The surrounding waters of Taiwan are evaluated as a moderate risk environment by Casualty Return, Lloyd’s Registry of Shipping. Among all types of maritime accidents, ship collisions occur most often, which has severe consequences, including ship damage, sinking and death of crews, and destruction of marine environments. It is, therefore, imperative to mitigate the risk of ship collision by exploring the risk factors and then providing preventive measures. This study invited domain experts to form a decision-making group, which helped with the risk assessment. The initial set of risk factors was selected from the literature. The expert group then identified seven representative risk factors using rough set theory (RST). The researchers worked with the experts to delineate the diagram of a bow-tie analysis (BTA), which provided the causes, consequences, and preventive and mitigation measures for ship collision incidents. The results show an integrated research framework for the risk assessment of ship collision that can effectively identify key factors and associated managerial strategies to improve navigation safety, leading to a sound marine environment.

The rise of marine recreational activities has led to a growing use of marine recreational powered platforms, raising safety concerns related to navigation. In Taiwan, the current regulatory system for such platforms remains fragmented and under debate. This study aims to support policy development by identifying key safety management priorities. This study utilized the four core components of the SHELL model, which include Software, Hardware, Environment, and Liveware, as the analytical foundation and identified 20 preliminary safety criteria through an extensive review of relevant literature. A Modified Delphi Method and DEMATEL analysis were applied to gather expert insights and prioritize 10 representative indicators. The resulting Influence Network Relation Map revealed that “Comprehensive Management Regulations” had the highest causal influence across all dimensions. Additionally, “Basic Navigation Concepts” and “Emergency Response and Safety Knowledge” were found to be the most central elements. Based on these findings, the study recommends targeted measures including enhanced regulation, improved training, radar monitoring, and spatial planning to reduce navigation risks and promote safer marine recreation. Building on the above findings, this study confirms the effectiveness of an innovative integration of the SHELL model and the DEMATEL method, which provides a structured and adaptive framework capable of systematically identifying systemic navigational risks in marine recreational activities.

Accurate multi-output forecasting for maritime navigation requires models that are robust to heterogeneous noise, outliers, and uncertainty miscalibration commonly present in Automatic Identification System (AIS) and meteorological–oceanographic (metocean) data. Conventional loss formulations and fixed weighting strategies often fail to balance inter-target errors and probabilistic reliability across multiple outputs. This paper presents a robust multi-output forecasting framework based on a Transformer encoder equipped with a per-target weighted Huber loss. To ensure stability and avoid manual tuning, an entropy-regularized Lagrangian formulation is first derived to obtain a closed-form solution for adaptive target weights, providing a deterministic and reproducible baseline. The per-target Huber thresholds are then optimized using a hybrid Grey Wolf Optimizer–Simulated Annealing (GWO–SA) strategy, which combines global exploration with stochastic local refinement on a validation objective. To further address regarding residual miscalibration, post-hoc calibration via temperature scaling and isotonic regression is applied to the predictive distributions. The key contribution of this work lies in the integration of a closed-form entropy-regularized weighting scheme with hybrid metaheuristic threshold optimization and post-hoc calibration. This integration enabling simultaneous improvements in robustness, inter-target balance, and probabilistic calibration for multi-output sequence forecasting. Comprehensive experiments on a multimodal AIS–metocean dataset show that the proposed approach consistently outperforms Gaussian NLL, fixed Huber loss, and single-metaheuristic baselines. The method can achieve lower MAE and MSE and significantly reducing NLL, CRPS, and Expected Calibration Error so can be resulting in accurate and well-calibrated predictive distributions. These experiment results demonstrate the suitability of the proposed framework for uncertainty-aware and safety-critical maritime decision-support applications.

For safety at sea, navigation, and traffic management, it is important to be able to accurately forecast the path of a ship. This research introduces a comprehensive multi-output deep learning model incorporating an attention mechanism for the simultaneous prediction of vessel position, course, and speed utilizing Automatic Identification System (AIS) data. The architecture merges bidirectional gated recurrent units (BiGRU) with the Hybrid Grey Wolf Optimizer–Simulated Annealing (GWO-SA) for loss-weight tuning and the Analytic Hierarchy Process (AHP) for prioritizing. Tests using real AIS data from Taiwan’s territorial waters show that the suggested BiGRU-AM-GWO-SA-AHP outperforms LSTM, GRU, and BiLSTM baselines in terms of spatial accuracy (latitude MAE 0.1798; mean Haversine distance 64.41 km), stable course prediction (cos(cog) RMSE 0.2747), and high speed prediction accuracy (SOG R² 0.9932). The results show that combining attention-based bidirectional recurrent modeling with hybrid optimization is a good way to make strong, real-time predictions about maritime trajectories.