A Novel Bayesian-Optimized Hybrid Recurrent Neural Network for Predicting Lettuce Harvest Time in Controlled-Environment Agriculture

Accurately predicting the harvest time of a crop remains a significant challenge in agricultural management and sustainable crop production, as it is essential for optimising harvesting schedules and preventing yield or quality losses caused by premature or delayed harvesting. To address this, a deep learning model integrating Gated Recurrent Unit (GRU) and Bidirectional Long Short-Term Memory (BiLSTM) networks was proposed to enhance prediction performance by capturing both short- and long-term dependencies within time series data while tuning the key hyperparameters using a Bayesian optimisation algorithm. The model was based on growth data of plant height, number of leaves, and projected area (the area of lettuce as captured from top-view images), along with environmental data such as light intensity, temperature, humidity, and the amount of carbon dioxide, using plant factory hydroponic lettuce as the research subject. To ensure rigorous evaluation, experiments were conducted under a plant-wise protocol (Plant ID) using grouped cross-validation, preventing any leakage of time windows from the same plant across folds. Under this evaluation setting, the proposed method demonstrated stable performance, achieving MSE = 2.376 ± 0.594, RMSE = 1.532 ± 0.194, MAE = 1.278 ± 0.159, and R² = 0.826 ± 0.029. These results indicate that combining hybrid recurrent architectures with automated hyperparameter optimisation provides an effective and reproducible approach for harvest-time forecasting and supports practical decision-making in controlled-environment agriculture.