Journal of The Korean Society of Living Environmental System ISSN:1226-1289(Print) 한국생활환경학회

This study developed an agricultural simulator that integrates a digital twin and a metaverse-based 3D virtual environment, and quantitatively verified its effectiveness by applying an XGBoost (Extreme Gradient Boosting)-based regression prediction model. Existing machine learning (ML)-based agricultural prediction research has tended to focus on improving prediction accuracy, but there has been a relative lack of a decision support structure that integrates synchronization with actual operating environments, scenario-based repeated experiments, and user interaction. Therefore, this study proposes an “agricultural metaverse simulator” that combines (1) building a digital twin model based on actual greenhouse data, (2) real-time visualization and interaction in a metaverse environment, and (3) integration of an XGBoost-based growth prediction model. The data were collected from the Rural Development Administration’s publicly available tomato data and data from three experimental greenhouses, and a total of 18 variables were configured: seven inputs (environment/control) and 11 outputs (growth/production). Preprocessing ensured reproducibility by specifying unit unification, missing value handling, and outlier correction (IQR: Interquartile Range) criteria. Model performance was evaluated using Root Mean Squared Error (RMSE) and Coefficient of Determination (R²), and results were presented by greenhouse unit (farm_code). Furthermore, the effectiveness of the simulator was rigorously verified through statistical significance tests (paired-samples t-tests and, when necessary, nonparametric tests) for differences between simulation results and field measurements.