福建平潭近海赤潮预警模型研究*

本文分析了福建省平潭近海海域2013—2019年水文、水质及气象数据的主成分结果, 筛选出5个气象因子和4个水质因子作为输入指标, 以藻密度为输出指标, 分别演算了KNN (K-nearest neighbor)、RF (random forest)、GBRT (gradient-boosted regression Trees)以及Bagging (bootstrap aggregating)4种赤潮预警回归模型。对2013—2019年的802 组海洋监测数据归一化处理后, 随机选取80%的数据作为模型的训练样本, 剩余的20%作为模型验证数据。其中, 以风速、气温、海平面气压、叶绿素a浓度组合为输入指标时, KNN回归模型演算结果的精度较高(R²=0.624, RMSE=0.821μg·L^-1, MAE=0.836μg·L^-1)。在没有叶绿素a浓度监测指标的海域, 构建了以叶绿素a浓度为输出指标, 气温、日照、风速、AOI(apparent oxygen increase)组合为输入指标的BP神经网络赤潮模型, 该模型也具有较好的预警精度(R²=0.651, RMSE=0.062μg·L^-1, MAE=0.033μg·L^-1)。本研究结果可为平潭海域的赤潮预警研究提供参考。

Study on the early warning model of red tide in the offshore area of Pingtan,Fujian province

We analyzed the principal components of hydrology, water quality, and meteorological data in Pingtan, Fujian province from 2013 to 2019. We selected 5 meteorological factors and 4 water quality factors. Our study establishes four early-warning model, KNN (K-nearest neighbor), RF (random forest), GBRT (gradient-boosted regression trees), Bagging (bootstrap aggregating) with meteorological factors and water quality factors as input indicators, and algal cell density as output indicators. After normalizing the 802 sets of marine monitoring data from 2013 to 2019, 80% of the data were randomly selected as the model training samples, and the remaining 20% were used as data of model verification. When temperature, wind speed, sea level pressure, and chlorophyll a are used as input parameters, the calculation result of KNN regression model is more accurate (R²=0.624, RMSE=0.821 μg·L^-1, MAE=0.836 μg·L^-1). In the sea area without chlorophyll a monitoring index, a BP neural network early-warning model with chlorophyll a concentration as the output index and temperature, sunshine, wind speed and AOI as input parameters was established, which has better warning accuracy (R²=0.651, RMSE=0.062 μg·L^-1, MAE=0.033 μg·L^-1). Our results can provide a reference for the red tide early warning research in the Pingtan coastal area.