| 摘要: |
| 针对我国近海总氮污染问题, 提出了一种基于卷积神经网络(Convolutional Neural Network, CNN)和可解释方法SHAP(SHapley Additive exPlanations)的河流入海总氮通量可解释预测模型, 模型耦合了马尔科夫链模拟的流域河网拓扑结构, 并充分利用多源时空数据。构建的模型应用于小清河, 将小清河流域气象、土地利用、土壤类型以及点源和非点源氮排放等多源数据, 通过马尔科夫链河网结构, 转换为三维输入数据。模型评估显示, 三维输入方式的模型在训练集和测试集上都表现出更高的准确性, 预测入海通量的相关系数达到了0.99。使用SHAP方法识别了影响模型预测的关键因素, 并分析了空间特征对模型预测的影响, 揭示了流域不同空间位置对入海总氮通量的影响差异。研究结果不仅提高了入海水质预测的准确性, 也为近海环境管理提供了科学依据。 |
| 关键词: 总氮通量预测 CNN模型 SHAP 深度学习 |
| DOI：10.11759/hykx20250114003 |
| 分类号:X522 |
| 基金项目:国家自然科学基金-山东省联合基金(U1906215) |
|
| Analysis of factors affecting total nitrogen flux into the sea from the Xiaoqing River based on CNN-SHAP |
|
FAN Zhicheng1,2, PENG Hui1,2, WANG Shuo3
|
|
1.Key Laboratory of Marine Environment Science and Ecology, Ministry of Education and College of Environmental Science and Engineering, Ocean University of China, Qingdao 266100, China;2.Shandong Provincial Key Laboratory of Marine Environment and Geological Engineering, Ocean University of China, Qingdao 266100, China;3.School of Environmental Science and Engineering, Shandong University, Qingdao 266237, China
|
| Abstract: |
| This study proposes an interpretable prediction model for total nitrogen flux from rivers to the sea to address the problem of total nitrogen pollution in China’s coastal waters. The model is based on a Convolutional Neural Network (CNN) and the SHAP (SHapley Additive exPlanations) methods. It couples the river network topology structure simulated by a Markov chain and fully utilizes multisource spatiotemporal data. For the purposes of this study, the model is applied to the Xiaoqing River. Multisource data, such as meteorology, land use, soil type, and point and nonpoint source nitrogen emissions in the Xiaoqing River Basin, are converted into three-dimensional input data based on the Markov chain river network structure. Model evaluation shows that the model with three-dimensional input performs better in both the training set and the test set, achieving higher accuracy. The correlation coefficient of the predicted inflow flux reaches 0.99. The SHAP method is used to identify the key factors that affect the model’s prediction and analyze the influence of spatial features on the prediction, revealing differences in the impact of different spatial locations in the basin on the total nitrogen flux to the sea. The research results not only improve the accuracy of the prediction of sea water quality but also provide a scientific basis for the management of the coastal environment. |
| Key words: total nitrogen flux prediction CNN model SHAP deep learning |
