《海洋预报》| 基于CNN-SVR网络的黄渤海近岸海域叶绿素a浓度预测

国家海洋预报台

2024-10-08 17:37发布于北京国家海洋预报台官方账号

基于CNN-SVR网络的黄渤海近岸海域叶绿素a浓度预测

全文请用PC端下载地址：

读书小笔记

作者：王晓霞^1 2 3 汪健平^1 3 王佳莹^1 3 孙珊^1 3 苏博^1 3 姜会超^1 3 朱明明^1 3

单位：1. 山东省海洋资源与环境研究院, 山东烟台 264006;
2. 自然资源部空间海洋遥感与应用重点实验室, 北京 100081;
3. 山东省海洋生态修复重点实验室, 山东烟台 264006

分类号：X145

出版年·卷·期（页码）：2024·41·第四期（77-87）

摘要：利用海洋卫星观测数据和黄渤海近岸海域实测生态水质数据，建立了一种基于卷积神经网络结合支持向量回归（Convolutional Neural Network-Support Vector Regression，CNN-SVR）的深度学习网络模型的叶绿素a浓度预测方法。采用皮尔逊方法对叶绿素a与环境动力因子和生态水质因子作相关分析，发现营养盐因子大多与叶绿素 a有显著相关性，水质因子如 pH、溶解氧、盐度等与叶绿素a的相关性不大；将黄渤海近岸海域划分为渤海南部与黄海北部、黄海中部，进行春夏、秋冬两个时期 1×1和 2×2两种卷积核大小的 CNN-SVR网络模型实验以及单因子敏感性分析试验。结果显示：卷积核大小为2×2时，CNN-SVR网络模型对训练数据的学习和对测试样本的预测检验效果都更优；渤海南部与黄海北部近岸海域模型预测效果更好。营养盐因子对模型预测能力的影响更显著，悬浮物等水质因子的影响相对较弱。单变量对模型预测的敏感性较弱，多变量整合具有互补性，改善了模型的预测效果。

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

Abstract：A chlorophyll-a (Chl-a) concentration prediction method based on the Convolutional Neural NetworkSupport 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.

Key words：CNN-SVR; chlorophyll-a concentration; sensitivity analysis of single variable; ocean satellite; ecological factors of water quality in ocean

读者也可以通过公众号菜单栏

“相关知识”板块进入《当期目录》和《过刊浏览》

查阅详细期刊内容