Enhancing Ammonia Concentration Prediction with a Transfer-Learning-Based Model: Application in a Pig Farm.

Area of Science:

• Agricultural Engineering
• Environmental Science
• Artificial Intelligence

Background:

• Intensification of swine farming raises concerns about ammonia (NH3) emissions.
• Smart farming technologies necessitate reliable NH3 monitoring without solely relying on expensive sensors.

Purpose of the Study:

• To develop an AI-based model for predicting NH3 concentrations in commercial pig houses.
• To evaluate the impact of data collection intervals and learning strategies on prediction accuracy.
• To compare standalone models with transfer learning models for NH3 prediction.

Main Methods:

• Development of an artificial intelligence prediction model for NH3 concentration.
• Comparison of standalone models trained on local data versus transfer learning models.
• Evaluation across various data collection intervals (10, 20, 30, 60 min) using Random Forest and XGBoost algorithms.

Main Results:

• Transfer learning models consistently outperformed standalone models across all tested data collection intervals.
• The best Random Forest and XGBoost models achieved high accuracy, with R2 of 0.969, RMSE ~1.0 ppm, and MAPE <5%.
• Accurate NH3 predictions were achieved even with sparse data using transfer learning.

Conclusions:

• Transfer learning offers a robust and data-efficient method for predicting NH3 concentrations in swine housing.
• This AI-driven approach supports sustainable and improved environmental management in the swine industry.
• The findings facilitate the integration of advanced monitoring in smart farming environments.