ABSTRACT

Sea surface temperature (SST) prediction plays an important role in planning marine operations and forecasting climate. With the rapid development of remote sensing technology, there are plenty of SST data available for scientific research. However, most previous studies ignored the quality of SST dataset and the importance of data at each time step, which limited the performance of prediction. Therefore, in order to fully exploit the features of SST data, we propose a model which combines cubic B-spline interpolation, attention mechanism and Long Short Term Memory network (LSTM), named CBSA-LSTM. In this model, we use cubic B-spline interpolation to enhance input data and make SST trend curve continuous derivative in time dimension, adjust time attention mechanism to let it more suitable for SST prediction, refine spatial attention mechanism to mainly focus on latitude, and then combine them with LSTM to predict daily SST. To our knowledge, it is the first attempt to use cubic B-spline interpolation to solve the problem of data quality for SST prediction. The experiment results indicate that the proposed model significantly improve the prediction performance and is reliable for SST prediction with high performance for wide time range and large spatial scope.

摘要

海面温度（SST）预测在计划海洋作业和预测气候中起着重要作用。随着遥感技术的飞速发展，有大量的SST数据可用于科学研究。但是，大多数先前的研究都忽略了SST数据集的质量和每个时间步长的数据重要性，这限制了预测的性能。因此，为了充分利用SST数据的特征，我们提出了一个将三次B样条插值，注意机制和长短期记忆网络（LSTM）相结合的模型，称为CBSA-LSTM。在此模型中，我们使用三次B样条插值法来增强输入数据，并使SST趋势曲线在时间维度上连续导数，并调整时间关注机制以使其更适合SST预测，完善空间注意力机制，主要关注纬度，然后将其与LSTM结合以预测每日的SST。据我们所知，这是首次尝试使用三次B样条插值来解决SST预测的数据质量问题。实验结果表明，所提出的模型可以显着提高预测性能，对于大范围的时间范围和大的空间范围内的海表温度预报是可靠的。