Source codes used to train models for Joint control of precipitation and CO2 on global long-term patterns of plant nitrogen availability

Plant nitrogen availability critically limits plant growth and thus constrains terrestrial carbon sequestration. Growing evidence points to a recent decline in plant nitrogen availability, yet the global consistency of this trend and its underlying drivers remain unclear. Here, we reconstruct the spatiotemporal trajectory of plant nitrogen availability (1980–2020) using four machine-learning algorithms applied to a global database of 37,268 foliar stable nitrogen isotopes measurements and ancillary climate data. We first find high spatial heterogeneity in foliar stable nitrogen isotopes, primarily linked to mean annual temperature. Then, we discover that foliar stable nitrogen isotopes declined primarily during the 1980s, and subsequently remained stable across 44% of land areas. Moreover, the dominant driver in temporal variations of foliar stable nitrogen isotopes shifts from CO2 (1980–1988) to precipitation (1989–2020). These patterns collectively reveal divergent plant nitrogen availability trajectories and highlight the increasing role of precipitation in shaping terrestrial nitrogen cycles over the past decades.Note:Please ensure that you have more than 7 GB of free disk space before unzipping the files. Each main folder includes a corresponding README file. We also provide the standalone code and the primary datasets, which can be downloaded and opened directly.Run the three Python scripts (All.py, All_holdout.py, Three_period.py) in sequence (the required Python packages need to be installed beforehand). The scripts will generate the corresponding folders containing model files, predictions, spatial weight tables, and other related outputs. After the models are trained, use them to generate global predictions at each grid point (Predict xxx.py). The prediction results are stored in the Predict_N15 folder.Subsequent analyses should be performed in R by running the provided R scripts. Because model training in Python takes a long time, the compressed package includes the pre-trained Python model outputs (in the models subfolder of each directory), which can be loaded and used directly. The results are fully reproducible because random seeds were set.If you encounter any errors, please make sure that the working directory is set correctly. For any questions, please contact: [e-mail] (Songbo Tang).

植物氮有效性（plant nitrogen availability）严重限制植物生长，进而制约陆地碳固存。越来越多的证据表明近期植物氮有效性有所下降，但这一趋势的全球一致性及其潜在驱动机制仍不明确。本文基于包含37268条叶片稳定氮同位素（foliar stable nitrogen isotopes）实测数据的全球数据库与辅助气候数据，结合四种机器学习算法，重建了1980—2020年植物氮有效性的时空演变轨迹。研究首先发现叶片稳定氮同位素存在显著的空间异质性，其主导影响因子为年平均气温。随后发现，叶片稳定氮同位素在20世纪80年代呈下降趋势，之后在44%的陆地区域保持稳定。此外，叶片稳定氮同位素时间变化的主导驱动因子从1980—1988年的二氧化碳（CO₂）转变为1989—2020年的降水。这些模式共同揭示了植物氮有效性轨迹的分异性，并强调了近几十年来降水在调控陆地氮循环中的作用日益增强。
注意：解压文件前请确保磁盘剩余空间不少于7GB。每个主文件夹均包含对应的README文件。本文同时提供独立代码与核心数据集，可直接下载并打开。
请按顺序运行三个Python脚本（All.py、All_holdout.py、Three_period.py），需提前安装所需Python依赖包。脚本运行后将生成包含模型文件、预测结果、空间权重表及其他相关输出的对应文件夹。模型训练完成后，可通过运行Predict xxx.py生成每个网格点的全球预测结果，预测结果将存储于Predict_N15文件夹中。
后续分析需通过R语言运行提供的R脚本完成。由于Python模型训练耗时较长，压缩包已包含各目录models子文件夹中的预训练Python模型输出，可直接加载使用。本研究结果具备完全可复现性，因所有随机种子均已预先设置。
若遇到任何错误，请确保已正确设置工作目录。如有任何疑问，请联系：[e-mail]（宋博唐）。