Explainable Machine Learning Framework for Predicting Yield and Plant Health in Controlled Environment Agriculture

Abstract

Controlled Environment Agriculture (CEA) has become an effective solution in order to produce crops sustainably by controlling the environment parameters like temperature, humidity, light intensity, nutrient concentration, and carbon dioxide level. Forecast the yield and monitor the health of the plants on real time basis remain as a challenge as there are complex interactions in environmental and plant conditions. The use of machine learning (ML) models in agricultural predictions has demonstrated positive results, but a number of existing solutions are a black box solution with very little interpretability, which has limited the adoption of the model and the trust of the user. In this study, an Explainable Machine Learning (XML) framework is proposed to predict the crop yield and health in CEA systems. It incorporates environmental data from sensors in the IoT network, along with plant images (RGB, multispectral and landmarks) captured by various camera types. Various ML Models are being used for predictive analytics like Random Forest, XGBoost, Support Vector Machine, and Long Short-Term Memory networks. For transparency, the SHAP values, LIME and feature importance methods are included when interpreting the model. Experimental outcomes show good predictability, interpretability and decision support for smart horticulture applications.