基于机器学习的长江流域农田氮径流流失负荷估算

定量解析长江流域农田氮径流流失特征是实现长江及其河口氮污染有效控制的关键科学基础.基于收集的长江流域570个旱地和434个水田田间氮径流流失数据组，采用相关性分析、结构方程模型、方差分解和机器学习方法，探究了影响旱地和水田总氮径流流失强度的主要因素，建立了基于机器学习的长江流域旱地和水田总氮径流流失强度预测模型，量化了长江流域农田总氮径流流失负荷.结果表明，径流深、施氮量和土壤氮含量是影响旱地总氮径流流失强度的主要因素；径流深和施氮量是水田总氮径流流失强度的主要影响因素.与分类与回归树、多元线性回归和增强回归树方法相比，采用随机森林算法构建的长江流域旱地和水田总氮径流流失强度预测模型具有更高的精度（R²为0.65~0.94）.基于随机森林算法的预测模型估算的2013年长江流域农田总氮径流流失负荷（以N计）为0.47 Tg ·a^-1（旱地：0.25 Tg ·a^-1；水田：0.22 Tg ·a^-1），中下游地区贡献了58%的流失负荷.模型预测5种防治情景下的长江流域农田氮流失负荷可削减2.4%~9.3%，其中减少径流量的削减效果最为显著.长江流域农田氮面源污染防治必须协同加强氮肥精准管理、减少农田径流量和提高土壤氮利用，且应将重点放在中下游地区.所发展的基于机器学习建模方法克服了氮径流流失强度与影响因素之间函数关系难以确定的问题，为估算区域或流域农田氮流失负荷提供了简便且可靠的方法.

Estimation of Cropland Nitrogen Runoff Loss Loads in the Yangtze River Basin Based on the Machine Learning Approaches

A quantitative understanding of cropland nitrogen (N) runoff loss is critical for developing efficient N pollution control strategies. Using correlation analysis, a structural equation model, variance decomposition, and machine learning methods, this study identified the primary influencing factors of total N (TN) runoff loss from uplands (n=570) and paddy (n=434) fields in the Yangtze River Basin (YRB) and then developed a machine learning-based prediction model to quantify cropland N runoff loss load. The results indicated that runoff depth, soil N content, and fertilizer addition rate were the major influencing factors of TN runoff loss from uplands, whereas TN runoff loss rate from paddy fields was mainly regulated by runoff depth and fertilizer addition rate. Among the four used machine learning methods, the prediction models based on the random forest algorithm presented the highest accuracy (R²=0.65-0.94) for predicting upland and paddy field TN runoff loss rates. The random forest algorithm based model estimated a total cropland TN loss load in the YRB of 0.47 Tg·a^-1 (upland:0.25 Tg·a^-1; paddy field:0.22 Tg·a^-1) in 2013, with 58% of TN runoff loss load derived from the midstream and downstream regions. The models predicted that TN runoff loss loads from croplands in YRB would decrease by 2.4%-9.3% for five scenarios, with higher TN load reductions occurring from scenarios with decreased runoff amounts. To mitigate cropland N nonpoint source pollution in YRB, it is essential to integrate efficient water, fertilizer, and soil nutrient managements as well as to consider the midstream and downstream regions as the high priority area. The machine learning-based modeling method developed in this study overcame the difficulty of identifying the functional relationships between cropland TN loss rate and multiple influencing factors in developing relevant prediction models, providing a reliable method for estimating regional and watershed cropland TN loss load.