Using Layer-wise Depth Recursive Prediction to Predict Underwater Sea Temperature

The impact of sea temperature (ST) variations on climate and ecosystems is profound, posing not only significant economic losses but also potential threats to life safety. Consequently, accurate ST prediction has become an extremely important research topic. Early ocean observation data were mostly limited to the sea surface, but with the advancement of the Argo program, it is now possible to obtain ST data at various depths. Many studies have focused on predicting ST, commonly employing machine learning and deep learning techniques. However, most of these studies use only surface parameters to predict ST at different depths, leading to a significant decrease in prediction accuracy as depth increases. To address this issue, this paper proposes a method called Layer-wise Depth Recursive Prediction (LDRP). This method first trains a model to predict the ST at the uppermost depth and then uses the prediction results from the upper depth as input for the model at the next depth level, thereby improving the prediction accuracy at subsequent depths. Using data provided by the Argo program, this study combines the LDRP method with a Transformer-based model to train and predict ST at different depths. The experimental results show that the predictions made using the LDRP method exhibit improved accuracy across all depth levels compared to predictions made without LDRP, and the prediction stability at different depths is also significantly enhanced. Overall, this research not only effectively improves the accuracy of ST predictions at various depths but also mitigates the decline in prediction accuracy that occurs with increasing depth in current methods. These results demonstrate the significant potential of the LDRP method for deep ST prediction, and future studies may consider applying this method to even deeper sea temperature predictions to further verify its stability and accuracy.