UAV-Enabled ISAC System for Fish School Detection Using Vision Transformer and Deep Reinforcement Learning

This paper proposes an integrated low Earth orbit (LEO) satellite and unmanned aerial vehicle (UAV)-enabled integrated sensing and communication (ISAC) framework that exploits the dual functionality of sensing and communication, enabling efficient detection of fish school zones and real-time data transmission to fishing vessel users in marine environments. This study focuses on two key components within the proposed framework: first, UAVs equipped with sensing radar perform fish school detection in marine environments by employing a region-based Vision Transformer (ViT) algorithm, using echo signals that are converted into range-doppler maps (RDMs). Second, to enhance system performance by addressing the challenges posed by UAV mobility and the need for adaptive beamforming, we formulate a joint optimization problem for UAV trajectory and beamforming. This problem is addressed using the proximal policy optimization (PPO) algorithm combined with the ViT model referred to as PPO-ViT, which incorporates the ViT detection results as part of the state input to guide optimal UAV trajectory planning and beamforming strategy. Simulation results demonstrate that the ViT-based detection achieves superior accuracy compared to other CNN models, while the PPO-ViT framework significantly improves average data rate and sensing precision over baseline methods.