Observing typhoon-driven upper ocean dynamics in the South China Sea using a virtual mooring underwater glider array: Methods and analysis

Abstract

Accurate and persistent observation of typhoon-induced upper ocean dynamics remains challenging due to strong currents and the limited station-keeping capability of autonomous platforms. This study proposes a learning-based virtual mooring framework using an underwater glider array to observe typhoon-driven processes in the South China Sea. A deep reinforcement learning approach based on the Dueling Double Deep Q-Network (D3QN) architecture is developed to improve station-keeping under dynamic flows and forecast uncertainties. Simulations involving 840,000 dive profiles across 400 sites show that the D3QN method achieves the lowest average station-keeping error (0.68 km), outperforming reactive and predictive strategies. Robustness tests confirm reliable performance under perturbed forecasts. Field experiments during typhoon conditions further demonstrate the feasibility of glider-based virtual mooring, and glider observations reveal pronounced upper-ocean responses, highlighting the framework’s utility for capturing typhoon-induced variability.