Real-Time Pond Water Assessment via Embedded Deep Learning and Visual Data Acquisition: A Practical Monitoring Approach for Aquaculture

Maintaining optimal water quality is essential in aquaculture, particularly for ornamental species such as Kois, where color clarity and environmental balance are directly linked to health and aesthetics. This article presents a complete protocol for designing, deploying, and operating a real-time water quality monitoring system based on underwater image classification using embedded deep learning. The proposed system integrates a low-cost underwater camera with a Raspberry Pi-based controller network in an autonomous water surface robot, enabling the capture, transmission, and classification of aquatic images through a lightweight neural network model. Five water condition categories -- ranging from clear to turbid -- are visually distinguished, with classification accuracy exceeding 99% on a custom dataset. The system is configured to run autonomously, offering continuous assessment and visual alerts for deteriorating conditions. This visual protocol covers device assembly, system calibration, model deployment, and real-time monitoring demonstrations. The method supports sustainable aquaculture practices by enabling non-invasive, continuous evaluation of aquatic environments using affordable, scalable technologies. The approach can be adapted to other aquaculture or environmental monitoring contexts with minimal modification.