Multi-scale collision risk assessment in restricted waters considering ship trajectory uncertainty

A comprehensive ship collision risk assessment is essential for intelligent supervision in restricted waterways. This study proposes a multi-scale risk assessment framework that incorporates trajectory uncertainty. At the micro scale, a Long Short-Term Memory network combined with an Unscented Kalman Filter (LSTM-UKF) is integrated into the Degree of Domain Violation (DDV) model to predict vessel trajectories and identify high-risk multi-ship encounter clusters. At the macro scale, Extreme Value Theory (EVT) is applied to the tail distribution of DDV to estimate the probability and frequency of collision events. The effectiveness of the proposed methodology is validated using AIS data collected from the Huangpu River. Results show that the proposed uncertainty-aware method significantly enhances the effectiveness of potential collision risk detection. In the Huangpu River case study, the estimated collision probability is 0.21 %, with an annual frequency of 11.54 incidents. Collision hotspots are primarily located near bends and turning zones, with cargo and passenger ships, particularly medium-sized vessels, exhibiting the highest risks. Additionally, the collision probability during high tide is 1.56 times higher than during non-high tide. These findings provide valuable support for refined traffic management and targeted risk mitigation in restricted waterways.