal dynamics to estimate global ER from 2000 to 2020. • There was an increasing trend in global ER over the past two decades, with significant spatial variations. • Across most global regions, carbon substrates dominate the interannual variability of ER. Ecosystem respiration (ER) is a fundamental component of the global carbon cycle, and its accurate estimation is essential for understanding global carbon budgets and climate change. In recent years, machine learning algorithms have been increasingly employed for ER estimation; however, existing ER data products derived from such methods exhibit considerable uncertainties in temporal dynamics. In this study, we reconstructed the time series of ER measurements using eddy covariance flux observations and corresponding driver data, developed ER prediction models annually, and produced a year-by-year global ER dataset (RCS_ER) at 0.05° spatial resolution during 2000-2020. Comparative analyses show that our product consistently outperforms existing global ER datasets in both its overall and year-by-year simulation accuracy. While the spatial distribution of RCS_ER is highly consistent with previous products, it reveals stronger interannual variability and an overall increasing trend over the past two decades, with significant spatial heterogeneity in trend patterns. Partial correlation analysis further indicated that carbon substrates is the dominant driver of ER interannual variability (IAV) in most global regions, while precipitation exerts a more substantial influence than temperature in modulating regional ER dynamics. These findings provide an important data foundation for studying the spatio-temporal variations of terrestrial carbon fluxes and improve our understanding of global carbon cycle processes.