arXiv:2508.09971v1 Announce Type: cross Abstract: Vision-driven autonomous river following by Unmanned Aerial Vehicles is critical for applications such as rescue, surveillance, and environmental monitoring, particularly in dense riverine environments where GPS signals are unreliable. We formalize river following as a coverage control problem in which the reward function is submodular, yielding diminishing returns as more unique river segments are visited, thereby framing the task as a Submodular Markov Decision Process. First, we introduce Marginal Gain Advantage Estimation, which refines the reward advantage function by using a sliding window baseline computed from historical episodic returns, thus aligning the advantage estimation with the agent's evolving recognition of action value in non-Markovian settings. Second, we develop a Semantic Dynamics Model based on patchified water semantic masks that provides more interpretable and data-efficient short-term prediction of future observations compared to latent vision dynamics models. Third, we present the Constrained Actor Dynamics Estimator architecture, which integrates the actor, the cost estimator, and SDM for cost advantage estimation to form a model-based SafeRL framework capable of solving partially observable Constrained Submodular Markov Decision Processes. Simulation results demonstrate that MGAE achieves faster convergence and superior performance over traditional critic-based methods like Generalized Advantage Estimation. SDM provides more accurate short-term state predictions that enable the cost estimator to better predict potential violations. Overall, CADE effectively integrates safety regulation into model-based RL, with the Lagrangian approach achieving the soft balance of reward and safety during training, while the safety layer enhances performance during inference by hard action overlay.