Edge AI-Powered Predictive Water-Quality Monitoring and Early Warning for Smart Aquaculture

Edge AI-Powered Predictive Water-Quality Monitoring and Early Warning for Smart Aquaculture 1Rishu Raj, 2Pryobroto Sarkar, 3Ritwik Kapat, 4Riya Panda, 5Debjyoti Adhikari, 6Ritwik Haldar 1,2,3,4UG Students, Dept. of Electronics and Communication Engineering, Haldia Institute of Technology, Haldia, Purba Medinipur, West Bengal, India 6Assistant Professor, Dept. of Information Technology, Haldia Institute of Technology, Haldia, Purba Medinipur, West Bengal, India 6Associate Professor, Dept. of Electronics and Communication Engineering, Haldia Institute of Technology, Haldia, Purba Medinipur, West Bengal, India ABSTRACT Water quality in aquaculture ponds varies with environmental conditions, biological activity, and daily cycles, so periodic manual measurement is inherently reactive and detects unfavourable conditions only after they have developed. This work presents an Edge Artificial Intelligence (Edge AI) predictive monitoring system for smart fish ponds that forecasts water temperature and pH and provides early risk indication without cloud-based inference. A real-world 24-hour pond dataset is processed by converting timestamps into a decimal hour-of-day feature. Decision Tree Regression and a Long Short-Term Memory (LSTM) network are trained and compared: the Decision Tree uses hour-of-day as its predictor, while the LSTM uses normalised temperature sequences as a temporal benchmark. For temperature, both models reproduce the diurnal pattern with almost identical accuracy, the Decision Tree achieving a coefficient of determination of 0.9989 and the LSTM 0.9987, with the LSTM marginally lower in absolute error. The Decision Tree is therefore selected not on accuracy but on deployability, since it trains in under a second, occupies a few tens of kilobytes, and executes on a microcontroller, whereas the LSTM could not be deployed on the target device. The selected models are converted to C+⁠+ using micromlgen and deployed on an ESP32, which synchronises time over NTP, performs local inference in under one millisecond, and displays four-hour forecasts on a 16 × 2 I2C LCD. A threshold-based mechanism issues early warnings when predicted values fall outside predefined safe ranges. The prototype demonstrates a compact TinyML approach to predictive aquaculture monitoring on inexpensive embedded hardware.