Integrating digital twin technology with deep reinforcement learning for sustainable marine fishery resource...

Marine fishery ecosystems face unprecedented pressure from overfishing and climate variability, and these mounting threats call for management tools that go beyond traditional static quota systems. This paper puts forward an integrated framework that couples digital twin (DT) technology with deep reinforcement learning (DRL) to tackle sustainable fishery resource management. We build a five-layer hierarchical architecture whose centerpiece is a high-fidelity digital twin that mirrors fishery dynamics through explicit state-transition and observation equations rather than abstract placeholders. A Proximal Policy Optimization (PPO) agent operates within this simulated environment, receiving multidimensional state inputs-resource stocks, oceanographic conditions, fleet operations-and optimizing a composite reward function whose weights we set to [Formula: see text] (economic), [Formula: see text] (ecological), and [Formula: see text] (sustainability). We conduct both comparative experiments and ablation studies using East China Sea fishery data spanning 2010-2023. The ablation study, which isolates the digital twin contribution by comparing PPO with and without DT integration, confirms that the DT alone accounts for a 31.9% reward improvement. Overall, our method achieves a resource recovery index (RRI) of 0.83, outperforming traditional maximum sustainable yield management by 97.6% and standard deep Q-networks by 36.1%. Spatial heatmaps and temporal effort-control time series generated from the learned policy reveal ecologically sensible seasonal and spatial harvest patterns. This research establishes a virtual-real fusion paradigm for intelligent fishery governance and provides decision-support tools for sustainability challenges in an era of accelerating environmental change.