831 / 2024-09-19 15:50:59
Using Deep Learning to Estimate the Temperature, Salinity, and Heat-Salt Exchanges across the Indonesian Seas
Deep learning,Indonesian Seas,salinity,heat-salt flux
Session 32 - Digital twins of the ocean (DTO) and its applications
Abstract Accepted
This study leverages deep learning techniques, a hybrid model combining ResNet and Transformer architectures, to predict the three-dimensional evolution of temperature and salinity in the Indonesian seas, based on available observations and forcing conditions. The ResNet component is employed to extract local spatial features, capturing small-scale structures and localized variations that are critical for accurate prediction, while the Transformer is particularly useful for identifying how temperature and salinity patterns evolve and are connected across larger spatial scales. The results show that the model can capture the primary characteristics of 3D temperature and salinity variations in the region, providing reasonable estimations even in subsurface layers with limited observational data. Additionally, the model provides estimates of heat-salt fluxes through the key straits, such as the Makassar, Lombok, and Ombai Strait. This work highlights the robustness of the hybrid ResNet-Transformer model in predicting both surface and subsurface conditions, offering a valuable tool for studying thermohaline dynamics in this region. Furthermore, the study underscores the potential for deep learning models to enhance the understanding of ocean dynamics and thermohaline processes in regions with scarce observations.