Scale-dependent drivers of water use efficiency across China: integrating stable isotopes, remote sensing, and machine learning

Water use efficiency (WUE) serves as a crucial metric for terrestrial carbon–water coupling, yet systematic gaps persist in understanding the spatial patterns and drivers of leaf-level intrinsic WUE (iWUE) versus ecosystem-scale WUE (WUE_Eco). Combining machine learning with 1,446 leaf δ¹³C_p records, we investigated the spatial heterogeneity and main drivers of iWUE and WUE_Eco across different life forms and climate zones in China. Results showed that inverse spatial patterns, where iWUE peaked in arid northwestern grasslands (60.46 μmol mol⁻¹). In contrast, WUE_Eco exhibited maxima in humid southeastern forests (1.82 g C/kg H₂O). Hierarchical partitioning and structural equation modeling revealed that elevation indirectly influenced iWUE (17.72 %) and WUE_Eco (25.64 %) through its modification of climatic conditions. Vegetation factors (e.g., leaf area index) and climatic factors (e.g., relative humidity) emerged as key drivers of iWUE (24.06 %) and WUE_Eco (15.31 %), primarily through their regulation of photosynthesis–transpiration coupling processes. Among four machine learning models, Random Forest has the best performance in iWUE prediction (R² = 0.73, NRMSE = 0.122, MBE =  − 0.078), providing a high-resolution national iWUE dataset. This study highlights the importance of scale in understanding carbon–water interactions and provides a valuable reference for water resource management under climate change.