Lightweight Deep Learning Model for Real-Time Anomaly Detection in Underwater Drone Surveillance Systems

Abstract

Maritime ecosystems face mounting pressure from human activities and environmental change, underscoring the urgent need for reliable underwater surveillance to safeguard biodiversity, critical infrastructure, and maritime security. Existing deep learning-based anomaly detection methods often depend on centralized processing, high computational resources, single-modality assumptions, and limited interpretability, which restrict their suitability for real-time underwater drone surveillance under noisy and bandwidth-constrained marine conditions. To address these limitations, we introduce QD2Net-XQ, a lightweight dual-path framework with federated learning and quantum-inspired hyperparameter optimization for multi-label underwater anomaly detection. The model applies standard sinusoidal encoding and Shapley-guided feature selection before parallel convolutional and recurrent-diffusion processing of local and long-range dependencies. An adaptive fusion mechanism combines the resulting representations, while the optimizer selects a hyperparameter configuration that balances predictive accuracy and computational efficiency. Evaluated on a large-scale public underwater-drone telemetry dataset, QD2Net-XQ achieves 98.6% accuracy, 98.3% precision, and a 96.9% PAC score, while requiring 16.4M parameters and having a $2PT$ complexity estimate of 0.984 G for a sequence length of $T=30$ . These results demonstrate its ability to outperform state-of-the-art baselines while maintaining fairness across clients. Overall, QD2Net-XQ offers a practical, explainable, and energy-efficient solution for next-generation maritime surveillance.