基于可解释集成学习的农业流域地表水总氮浓度预测及关键驱动因子识别：以杭埠河流域为例

Interpretable ensemble learning-based prediction of surface-water total nitrogen concentration and identification of key driving factors in an agricultural watershed: a case study of the Hangbu River Watershed

摘要

【目的】为提高农业流域地表水总氮（total nitrogen，TN）浓度预测精度并识别关键驱动因子，以巢湖重要支流杭埠河流域为研究对象开展分析。【方法】基于 2024 年 5 月—2025 年 9 月研究区 62 个采样点的监测数据，共采集 558 份水样；经缺失值筛查后，保留 535 个有效样本，并构建包含气象、水分、植被、土壤和景观格局信息的解释变量数据集。比较岭回归（ridge）、支持向量回归（support vector regression，SVR）、随机森林（random forest，RF）、极端随机树（extremely randomized trees，ET）和极端梯度提升（extreme gradient boosting，XGBoost）等模型，并进一步构建基于折外预测（out-of-fold，OOF）的 Stacking 集成模型；采用 SHapley Additive exPlanations （SHAP）方法识别关键驱动因子。【结果】不同模型对杭埠河流域地表水 TN 浓度的预测效果存在明显差异，其中 Stacking 模型的测试集平均决定系数（R²）、均方根误差（RMSE）和平均绝对误差（MAE）分别为 0.810 6 ± 0.047 6、0.897 4 ± 0.096 4 和 0.625 6 ± 0.033 3，该模型在 R²、RMSE 及结果稳定性方面表现较优。SHAP 分析表明，归一化植被指数（NDVI）、温度、降水、蒸散量和土壤湿度是影响 TN 浓度预测的主要因子。【结论】杭埠河流域地表水 TN 浓度可通过集成学习方法实现较高精度预测，其预测结果与植被状态、气象条件及流域水分环境密切相关。研究结果可为流域水质预测、关键变量识别及农业面源污染防控提供参考。

Abstract

[Objective] To improve the prediction accuracy of surface-water total nitrogen (TN) concentration in agricultural watersheds and identify key driving factors, the Hangbu River Watershed, an important tributary of Chaohu Lake, was selected as the study area. [Method] Based on monitoring data collected from 62 sampling sites in the study area from May 2024 to September 2025, a total of 558 water samples were obtained. After screening for missing values, 535 valid samples were retained, and a dataset containing explanatory variables related to meteorology, water conditions, vegetation, soil, and landscape pattern was constructed. Ridge regression (Ridge), support vector regression (SVR), random forest (RF), extremely randomized trees (ET), and extreme gradient boosting (XGBoost) were compared, and an out-of-fold (OOF)-based Stacking ensemble model was further developed. SHapley Additive exPlanations (SHAP) was used to identify key driving factors. [Result] The prediction performance of different models for surface-water TN concentration in the Hangbu River Watershed varied considerably. Among them, the Stacking model achieved an average test coefficient of determination (R²), root mean square error (RMSE), and mean absolute error (MAE) of 0.810 6 ± 0.047 6, 0.897 4 ± 0.096 4, and 0.625 6 ± 0.033 3, respectively, showing relatively better performance in terms ofR², RMSE, and result stability. SHAP analysis indicated that normalized difference vegetation index (NDVI), temperature, precipitation, evapotranspiration, and soil moisture were the major factors affecting TN concentration prediction. [Conclusion] Surface-water TN concentration in the Hangbu River Watershed can be predicted with relatively high accuracy using ensemble learning methods, and its variation is mainly driven by vegetation status and meteorological-hydrological conditions. The results can provide a reference for watershed water-quality prediction, identification of key driving factors, and control of agricultural non-point source pollution.

关键词

农业流域 / 总氮浓度 / 关键驱动因子 / 可解释集成学习 / SHAP

Key words

agricultural watershed / total nitrogen concentration / key driving factors / interpretable ensemble learning / SHAP

参考文献

[1] Conley D J, Paerl H W, Howarth R W, et al. Controlling eutrophication: nitrogen and phosphorus[J].Science, 2009, 323(5917): 1014-1015.

[2] 嵇晓燕, 王姗姗, 杨凯, 等. 2016—2020 年中国地表水中总氮浓度时空变化特征分析[J].生态环境学报, 2022, 31(6): 1184-1192.

[3] 秦先燕, 吴剑雄, 郑涛, 等. 店埠河流域地表水氮素时空分布和来源解析[J].环境科学, 2024, 45(12): 7100-7110.

[4] 陈传保 . 杭埠河舒城段水质现状评价[J].安徽农业科学, 2011, 39(16): 9774-9775.

[5] 徐洁, 耿伯彰, 娄英斌, 等. 基于 LSTM 模型的地表水总氮预测模型[J].中国环境监测, 2025, 41(2): 63-68.

[6] Zhang K, Wang X Y, Liu T L, et al. Enhancing water quality prediction with advanced machine learning techniques: an extreme gradient boosting model based on long short—term memory and autoencoder[J].J Hydrol, 2024, 644: 132115.

[7] Lundberg S M, Lee S I . A unified approach to interpreting model predictions[C]//Proceedings of the 31st International Conference on Neural Information Processing Systems. ACM, 2017: 4768-4777.

[8] Ruan J L, Cui Y, Song Y C, et al. A novel RF—CEEMD—LSTM model for predicting water pollution[J].Sci Rep, 2023, 13: 20901.

[9] Elsayed A, Rixon S, Levison J, et al. Machine learning models for prediction of nutrient concentrations in surface water in an agricultural watershed[J].J Environ Manag, 2024, 372: 123305.

[10] 水质总氮的测定碱性过硫酸钾消解紫外分光光度法: HJ 636—2012[S].

[11] Pei W, Xu Q Y, Lei Q L, et al. Interactive impact of landscape composition and configuration on river water quality under different spatial and seasonal scales[J].Sci Total Environ, 2024, 950: 175027.

[12] Liu H Y, Meng C, Wang Y, et al. Establishing the relationship between the integrated multidimensional landscape pattern and stream water quality in subtropical agricultural catchments[J].Ecol Indic, 2021, 127: 107781.

[13] Xu S, Li S L, Zhong J, et al. Spatial scale effects of the variable relationships between landscape pattern and water quality: Example from an agricultural karst river basin, Southwestern China[J].Agric Ecosyst Environ, 2020, 300: 106999.

[14] Chen T Q, Guestrin C . XGBoost: a scalable tree boosting system[C]//Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. ACM, 2016: 785-794.

[15] Li Z Q . Extracting spatial effects from machine learning model using local interpretation method: an example of SHAP and XGBoost[J].Comput Environ Urban Syst, 2022, 96: 101845.

[16] Rixon S, Levison J, Binns A, et al. Spatiotemporal variations of nitrogen and phosphorus in a clay plain hydrological system in the Great Lakes Basin[J].Sci Total Environ, 2020, 714: 136328.

[17] Yuan Y, Zhou C J, Wu J W, et al. An interpretable deep learning framework for river water quality prediction: a case study of the Poyang Lake Basin[J].Water, 2025, 17(16): 2496.

[18] Newcomer M E, Bouskill N J, Wainwright H, et al. Hysteresis patterns of watershed nitrogen retention and loss over the past 50 years in United States hydrological basins[J].Glob Biogeochem Cycles, 2021, 35(4): e2020GB006777.

[19] Xu Y T, Li P, Ma F M, et al. Watershed landscape characteristics and connectivity drive river water quality under seasonal dynamics[J].J Clean Prod, 2024, 473: 143533.

[20] Lu C Q, Zhang J E, Tian H Q, et al. Increased extreme precipitation challenges nitrogen load management to the Gulf of Mexico[J].Commun Earth Environ, 2020, 1: 21.

[21] Ballard T C, Sinha E, Michalak A M . Long—term changes in precipitation and temperature have already impacted nitrogen loading[J].Environ Sci Technol, 2019, 53(9): 5080-5090.

[22] Hess L J T, Hinckley E S, Robertson G P, et al. Rainfall intensification increases nitrate leaching from tilled but not no—till cropping systems in the U.S. Midwest[J].Agric Ecosyst Environ, 2020, 290: 106747.

[23] Zhu X Q, Miao P, Zhu H, et al. Extreme precipitation accelerates nitrate leaching in the intensive agricultural region with thick unsaturated zones[J].Sci Total Environ, 2024, 918: 170789.

[24] Wu J H, He S J, Lu J . Multi—scale effects of topography and landscape pattern on riverine nitrogen and phosphorus nutrients in an agricultural watershed[J].Landsc Ecol, 2025, 40(6): 112.

[25] Meng C, Liu H Y, Wang Y, et al. Landscape pattern exhibits threshold—driven effect on nitrogen export of typical land use in subtropical hilly watershed under specific hydrological regimes[J].J Clean Prod, 2023, 419: 138322.

[26] Lyu C J, Li X J, Yuan P, et al. Nitrogen retention effect of riparian zones in agricultural areas: a meta—analysis[J].J Clean Prod, 2021, 315: 128143.

[27] Takefuji Y. Limitations of SHAP—based interpretations in environmental and membrane filtration applications[J].Water Res, 2026, 288: 124766.

[28] IPCC Intergovernmental Panel on Climate Change. Climate Change 2022 — Mitigation of Climate Change: Working Group III Contribution to the Sixth Assessment Report of the Intergovernmental Panel on Climate Change[M]. Cambridge: Cambridge University Press,2023.

[29] Zhang J E, Lu C Q, Feng H L, et al. Extreme climate increased crop nitrogen surplus in the United States[J].Agric For Meteor, 2021, 310: 108632.

基金资助

安徽省自然科学基金项目(2308085MD114)

国家自然科学基金项目(32171573)

安徽省高校自然科学研究项目(2023AH051002)