A Hybrid AIoT Multi Agent Framework for Autonomous Underwater Exploration and Search and Rescue
Journal of Intelligent Decision Making and Information Science · 09.09.2026 · 3 (18s) · 3197-3224
DOI: 10.17762/jidmis.v3.6080 ↗Kurz gefasst
Fachartikel in der Zeitschrift Journal of Intelligent Decision Making and Information Science, online erschienen am 09.09.2026. Verfasst von Dhaker Boubaker Mnari. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
Underwater object detection remains a challenging problem due to poor visibility, light scattering, and highly dynamic environmental conditions, which significantly degrade the performance of conventional computer vision systems. While recent deep learning-based detectors such as YOLO and SSD have improved detection accuracy, most existing approaches focus on isolated model optimization and neglect system-level intelligence, sensor integration, and real-time coordination in real-world underwater environments. This paper proposes a hybrid AIoT multi-agent framework that integrates IoT sensing, deep learning-based perception, and large language model (LLM)-driven orchestration for robust underwater object detection and decision-making. The system relies on an MQTT-based communication layer to enable lightweight and scalable data exchange between distributed underwater sensors and edge computing units. A data fusion module is introduced to integrate heterogeneous inputs, including visual data and environmental sensor information, enhancing robustness under degraded conditions. At the core of the proposed framework, a LangGraph-based multi-agent orchestration mechanism coordinates specialized agents responsible for preprocessing, perception, data fusion, and decision-making. The perception module is implemented using a custom YOLOv11 object detection model, and its performance is benchmarked against SSD and YOLOv8 under identical experimental conditions. A lightweight LLM (e.g., Qwen or Phi-mini) is employed as a reasoning agent to dynamically manage workflow execution and interpret detection outputs. In addition, the system incorporates Langfuse-based observability, enabling detailed tracking of agent interactions, LLM calls, and execution traces for reproducibility and performance analysis. The study also investigates the impact of underwater image enhancement techniques on detection performance, revealing that the proposed model remains robust even without preprocessing enhancements. Experimental results on a dataset of 13,000 underwater images demonstrate that the proposed YOLOv11-based framework achieves superior performance in terms of Precision, Recall, mAP@50, and mAP@50-95 compared to baseline models. Furthermore, results indicate that image enhancement does not significantly improve detection accuracy, highlighting the inherent robustness of the proposed AIoT multi-agent architecture.
Zitieren
APA
Mnari, D. B. (2026). A Hybrid AIoT Multi Agent Framework for Autonomous Underwater Exploration and Search and Rescue. Journal of Intelligent Decision Making and Information Science, 3(18s), 3197-3224. https://doi.org/10.17762/jidmis.v3.6080
Vancouver
Mnari DB. A Hybrid AIoT Multi Agent Framework for Autonomous Underwater Exploration and Search and Rescue. Journal of Intelligent Decision Making and Information Science. 2026;3(18s):3197-3224. doi:10.17762/jidmis.v3.6080
BibTeX
@article{Mnari2026,
author = {Mnari, Dhaker Boubaker},
title = {A Hybrid AIoT Multi Agent Framework for Autonomous Underwater Exploration and Search and Rescue},
journal = {Journal of Intelligent Decision Making and Information Science},
year = {2026},
volume = {3},
number = {18s},
pages = {3197--3224},
doi = {10.17762/jidmis.v3.6080}
}
EUVIMED