End-to-End Nomadic Ocean Forecasting: Bridging Compression, Assimilation, and Task-Specific Metrics Using Deep Learning

Nomadic ocean forecasting in bandwidth-constrained maritime environments demands innovative solutions to reconcile high-resolution forecasts with sparse onboard observations. Traditional workflows rely on independent data transmission and data assimilation schemes, and struggle to preserve critical features for operational tasks such as underwater acoustics. This study proposes an end-to-end deep learning framework integrating task-aware compression and data assimilation to address these challenges. Using high-resolution NEMO ocean model simulations in a Gulf Stream region, we evaluate two case studies: (1) Convolutional autoencoders (CAEs) for 3D+t sound speed profile (SSP) compression, demonstrating superior performance over principal component analysis (PCA) at compression ratios above 20, with a 31% reduction in reconstruction error; and (2) 4DVarNet-based neural reconstruction of surface duct width (SDW) fields using sparse in situ measurements and sea surface temperature (SST) synergies. The multimodal 4DVarNet-SST achieves superior spatiotemporal resolution (minimum spatial scale: 1.81°, temporal scale: 33.87 days) and a 11% improvement in root-mean-square error (RMSE) over optimal interpolation. By unifying compression, assimilation, and task-aware metrics, this work advances relocatable forecasting for mission-critical applications, emphasizing the role of end-to-end learning in preserving operationally relevant oceanographic features.