Integrating Machine Learning in Greenhouse Modeling: A Comparative Study of Advanced Time-Series Strategies

Abstract

Greenhouse cultivation offers a promising solution for sustainable agriculture by enabling controlled environments that optimize plant growth. However, accurately predicting crop growth in these complex, dynamic systems remains a significant challenge. This study employs advanced machine learning techniques Long Short-Term Memory (LSTM) networks, Gated Recurrent Units (GRU), and Support Vector Regression (SVR)to model the growth of dwarf tomatoes under controlled greenhouse conditions. Using a dataset collected during the 4th International Autonomous Greenhouse Challenge, which includes environmental factors and plant growth metrics, the study evaluates the performance of these models in predicting growth dynamics.

Results demonstrate that GRU outperforms LSTM and SVR, achieving the highest predictive accuracy with the lowest Root Mean Square Error (RMSE) and the highest Coefficient of Determination (R²) across most features. The GRU model effectively captures complex temporal dependencies and nonlinear relationships, making it a robust tool for greenhouse crop management. LSTM also shows competitive performance, particularly for specific growth parameters, while SVR exhibits limited ability to handle dynamic temporal patterns.

This study highlights the potential of deep learning models in advancing precision agriculture, enabling improved decision-making in crop yield optimization and resource efficiency. Future work will focus on incorporating additional environmental variables, expanding datasets, and exploring advanced architectures to further enhance model performance and generalizability.