Deep Learning-Based Enhanced Cluster Head Selection for Underwater Wireless Sensor Networks

Underwater wireless sensor networks (UWSNs) are subject to unique operational challenges, including constrained energy availability, dynamic topological structures, and unreliable acoustic communication. These factors significantly impact the efficiency and stability of data collection processes, particularly the selection of cluster heads (CHs), which plays a vital role in prolonging network functionality. This paper presents GAT-CHS, a cluster head selection algorithm that integrates graph-based attention mechanisms with deep reinforcement learning to adaptively optimize clustering decisions in underwater environments. The proposed approach encodes critical node attributes into a spatial–topological representation, applies multi-head attention to quantify inter-node relevance, and utilizes a Deep Q-Network (DQN) to determine CH roles based on long-term network performance. The algorithm is evaluated across a range of simulated UWSN scenarios reflecting varying node distributions and environmental conditions. Results show that GAT-CHS reduces energy usage by 24%, improves network longevity by 36%, and achieves a packet delivery ratio of 98.6%. These findings underscore the model’s robustness and scalability, establishing GAT-CHS as a promising direction for next-generation clustering in complex underwater sensor deployments.