A Closed-Loop Agentic Framework for Marine Litter Monitoring: Data Synthesis, Augmentation, and Performance-Guided Optimization

Data scarcity remains a major obstacle to robust computer vision for marine litter monitoring, especially for emerged litter detection and temporal tracking in UAV-based imagery. This paper introduces a closed-loop agentic framework that autonomously generates, annotates, and refines synthetic training data to improve downstream detection and tracking performance. The system is supervised by either a remote or local reasoning agent, which coordinates data generation, annotation, training, and evaluation within an iterative optimization loop. The framework combines SDXL Inpainting for foreground object synthesis and LTX-2.3 for high-fidelity temporal sequence generation. Automated annotations are produced using FastSAM and ByteTrack, providing mask and tracking ground-truth labels suitable for supervised training. The synthesized data are used to train and evaluate YOLOv8 Nano models for marine litter detection and multi-object tracking. Performance metrics, including mAP and MOTA, are fed back to the agent to refine prompts, generation settings, and training configurations. Experimental results show that the proposed workflow improves performance, highlighting the potential of agentic data synthesis and augmentation for automated marine litter monitoring in data-scarce scenarios.