The surveillance of illegal fishing activities is a critical issue for the management of marine resources. Here, we address the monitoring of fishing activities from space, namely from VMS data (Vessel Monitoring System). We propose and evaluate a novel method for the recognition of the fishing vessel gear type from VMS trajectories as a mean of detecting abnormal uses of undeclared fishing gear. Our approach combines supervised and unsupervised data mining techniques. It consists of a three-step framework: i) for each gear type, we first train a Gaussian Mixture Model (GMM) for the unsupervised identification of the relevant regimes from the series of the speed and turning angle computed from the VMS data, ii) for any given VMS data series, we apply the fitted gear-specific GMMs to extract some associated regime related features (e.g., time spent in each regime, mean duration of each regime period), iii) we combine the resulting feature vector with other relevant features (e.g., mean position and sinuosity index) within a supervised classification framework and we train a supervised classified, namely RF (Random Forest) and SVM (Support Vector Machine), for the automated recognition of the fishing gear from monthly VMS data. Overall, we report mean correct recognition rates around 94.59%, which demonstrates the relevance of the proposed approach.