基于全局跨尺度时空注意力的深度神经网络海表面温度预测模型  

A deep neural networks prediction model for sea surface temperature based on global cross-scale spatial-temporal attention

作　　者：杜家豪 聂婕[1] 叶敏 宋德瑞[2] 高振 魏志强[1] DU Jiahao;NIE Jie;YE Min;SONG Derui;GAO Zhen;WEI Zhiqiang(College of Information Science and Engineering,Ocean University of China,Qingdao 266100,China;National Marine Environmental Monitoring Center,Dalian 116023,China;School of Mathematical Sciences,Ocean University of China,Qingdao 266100,China)

机构地区：[1]中国海洋大学信息科学与工程学部,山东青岛266100 [2]国家海洋环境监测中心,辽宁大连116023 [3]中国海洋大学数学科学学院,山东青岛266100

出　　处：《海洋环境科学》2023年第6期944-954,共11页Marine Environmental Science

基　　金：国家重点研发计划项目(2021YFF0704000);山东省中央引导地方科技发展资金项目(YDZX2022028);国家自然科学基金项目(62172376)。

摘　　要：海表面温度(sea surface temperature, SST)是海洋与大气之间相互作用的关键因素,海温控制着全球大气和海洋生态系统的变化。准确预测海表面温度的演变对治理全球大气系统和海洋生态系统都具有重要的意义。为了对SST数据的空间自相关性准确建模,本文提出了基于全局跨尺度时空注意力的深度神经网络海表面温度预测模型(deep neural network based on global cross-scale spatiotemporal attention, GCSA-DNN)。模型分为3个部分,从长时序数据中提取时序依赖特征的时序建模模块,从SST序列均值中提取空间分布规律特征的多尺度局部空间建模模块和基于全局跨尺度的时空注意力融合模块,实现每个网格点对全局自相关性的建模。本研究选择空间分布规律不同的东海和南海海域数据,对1981年9月1日至2022年4月7日美国国家海洋和大气管理局(nationaloceanicand atmospheric administration,NOAA)的数据进行了预测分析,共14829条数据,其中1981年9月1日至2021年8月31日的70%数据用于训练,30%用于验证,2021年9月1日至2022年4月7日的数据用于测试。结果表明,在不同的实验条件下该模型可以准确捕捉SST数据在时空过程中的演变规律,在东海和南海SST数据集上其准确度相较于卷积长短时记忆神经网络(convolutional long short-term memory, ConvLSTM)分别提高了14.07%和14.18%,提升了SST预测的准确度。Sea surface temperature(SST)is a crucial factor in the interaction between the ocean and the atmosphere,controlling the changes of global atmospheric and oceanic ecosystems.Accurately predicting the evolution of SST is of great significance in managing global atmospheric and oceanic systems.To accurately model the spatial autocorrelation of SST data,this paper proposes a deep neural network based on global cross-scale spatiotemporal attention(GCSA-DNN)for predicting SST.The model consists of three parts:a temporal modeling module that extracts temporal dependent features from long time series data,a multi-scale local spatial modeling module that extracts spatial distribution pattern features from SST series mean values,and a global cross-scale spatiotemporal attention fusion module that models the global autocorrelation of each grid point.In this study,the data of the East China Sea and South China Sea with different spatial distribution patterns are selected,and the national oceanic and atmospheric administration(NOAA)data from September 1,1981 to April 7,2022 are predicted and analyzed with a total of 14829 data,of which 70%of the data from September 1,1981 to August 31,2021 were used for training,30%for validation,and the data from September 1,2021 to April 7,2022 were used for testing.The results show that the model can accurately capture the evolution pattern of SST data in spatial-temporal processes in different experimental conditions,with an accuracy improvement of 14.07%and 14.18%on the East China Sea and South China Sea SST datasets,respectively,compared with the convolutional long short-term memory(ConvLSTM)model,achieving an improvement in SST prediction accuracy.

关 键 词：海表面温度 时空预测 多尺度 深度神经网络 注意力机制 

分 类 号：X16[环境科学与工程—环境科学]