《海洋预报》| 基于Stacking集成学习的日尺度海洋三维温度场重构

摘要：提出了一种基于Stacking集成学习的日尺度海洋三维温度场重构方法，旨在通过融合多源海面观测数据与历史温盐观测数据，提升温度场重构的精度与鲁棒性。本研究采用随机森林、线性回归、决策树、K近邻、多层感知器和梯度提升树等多种基模型，结合Optuna超参数优化技术，构建了Stacking集成学习模型。结果表明：Stacking模型在0～1 000 m深度范围内的深度平均均方根误差（RMSE）为0.840℃，显著优于单一基模型（最优基模型平均RMSE为0.992℃）。通过与黑潮延伸体区域浮标观测站和GLORYS再分析数据的对比，验证了重构温度场在时间序列平滑性和温度变化捕捉能力上的优越性。

关键词：集成学习 重构 温度场 Stacking

Abstract：In this paper, we propose a daily-scale ocean 3D temperature field reconstruction method based on Stacking integrated learning, aiming to improve the accuracy and robustness of the temperature field reconstruction by integrating multi-source sea surface observation data and historical temperature and salt observation data. The study adopts various base models such as random forest, linear regression, decision tree, K-nearest neighbor, multilayer perceptron and gradient boosting tree, and combines them with Optuna's hyperparameter optimization technique to construct the Stacking integrated learning model. The experimental results show that the average Root Mean Square Error(RMSE) of the Stacking model is 0.840 ℃ in the depth range of 0 to 1 000 m, which is significantly better than that of the single base model(the average RMSE of the optimal base model is 0.992 ℃). The superiority of the reconstructed temperature field in terms of time-series smoothing and the ability to capture temperature variations is verified by comparing with the Kuroshio Extension Observatory(KEO) and GLORYS reanalysis data.

Key words：ensemble learning; reconstruction; temperature field; Stacking

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

未来两天海洋环境预报

35个美丽海岛环境预报