Ammonia Detection and Prediction in a Poultry Farm using Machine Learning

The concentration of ammonia within poultry sheds is a constant risk to flock health, productivity, and profitability, but most farms only make the occasional manual measurement that does little to provide a predictive sense. We created a low-cost monitoring platform that integrates Internet of Things (IoT) sensing and machine-learning analytics to enable real-time, predictive ammonia management. The sensing layer consists of two ESP32 microcontrollers: each controller combines an MQ-135 gas sensor (for ammonia) with a BME280 sensor that measures temperature, humidity, and barometric pressure. Readings are uploaded to the ThingSpeak cloud and stored, visualised, and used to feed predictive models every 30 seconds. The ammonia forecasting model was an ARIMA model used to forecast ammonia up to 2 hours ahead, and the image classifier was a CNN, with captured flock photos used to assess whether the flock is diseased. The sensors recorded maximum ammonia concentrations of 14.55 ppm and 16.64 ppm in live farm trials; the ARIMA model maintained a prediction error of ±2.0 ppm for a 30-min prediction horizon, and the CNN model predicted Coccidiosis with 99.72% confidence. As a whole, this evidence supports the conclusion that the platform provides consistent and meaningful intelligence that can be used to take proactive steps to control the environment in poultry facilities of any size.