克鲁伦河流域生态环境安全在中蒙两国受到越来越多关注，掌握流域土壤全氮（STN）和土壤全磷（STP）含量对于准确估算流域面源污染（NPS）负荷、研究流域资源环境状况与可持续发展具有重要意义。传统采样方法在获取大范围的STN和STP含量时耗时耗力、STN与STP存在空间异质性、STN和STP与辅助变量间的关系也存在空间异质性等。单一的全局模型无法拟合复杂的异质性关系，而局部建模方法难以克服维度灾难问题，因此本文引入了两点机器学习（TPML）方法。该方法首先基于点对差异建立全局模型，然后基于全局模型的预测差异构建局部模型，能够将样本量从n扩充至n²，可利用有限的采样点数据实现高精度大范围的STN和STP含量预测。本文结合地形、气候、土壤属性、植被及空间位置等共18个辅助变量，采用TPML方法，制作了流域STN和STP含量分布数据集。并基于十折交叉验证方法证实了TPML方法相较于普通克里格（OK）方法，预测精度提高超过10%。TPML方法预测STN含量的平均绝对误差（MAE）均值和平均均方根误差（RMSE）分别为0.309%、0.456%，随机森林（RF）、反距离加权（IDW）与OK方法预测STN含量的平均MAE分别为0.329%、0.247%与1.864%，平均RMSE分别为0.468%、0.387%、1.976%。TPML方法预测STP含量的平均MAE和平均RMSE分别为0.640%和0.861%，RF、IDW与OK方法预测STP含量的平均MAE分别为0.643%、0.396%与1.357%，平均RMSE分别为0.862%、0.523%与1.651%。

The ecological and environmental security of the Kherlen River Basin has attracted more and more attention in China and Mongolia. It is of great significance to investigate the contents of soil total nitrogen (STN) and soil total phosphorus (STP) in the basin for accurately estimating the load of non-point sources (NPS) and studying the state of resources and environment and sustainable development. It is time-consuming and labor-intensive to obtain a wide range of STN and STP contents using traditional sampling methods, while STN and STP not only have spatial heterogeneity, but also have heterogeneity in their relationships with auxiliary variables. Moreover, a single global model cannot fit complex heterogeneous relationships, and it is difficult for the local modeling method to overcome dimensional disaster problems. Therefore, the two-point machine learning (TPML) method is introduced in this paper. The TPML method first establishes a global model based on the difference of paired points, and then constructs a local model based on the prediction difference of the global model. It can expand the sample size from n to n², achieving the prediction of high-precision and large-scale STN and STP contents using limited sampling points. Based on 18 auxiliary variables of topography, climate, soil properties, vegetation and spatial location, etc, the study produced the distribution dataset of STN and STP contents in the basin using the TPML method. Futhermore, using the ten-fold cross-validation method, the study confirmed that the prediction accuracy of TPML model is more than 10% higher than that of Ordinary Kriging (OK) model. The mean absolute deviation (MAE) and mean root mean squared error (RMSE) of STN content predicted by the TPML method are 0.309% and 0.456% respectively. The mean MAE of STN content predicted by random forest (RF), inverse distance weighted (IDW) and OK methods is 0.329%, 0.247% and 1.864%, and the mean RMSE is 0.468%, 0.387% and 1.976%, respectively. The mean MAE and mean RMSE of STP content predicted by TPML method are 0.640% and 0.861%. The mean MAE of STP content predicted by RF, IDW and OK methods is 0.643%, 0.396% and 1.357%, and the mean RMSE is 0.862%, 0.523% and 1.651%, respectively.