Ships play a critical and indispensable role in various human activities at sea. Thanks to the extensive adoption of the Automatic Identification System (AIS) in the past two decades, large amounts of ship trajectory data are being collected over time continuously. These data record detailed ship movement and thus support many significant applications, such as trajectory prediction, collision avoidance, and maritime traffic surveillance.

However, despite its wide application value, analyzing ship trajectories still faces challenges. Specifically, this thesis focuses on three challenging tasks. The first task tackles the challenge of degraded AIS data quality caused by large gaps which can last up to hours or even days. The second and third tasks concern environmental applications of AIS data and work on two challenges respectively: detection of complete fishing activities and large-scale experimental comparison of shipping CO2 estimation methods.

First, we address the problem of imputing large AIS gaps. The main limitation of existing trajectory imputation methods lies in the lack of temporal information in the imputed trajectories. In light of this, we resort to an approach based on multiple coastal cameras. We propose MCbSLE, an algorithm that estimates ship locations as a set of spatial polygons. In this way, MCbSLE takes into account the uncertainty in conversion from pixel coordinates to longitude/latitude coordinates. Built on top of MCbSLE, we propose TrajImpMC, a tracking-based trajectory imputation framework for filling large AIS gaps. TrajImpMC combines speed constraints and Kalman filters, and can return imputed trajectories that contain both spatial and temporal information. Extensive experiments on real datasets demonstrate the effectiveness of TrajImpMC in handling large AIS gaps and complex ship movement.

Second, we study the problem of detecting complete fishing activities from ship trajectories. Existing trajectory segmentation methods fall short when dealing with the diverse ship movement patterns during fishing, which differ greatly from one gear type to another. Moreover, their segmentation results usually lack label information indicating whether a ship is fishing or not. We first propose TPoSTE, a visualization-based technique to help design features distinguishing fishing movement patterns from non-fishing movement patterns. TPoSTE works by plotting occurrences of different spatiotemporal events of interest in parallel. Next, we propose WBS-RLE, a novel window-based trajectory segmentation algorithm. WBS-RLE uses a classifier to label each trajectory window as fishing or non-fishing, and then employs run-length encoding to merge close fishing trajectory windows into one complete fishing activity. Experimental results on real datasets show that WBS-RLE is able to return segmentation results that are more semantically meaningful than existing trajectory segmentation methods.

Third, we work on the research gap that a large-scale experimental comparison of shipping CO2 estimation methods is lacking in the literature. To this end, we propose a general data-driven evaluation framework. This framework is based on data integration of three data sources: ship technical data, AIS trajectory data, and weather data. Together with emission models, these data are fed into three novel modules that perform analysis at both grid and trajectory levels as well as use annually aggregated MRV data as ground truth. Extensive experiments are conducted on one-month Danish AIS data to demonstrate the utility of the framework, and insights into five popular shipping CO2 emissions models are presented.

Put together, these contributions advance ship trajectory analytics by offering better AIS data quality, accurate fishing activity detection, and rich trajectory-based analysis of shipping CO2 emissions.