Abstract
Accurate identification of fishing vessel operations is vital for sustainable fishery management. Existing methods inadequately exploit spatiotemporal contextual information in vessel trajectories and fail to effectively fuse multimodal data. To address this, this study proposes a novel framework integrating Geohash-based geocoding with embedding techniques inspired by natural language processing to extract spatiotemporal features from trajectory sequences. We develop a multi-branch 1D convolutional neural network (MB-1dCNN) to minimize feature engineering dependency while enhancing operational-type recognition. Comparative experiments evaluate Geohash encoding lengths and network architectures (single-branch vs. multi-branch, fully-connected vs. 1D-CNN). Results indicate optimal Geohash encoding at length 5. The multi-branch structure significantly outperforms single-branch counterparts, and MB-1dCNN demonstrates superior performance over multi-branch model with fully connected layers (MB-FCNN), achieving additional gains in accuracy and F1-score. Key findings reveal: (1) 1D-CNN processing surpasses fully-connected networks in sequential feature extraction, (2) Multi-branch architectures enhance information fusion capabilities. The proposed MB-1dCNN establishes state-of-the-art performance for trajectory-based fishing operation recognition, offering valuable insights for spatial computing applications in maritime surveillance.