摘要
海表面温度预报在海洋相关领域具有重要的实用价值,随着遥感信息采集技术的不断发展和完善,区域内海表面温度数据采集的完整性得到了保障。现今大多数方法在预报海表面温度时,只考虑了海表面温度的时间相关性,并未利用其空间相关性,使得预报精度受到限制。针对该问题,本文将区域内每天的海表面温度数据作为一个矩阵输入模型,便于时间和空间信息的提取,并提出了CA-ConvLSTM模型来预报海表面温度。该模型首先利用卷积层对海表面温度矩阵进行局部特征提取,然后通过注意力模型为矩阵序列分配权重,将权重与矩阵序列对应相乘得到加权特征序列,最后,利用ConvLSTM进行预报,获得未来一天或五天内的海表面温度。通过实验确定模型的结构、输入尺寸和k值,再将CA-ConvLSTM与SVR、LSTM和ConvLSTM进行对比。实验结果表明:CA-ConvLSTM的均方根误差(Root Mean Square Error,RMSE)和预报精度(Prediction Accuracy,PACC)指标均要优于其他三种预报方法,验证了本文方法的有效性。
Sea surface temperature prediction has important practical value in marine related fields.With the continuous development and improvement of remote sensing information collection technology,the integrity of sea surface temperature data collection in the region has been guaranteed.At present,most methods which are used to predict the sea surface temperature only consider the time dependence of sea surface temperature and do not utilize the spatial correlation,so the prediction accuracy is limited.In response to this problem,this paper uses the daily sea surface temperature data in the region as a matrix input model to facilitate the extraction of time and space information.The CA-ConvLSTM model is proposed to predict sea surface temperature.Firstly,the convolution layer is used to extract the local features of the matrix sequence;then,the attention model is used to assign weights to the matrix sequence,and the weights are multiplied by the matrix sequences to obtain the weighted feature sequences;finally,ConvLSTM is used to predict sea surface temperatures for the next day or five days.Through experimental analysis,structure,input size and k value of the model are determined,and then CA-ConvLSTM with SVR,LSTM and ConvLSTM are compared.The experimental results show that the RMSE and PACC indicators of CAConvLSTM are better than the other three prediction methods,which verifies the effectiveness of the proposed method.
作者
查铖
贺琪
宋巍
郝增周
黄冬梅
胡泽煜
ZHA Cheng;HE Qi;SONG Wei;HAO Zengzhou;HUANG Dongmei;HU Zeyu(Department of Information Technology,Shanghai Ocean University,Shanghai 201306,China;Shanghai University of Electric Power,Shanghai 200090,China;State Key Laboratory of Satellite Ocean Environment Dynamics,Second Institute of Oceanography,Ministry of Natural Resources,Hangzhou 310012,China)
出处
《海洋通报》
CAS
CSCD
北大核心
2020年第2期191-199,共9页
Marine Science Bulletin
基金
海洋大数据分析预报技术研发基金(2016YFC1401902)
国家海洋局数字海洋科学技术重点实验室开放基金(B201801029)
上海市高校特聘教授(东方学者)项目(TP2016038)
上海市科委部分地方院校能力建设项目(17050501900)。
关键词
时间序列
海表面温度预报
空间相关性
注意力机制
time series
sea surface temperature prediction
spatial correlation
attention mechanism
