| 摘要: |
| 海表面二氧化碳分压(pCO2)的未来变化趋势, 对统计评估全球碳收支以及理解全球气候变化背景下的海洋酸化现象至关重要。目前传统的海面pCO2预测方法大部分基于有限的实测数据, 然而实测数据存在着时间和地理方面的制约, 且计算成本较高。近年来, 随着时空观测数据的爆炸性增长, 基于深度学习的数据驱动模型在海表面pCO2预测方面中表现出良好的潜力。然而, 由于多种环境因素与海表面pCO2之间的关系错综复杂, 到目前为止尚无十分简单有效的相关模型来对海表面pCO2进行预测。为应对这一挑战, 利用时空卷积长短时记忆神经网络(ST-ConvLSTM)模型, 通过海面温度(sea surface temperature, SST)、海面盐度(sea surface salinity, SSS)、叶绿素a浓度(chl a)和海面pCO2数据, 预测南海的海面pCO2, 并将2019年1~12月的数据作为测试集对模型的表现进行了验证。结果显示, ST-ConvLSTM模型的预测因子均方根误差、平均绝对误差和决定系数分别为0.981 Pa、0.711 Pa和0.997。对比卷积LSTM (ConvLSTM)、随机森林和广义回归神经网络(generalized regression neural network, GRNN)三种方法, 证实本文所提出的方法在解决南海pCO2预测问题上是可靠的。 |
| 关键词: ST-ConvLSTM模型 中国南海 海表面二氧化碳分压 深度学习 |
| DOI：10.11693/hyhz20230300074 |
| 分类号: |
| 基金项目:国家自然科学基金项目,41830533号 |
附件 |
|
| SPATIAL AND TEMPORAL PREDICTION OF PCO2 IN THE SOUTH CHINA SEA BASED ON ST-CONVLSTM |
|
GAO Yu, LI Shuang, HAO Peng, SONG Jin-Bao
|
|
Ocean College, Zhejiang University, Zhoushan 316021, China
|
| Abstract: |
| Understanding the future trends in the partial pressure of carbon dioxide (pCO2) at sea surface is crucial to assess statistically the global carbon balance and ocean acidification in the context of global climate change. Most of the current traditional sea surface pCO2 prediction methods are based on limited real-world data, and require temporal and geographical constraints on the real-world data and high computational costs. In recent years, with the explosive growth of spatiotemporal observational data, data-driven models based on deep learning have shown good potential in sea surface pCO2 prediction. However, due to the complex relationship between multiple environmental factors and sea surface pCO2, there is no simple and effective relevant model in this regard so far. We deveveloped a spatio-temporal convolutional long and short-term memory neural network (ST-ConvLSTM) model to predict sea surface pCO2 in the South China Sea from sea surface temperature (SST), sea surface salinity (SSS), chlorophyll a concentration (chl a), and sea surface pCO2 data, and the model was validated using data from January to December 2019 as test set. Results show that the prediction factors, and Root Mean Square Error, Mean Absolute Error and Coefficient of Determination of the model are 0.981 Pa, 0.711 Pa, and 0.997, respectively. Among three methods of convolutional LSTM (ConvLSTM), random forest, and generalized regression neural network (GRNN), our method is most reliable in the pCO2 prediction in the South China Sea. |
| Key words: ST-ConvLSTM the South China Sea sea surface pCO2 prediction deep learning |
