Real-Time Lettuce Biomass Forecasting in Indoor Vertical Farming Using Machine Learning Framework

Highlights Real-time lettuce biomass was estimated using RGB images in a vertical farming system. A two-stage pipeline combined CNN-based prediction with XGBoost and LSTM for forecasting. Results showed high accuracy in biomass prediction (R2 = 99.56%) and forecasting (R2 = 95.10% for XGBoost and 94.31% for LSTM). Low variability of environmental data caused identical outcomes of model performance. Study shows potential of two-stage lightweight deep learning for yield forecasting. ABSTRACT. Lettuce is one of the most commonly cultivated crops in indoor vertical farming systems utilizing artificial lighting. Accurate forecasting of its biomass yield is essential for aligning production with market demand and optimizing resource use. This study presents a two-stage machine learning framework combining a Convolutional Neural Network (CNN) for real-time biomass estimation, followed by both an Extreme Gradient Boosting (XGBoost) and a Long Short-Term Memory (LSTM) model for future biomass forecasting. Using hourly top-down RGB images collected over a 30-day cultivation cycle, the CNN achieved high prediction accuracy with an R 2 of 99.56% and a root mean square error (RMSE) of 3.65 g. The XGBoost and LSTM models, utilizing a sliding-window approach, achieved R 2 value of 95.10% and 94.31%, respectively, and a forecasting RMSE of 6.63 g and 6.71 g, respectively. These results demonstrate the potential of machine learning models for non-destructive, real-time yield forecasting in indoor farming, providing a foundation for data-driven resource management and production planning in indoor vertical farming systems.