Timely prediction of maritime casualties resulting in pollution occurrence remains unsolved in academia, as the significant data imbalance between non-polluting and polluting casualties poses a challenge to prediction efficacy. This study proposes an ensemble method for predicting polluting maritime casualties and exploring the contributing features to pollution. In the data preprocessing phase, key features related to casualties and vessels are extracted and encoded into model variables; in the data augmentation phase, Variational Autoencoder is employed to generate synthetic samples from the minor class, effectively mitigating the impact from data imbalance; and in the pollution indicator classification phase, machine learning models are trained on the balanced dataset to label a casualty as “polluting” or “non-polluting”. A dataset containing 25,414 worldwide maritime casualties from 2013 to 2023 is utilized for method validation. Several state-of-the-art data balancing techniques serve as baselines for comparison with the VAE on the quality of generated synthetic data. The model trained on the VAE dataset achieves the most satisfactory performances, demonstrating the superiority of VAE in augmenting data quantity and diversity. “Casualty cause”, “Vessel age” and “Vessel type” are revealed as the top three contributing features to pollution. Several insights are discussed for precautionary measures and policy development.