The deployment of unmanned surface vehicles (USVs) for autonomous maritime operations requires robust target tracking based on multimodal marine sensing under stochastic ocean environments. However, conventional tracking methods suffer from high-dimensional sensor noise and limitations in onboard sensing resources and the computational burden of multiagent coordination. We developed the dynamic marine attention network (DyMAN), a sensing-oriented multiagent deep reinforcement learning framework integrated with an ensemble hard attention mechanism and game theory. By utilizing a centralized critic and distributed actor architecture, DyMAN mitigates multiagent nonstationarity while ensuring decentralized execution. To address the limitations of onboard hardware, a hard attention mechanism is integrated to prioritize informative sensor data, effectively filtering environmental noise from light detection and ranging and radio detection and ranging data, and improving the efficiency of multimodal sensor data processing by focusing computational resources on important targets. The Nash equilibrium is also integrated to ensure stable cooperative sensing and coordination and reduce decision conflicts among the USVs in the fleet. The experimental results demonstrate that DyMAN outperforms baseline algorithms, including multiagent deep deterministic policy gradient, deep deterministic policy gradient, and proximal policy optimization in terms of cumulative reward and convergence stability. DyMAN maintains a stabilized minimum interagent distance of approximately 40 units, which is a 50% improvement in formation stability compared with that in early training, and significantly reduces tracking error and energy consumption. The results of DyMAN provide a computational framework for distributed marine sensing nodes, enhancing the reliability and efficiency of multimodal sensing systems in complex maritime environments.