Nighttime port safety surveillance and management via a policy-aware visual deep reinforcement learning approach

The maritime industry is undergoing rapid digital transformation, yet maintaining safety oversight under adverse conditions, particularly at night, remains a persistent management challenge. Traditional nighttime supervision relies heavily on manual patrols and passive sensor monitoring, which can delay the identification and review of possible safety violations. This study proposes a policy-aware visual reinforcement learning approach that translates port rules, including spatial-zone restrictions and safety-equipment requirements, into quantifiable training objectives. In the controlled port-like dataset examined here, the approach identified safety violations under low-visibility conditions and maintained strong performance with few training examples. These findings demonstrate technical feasibility under the tested conditions; they do not establish deployment readiness or governance effectiveness in an operating terminal. If confirmed through staged real-port validation, the observed data efficiency may reduce site-specific data-labelling and adaptation burdens. Operational value would also depend on human review, workforce acceptance, liability allocation, false-alert management, and regulatory oversight.