Machine Learning on Low-Cost Edge Devices for Real-Time Water Quality Prediction in Tilapia Aquaculture.

This study presents the deployment of Machine Learning (ML) models on low-cost edge devices (ESP32) for real-time water quality prediction in tilapia aquaculture. A compact monitoring and control system was developed with low-cost sensors to capture key environmental parameters under field conditions in Northern Thailand. Three ML models-Multiple Linear Regression (MLR), Decision Tree Regression (DTR), and Random Forest Regression (RFR)-were evaluated. RFR achieved the highest accuracy (R2 > 0.80), while MLR, with moderate performance (R2 ≈ 0.65-0.72), was identified as the most practical choice for ESP32 deployment due to its computational efficiency and offline operability. The system integrates sensing, prediction, and actuation, enabling autonomous regulation of dissolved oxygen and pH without constant cloud connectivity. Field validation demonstrated the system's ability to maintain DO within biologically safe ranges and stabilize pH within an hour, supporting fish health and reducing production risks. These findings underline the potential of Edge AIoT as a scalable solution for small-scale aquaculture in resource-limited contexts. Future work will expand seasonal data coverage, explore federated learning approaches, and include economic assessments to ensure long-term robustness and sustainability.