Multi-Feature Attention Based LSTM Network for Sea Surface Temperature Prediction

Abstract:
Sea surface temperature (SST) prediction is an important part of atmospheric and Marine science and has been widely used in many fields including climate studies and oceanic environmental protection. Previous works have attempted to exploit the spatial and temporal features of the SST to make prediction, but rarely consider other SST-related features. This paper builds a multi-feature model based on long short term memory (LSTM) neural network. We model the time sequence through the LSTM layer and integrated external features of the ocean. These external features include the SST-related marine hydrological elements (salinity, water velocity) and features derived from SST (spatial, periodic and increment features). In addition, feature attention mechanism is used to make the model adaptively and dynamically assign weights to external features according to their importance in SST prediction. Experimental results on the high resolution data set of Bohai Sea area demonstrate the effectiveness of the proposed model.