A chlorophyll-a(Chl-a)concentration prediction method based on the Convolutional Neural Network-Support Vector Regression(CNN-SVR)model is developed using satellite observations and in-situ ecological water quality measurements in near-shore waters of the Yellow and Bohai Seas.Firstly,we use Pearson method to establish correlation between Chl-a concentration and factors of environmental dynamics and ecological water quality.It is found that Chl-a concentration correlates significantly with nutrient salt factors,while poorly with water quality factors such as pH,dissolved oxygen,salinity.Then,we divide two regions,one is nearshore waters of the southern Bohai Sea and northern Yellow Sea,and the other one is nearshore waters of the central Yellow Sea.We also divide two periods:spring‒summer and autumn‒winter.We perform the CNN-SVR model experiments with two convolutional kernel sizes,1×1 and 2×2,as well as the single factor sensitivity analysis experiment.The results show that the CNN-SVR network model has better learning of the training data and better prediction of the test samples when the convolution kernel size is 2×2.The CNN-SVR network model performs better in nearshore areas of the southern Bohai Sea and northern Yellow Sea.Compared to water quality factors,the nutrient salt factors have larger impacts on the model's prediction ability.The sensitivity of single factor to model's prediction ability is weak,while multiple variables exhibit complementary feature which improves the model's prediction ability.
关键词
卷积神经网络结合支持向量回归模型/叶绿素a浓度预测/单因子敏感性分析/海洋卫星/海洋生态水质因子
Key words
CNN-SVR/chlorophyll-a concentration/sensitivity analysis of single variable/ocean satellite/ecological factors of water quality in ocean
维普
万方数据