Establishing the global isoscape of leaf carbon in C3 plants through the integrations of remote sensing, carbon, geographic, and physiological information

The carbon isotope composition (δ¹³C_Leaf) of C₃ plant leaves provides valuable information on the carbon-water cycle of vegetation and their responses to climate change within terrestrial ecosystems. However, global applications of δ¹³C_Leaf are hindered by a lack of global long-term spatial maps (isoscapes) that capture vegetation δ¹³C_Leaf variations. The ways in which δ¹³C_Leaf varies over time and across regions are still unknown. In this study, we collected leaf carbon isotope samples across the globe and selected the optimal predictive model from three machine learning algorithms to construct long-term annual global δ¹³C_Leaf isoscapes at a spatial resolution of 0.05° for natural C₃ plants between 2001 and 2020. We also assessed the potential of remotely sensed spectral bands, atmospheric CO₂ characteristics, geographic, and physiological information to estimate the δ¹³C_Leaf of the global C₃ plants. Our results show that the random forest (RF) algorithm can more accurately construct the δ¹³C_Leaf isoscape (R² = 0.61, Nash Sutcliffe = 0.61, RMSE = 1.21‰, MAE = 0.91‰) than the multilayer perceptron (MLP) and support vector machine (SVM). The inclusion of atmospheric CO₂ characteristics, geographical, and physiological information greatly improves prediction compared to relying only on spectral bands. Among the variables, elevation, band 3 spectral reflectance, and solar-induced chlorophyll fluorescence were the three most important variables for constructing the isoscape model, and their relative importance all exceeded 85%. The predicted isoscape revealed strong spatial heterogeneity of δ¹³C_Leaf for C₃ plants at a global scale between different continental regions, with enriched values occurring in high-altitude cold and arid regions, and depleted values occurring in warm, humid, or tropical regions. For the first time, we estimated the global depleted rate of δ¹³C_Leaf in C₃ plant (−0.0491 ± 0.07 ‰ year⁻¹ from 2001 to 2020). Over the past two decades, 86.49% and 1.00% of global grid cells showed more depleted and more enriched trends in δ¹³C_Leaf of C₃ plants, respectively. Our results demonstrated the potential for establishing isoscapes by combining remote sensing, atmospheric CO₂ characteristics, physiological, and geographic variables using the RF machine learning algorithm. The isoscape products generated in this study are valuable for assessing carbon-water coupling of terrestrial ecosystems and for improving land surface models.