Maritime transportation remains a critical and of the most efficient global transportation method. However, different forms of maritime accidents, such as ship collisions, still present a great problem. They result in catastrophic consequences, including loss of human lives, pollution and destruction of the ecosystem, and economic losses. Therefore, developing effective maritime risk assessment and management strategies is highly desirable to prevent such accidents and mitigate the consequences in the event they occur.
Numerous models have been proposed to predict ship collision risks. These models provide reliable information used in decision-making processes to improve maritime safety by preventing accidents and enhancing pollution preparedness and response. They are also useful in predicting trends in ship collisions and accident risks by identifying the risk spots and possible collision scenarios. With the advances in technology, automatic identifiable systems (AIS) have been of great benefit in detecting critical encounters and near misses in ship traffic data and estimating collision risks. Typically, the operation of the existing models can be categorized in two: those calculating the collision risk directly from the near-miss detection results and those relying on using algorithms to detect near misses from AIS data which are subjected to experts’ interpretation and judgment.
Relying on expert knowledge to interpret ship traffic data limits the practical applications of most existing approaches as it is prone to errors. Additionally, determining hotspots for accident risks and detecting the variation in the collision risk levels requires assessment and interpretation of a large number of near misses, which is time-consuming and expensive. To address these challenges, Professor Qing Liu from the University of Hamburg in collaboration with Professor Weibin Zhang and Mr. Xinyu Feng from Nanjing University of Science and Technology and Professor Floris Goerlangt from Dalhousie University presented a novel method for automatic interpretation and classification of ship collision risks. The main objective was to develop a cost-effective strategy for improving navigational risk assessment and management. The work is published in the journal, Reliability Engineering and System Safety.
In brief, the presented method relied on Convolutional Neural Networks (CNNs) to recognize and interpret, in terms of collision risks, the images obtained from the AIS data. The model’s predictive accuracy was investigated in the presence and absence of additional navigational information. The reliability of the obtained data was validated by comparing it to the estimates derived from training data. Furthermore, data from the Baltic Sea was used as a case study to test and validate the applicability of different design model alternatives.
Results showed that the CNN-based model automatically classified the ship collision risks at different scenarios by mimicking the purported expert judgments on the risk levels. Based on the training dataset and the case study, the CNN design model proved effective for fast and accurate interpretation of the traffic images and subsequent classification of the collision risk levels. Compared to the existing methods, this method was advantageous in reducing the resources required to interpret and predict the collision risk estimates. The results also confirmed that additional vessel navigational information and the traffic images indeed improved the model’s predictive accuracy.
In summary, a new method based on CNNs for the classification of ship collision risks from the collision encounters was reported. Based on the study findings, the CNN-based model met the specific design requirements for effective, fast, and accurate interpretation and prediction of the collision risks. Moreover, it also significantly reduced the resource requirement and enabled automatic ship collision risk classification. Therefore, this work contributes to developing a cost-effective method for identifying hot spot areas and detecting accident trends. Accordingly, Professor Qing Liu, the corresponding author explained that the proposed model provides a reliable route for future waterways risk analysis and management.
Zhang, W., Feng, X., Goerlandt, F., & Liu, Q. (2020). Towards a Convolutional Neural Network model for classifying regional ship collision risk levels for waterway risk analysis. Reliability Engineering & System Safety, 204, 107127.