Integrating Ensemble Learning and Explainable AI in a Smart Agriculture Information Technology Platform for Improved Fruit Quality Prediction

Agricultural industry has been challenged by the demands to produce high quality products with less waste and better utilization of limited resources. This research aims to establish a new platform with an ensemble learning way of thinking combined with XAI (Explainable Artificial Intelligence) on smart agriculture information technology platform to conduct fruit quality prediction. In this research, by using 4,000 apple samples with eight physiochemical properties (Size Weight sweet, crunch Juicy ripe acid good/bad), we design and compare several machine learning models. The combination of RF, GBDT and XGBoost classifiers exhibit the best prediction accuracies of 94.2%, over 89.7% from single models. More importantly, SHAP and LIME are utilized for the first time in agriculture experiments to reveal the underlying reason of model's decisions, Therefore, the physiochemical factors that relate to quality prediction can be derived. With interpretability, this system can effectively assist production, sorting and supply chain management by data-driven quality grading decisions. The system exhibited the feasibility to be implemented in practice and could decrease manual inspection costs by 40%.