Adaptive hierarchical representation of marine environmental variables based on mixture of experts

The representation of marine environmental variables is the basis of underwater acoustic information processing and ocean dynamics analysis. However, the marine physical field shows significant non-stationarity and spatial heterogeneity on the vertical scale. Under a parameter-sharing paradigm, the traditional monolithic deep learning model is prone to encountering a representation capacity bottleneck when dealing with stratified physical characteristics. To solve this problem, this paper proposes an adaptive hierarchical representation framework based on the Mixture of Experts (MoE) model, which aims to achieve deep feature decoupling and fine modeling of high-dimensional oceanic fields. The core innovation of this architecture is the design of a physics-informed gating network driven by ocean dynamics, which utilizes physical property gradients to dynamically distribute the computational workload of vertical profiles to specific expert subnetworks. In order to enhance the generalization ability of the model under sparse monitoring conditions, this paper introduces a multi-granularity contrastive learning strategy, which strengthens the model’s ability to capture the spatio-temporal evolution characteristics of complex phenomena such as mesoscale eddies through multi-scale association constraints. In addition, the model integrates a physics-informed regularization term, forcing the characterization process to follow vertical sound speed profiles and the principle of thermodynamic consistency. The experimental results show that the proposed framework significantly outperforms conventional methods in improving representation accuracy and physical fidelity, providing a robust solution for high-fidelity modeling in complex marine environments.