Modeling Coastal Eutrophication at Florida Bay using Neural Networks
Modeling Coastal Eutrophication at Florida Bay using Neural Networks
复制标题
DOI：
10.2112/06-0646.1
复制
发表时间：
2008
期刊：
影响因子：
--
通讯作者：
A. Melesse;J. Krishnaswamy;Keqi Zhang
中科院分区：
文献类型：
--
作者：
A. Melesse;J. Krishnaswamy;Keqi Zhang
摘要近岸海域营养盐负荷和富营养化是造成水质退化和海洋生物丧失的原因，导致生态失衡。理解富营养化程度并将其作为环境参数的函数进行建模，有助于沿海生态系统的管理。由于确定性模型和经验模型在准确预测藻类水华水平方面的局限性，以及水质和环境参数与叶绿素水平之间的非线性关系，需要一种使用机器学习和数据驱动建模的新方法。利用人工神经网络(ANN)的多层感知器-反向传播(MLP-BP)算法对佛罗里达州海湾两个水质监测站(FLAB03和FLAB14)监测的水质参数进行了富营养化(叶绿素a)预测。基于月养分(总磷、亚硝酸盐、氨氮)和其他水分数据(温度、浊度和溶解氧)与叶绿素a水平的相关性，选择了输入输出数据结构。研究了7个输入数据场景，并用4个指标对模型性能进行了比较。从1992年到2004年的月度数据被划分为训练和测试子集。结果表明，所选输入对叶绿素a的预测效果较好，平均R2为0.856，模型效率(E)为0.582。单独使用前件叶绿素a的预测结果稳定，误差较小，由于训练更容易和更有效，预测效果更好。研究还发现，ANN在FLA03上的表现好于FLA14上的表现。结果表明，MLP-BP技术适用于赤潮的监测和预报，对沿海流域管理具有重要意义。
Abstract Nutrient loading and eutrophication in coastal waters are the causes of water quality degradation and loss of marine biota, which has led to ecological imbalance. Understanding and modeling the level of eutrophication as a function of environmental parameters can be beneficial to coastal ecosystem management. The limitation of deterministic and empirical models in accurately predicting the level of algal blooms, and the nonlinear relationship between the water quality and environmental parameters and that of the level of chlorophyll a necessitate a new approach using machine learning and data-driven modeling. A multilayer perceptron-back propagation (MLP-BP) algorithm of artificial neural network (ANN) was used to predict the level of eutrophication (chlorophyll a) from water quality parameters monitored at two Florida Bay water quality monitoring stations (FLAB03 and FLAB14). Based on the correlation of monthly nutrients (total phosphate, nitrite, ammonium) and other water data (temperature, turbidity, and dissolved oxygen) to the level of chlorophyll a, an input-output data structure was selected. Seven input data scenarios were studied, and model performance was compared using four indices. Monthly data from 1992 to 2004 were partitioned into training and testing subsets. Results show that chlorophyll a was predicted well with the selected inputs, with an average R2 and model efficiency (E) of 0.856 and 0.582, respectively. Prediction with antecedent chlorophyll a alone gave a stable result with smaller error and higher performance attributed to easier and more efficient training. It was also found that ANN performed better at FLA03 than at FLA14. It is shown that the MLP-BP technique is applicable to the monitoring and prediction of algal blooms and will be crucial to coastal watershed management.