The development of long-endurance autonomous marine vehicles, such as autonomous underwater gliders and sailboats, enables the formation of a persistent and mobile ocean sensing network. Utilizing the sparse sensing data from these networks for rapid and accurate prediction of spatiotemporal ocean fields is a crucial issue. This paper presents a comprehensive methodology to achieve this objective. Recognizing the multiscale characteristics of ocean environments, the multiresolution Dynamic Mode Decomposition (MrDMD) method is employed to establish a computationally lightweight reduced-order ocean model, leveraging historical prediction data from numerical ocean models. Real-time sparse sensing data from autonomous marine vehicles are then integrated using the ensemble Kalman filter, sequentially refining the coefficients of the MrDMD modes to align the model with current ocean conditions. Additionally, the parameter estimation process employs the Group Lasso method to achieve sparsification of the model, thereby enhancing computational efficiency. Moreover, to enhance prediction accuracy, deep reinforcement learning is introduced to optimize the sensing locations of the vehicles guided by the learned model. The integration of data-driven ocean prediction, online model learning, alongside optimization of sensing locations, forms a dynamic data-driven application system. Computer simulations demonstrate the effectiveness of the proposed methodology, enabling rapid and accurate data-driven ocean prediction using sparse sensing data from a marine vehicle network.