Prediction of the temperature profiles using the generative model and the stacking ensemble method in a marginal sea

Various machine learning and neural network models using remote sensing data are useful for understanding and predicting the ocean status. Most machine learning methods have focused on predicting sea surface status because satellite data mainly provide only sea surface information. Subsurface data have been sparsely observed owing to difficulty accessing the ocean subsurface. Machine learning could be a viable option for predicting the subsurface status. Sea temperature profiles have been predicted using various combinations of machine learning methods and ensemble models. However, ocean observation data are insufficient for machine learning to resolve the complex spatiotemporal variability of sea temperature in marginal seas. We propose data augmentation of observation datasets to overcome sparse subsurface data in marginal seas. In the proposed method, remote sensing data are integrated with an augmented dataset to predict the temperature profile based on the stacking ensemble method. Further, we developed a neural network generative model-based solution to augment the observational dataset and make the preparation of datasets for training easier. The performance of the prediction model was evaluated using the observed dataset. The prediction model's performance was evaluated using the observed dataset. With artificially augmented data, the model showed improved accuracy, reducing RMSE from 2.51 to 1.20 in February, from 3.60 to 2.45 in August, demonstrating enhanced predictive capability and generalization. This paper is based on author's dissertation carried out at the SNU (Jung, 2022).