Analyzing fish detection and classification in IoT-based aquatic ecosystems through deep learning

Fish, a crucial source of protein, contribute approximately 17% of the global animal protein intake. In Malaysia, the fisheries sector serves as a significant economic driver, contributing around 11.22 billion Malaysian ringgit to the nation’s GDP in 2021. However, overfishing poses a substantial threat, leading to the depletion of marine life, disruption of ecosystems, potential food shortages, and unemployment. The advent of the Internet of Things (IoT) offers transformative potential for aquaculture, enhancing productivity, reducing waste, and promoting sustainability. This research underscores the viability of IoT-based smart aquaculture, focusing on different species and colours, such as red tilapia, black tilapia, and sailfin catfish. For monitoring aquaculture activities, the ESP32 Devkit is employed to collect data on temperature, dissolved oxygen, and pH levels, as well as to operate fish pellet feeders. Fish detection is facilitated using the NVIDIA Jetson Nano, an underwater camera, and various Darknet architectures within the You Only Look Once (YOLO) version, including YOLOv3, YOLOv3-Tiny, YOLOv4, and YOLOv4-Tiny. Future enhancements aim to monitor fish growth sizes, behaviour, and diseases, as well as identify waterborne pathogens, control pH, dissolved oxygen, and chlorine levels. The ongoing evolution of machine learning, deep learning, and transfer learning can facilitate the production of safe, high-quality, and abundant protein sources in well-regulated environments. This integration of technology into aquaculture signifies a promising step towards sustainable fisheries management, potentially mitigating the adverse effects of overfishing while ensuring the continued provision of essential protein sources. The study’s findings highlight the transformative potential of IoT and machine learning technologies in enhancing aquaculture productivity and sustainability. Among the tested models, YOLOv4-Tiny demonstrated the highest Average Precision (AP) and F1-score (approximately 1.00), making it the most suitable for real-time implementation due to its balance of accuracy and computational efficiency.