Interpretable machine learning unveils threshold responses and spatial patterns of global soil respiration

Soil respiration (Rs) represents the largest carbon flux from land to the atmosphere and is important for assessing the terrestrial carbon cycle and studying climate change. In this study, we propose an interpretable machine learning prediction of global Rs (IMPGRs) based on explainable artificial intelligence (XAI) technology to interpret tree-integrated global Rs prediction models, explore the factors driving global Rs responses, and predict their distribution patterns. The IMPGRs model showed superior performance in predicting Rs, capturing two meaningful non-linear (‘J’- and ‘U’-type) relationships between Rs and environmental variables. We found that a soil temperature of 20.9°C represented an important threshold for the ‘thermal adaptation’ of Rs. Moreover, this phenomenon varied significantly across climatic zones and ecosystems, and the threshold was positively correlated with precipitation. The response of global Rs to the leaf area index (LAI) was not a simple positive correlation, and contrasting results were observed both inside and outside the Tropic of Cancer. Global Rs values (688.43 g C m⁻² year⁻¹) and their distribution were predicted using IMPGRs, with forest soils releasing the most carbon dioxide (CO₂; 42.84 Pg C year⁻¹) and accounting for 45.7 % of the global Rs. Additionally, we found significant biases in the annual Rs calculated by area weighting based on climate and ecosystem classifications because these factors characterise spatial heterogeneity differently. Such dynamics should be considered when modelling global Rs and analysing the results because they can help improve the estimation accuracy of global Rs prediction models.