AIoT platform integrated with smart breeding and omics cloud database for Taiwan tilapia

The ongoing depletion of marine fishery products has heightened the importance of aquaculture, with farming tilapia emerging as a key substitute for protein sources. Conventional breeding approaches for Taiwan tilapia are often protracted in duration and susceptible to genetic deterioration. To address these challenges, we propose a multi-year research project that aims to build a platform integrated with a smart breeding system and an omics cloud. State-of-the-art artificial intelligence (AI) models and high-throughput IoT are employed to offer promising solutions. Global advancements in selective breeding further highlight the significance of this work in sustaining Taiwan's tilapia industry. Funded by Taiwan's Ministry of Agriculture, the research focuses on exploring genotype-phenotype relationships to pave the way for more efficient breeding procedures and, ultimately, optimal breeding strategies. This paper reports the first-year results of phenotype work. Novel algorithms are developed to integrate with the AI models, enabling real-time monitoring of fish behavior and environmental parameters, thereby facilitating selective breeding procedures. Key findings include: (1) for fish aged 1–3 months, YOLO11s achieved 0.802 of mAP @50 and 0.9316 of mAP @50 for body detection; (2) for 3–5 months fish, YOLOv12n outperformed others with 0.767 of mAP @50 ; (3) for anomaly behavior recognition, F1-score and precision rate reached 97 % and 99 %, respectively; and (4) size measurement using stereo vision yielded an error rate of 3.42 %. All computing tasks, including object detection, behavior recognition, and size measurement, can be performed on the Jetson Xavier series, enabling cost-effective edge computing for on-site aquaculture applications. • The AIoT omics cloud database platform enhances breeding efficiency and sustainability for Taiwan tilapia. • Can detect fish of 1–3 and 3–5 months with mAP @50 = 0.802 and 0.767, respectively. • For anomaly behavior recognition, F1-score can reach 99%; for the size measurement, it achieves 3.42% in error rate. • All AI tasks are executable on edge computing devices (e.g., Jetson Xavier, 39.6 FPS). • As farming production exceeds capture fisheries, the study of relationships between genotype and phenotypic traits.