The human-in-the-loop workflow for recognizing behavioural patterns in vessel trajectories combines feature engineering with interactive visual analytics to build classification models that are both flexible and tolerant to data noise. The workflow partitions continuous vessel trajectories into overlapping episodes of appropriate duration (3-hour windows with 1-hour shifts), then derives interval-based synoptic features capturing behavioural aspects such as speed levels, movement curvature, and spatial context (e.g., distance to ports) that distinguish trawling activities from other movements. Through iterative visual exploration, domain experts refine the feature space by evaluating whether episodes clustered by feature similarity exhibit coherent behavioural patterns, adjusting or adding features (such as applying logarithmic transformations to handle skewed distributions) until groups are interpretable and well-separated. A dimensionality reduction projection (UMAP) enables simultaneous visual inspection of feature similarity across all episodes, while coordinated views showing trajectory shapes on geographic maps and feature distributions in histograms allow experts to efficiently examine and label cluster cores as representative examples of different pattern types. This iterative refinement process—cycling between feature engineering, clustering, visualization, pattern interpretation, and quality assessment—continues until labelled examples enable reliable automated classification, as validated through k-nearest neighbour testing. The approach not only successfully identified expected trawling patterns but also revealed three distinct trawling subtypes (wide curves, tight loops, and straight movements with 180-degree turns) that would have remained hidden without the flexibility and discovery capabilities of the human-in-the-loop methodology.