Adaptive spectral-thermal illumination management for protected tomato cultivation: a fused deep learning and...

This study presents an intelligent greenhouse lighting control framework that integrates a CNN-ELM photosynthesis prediction model with MOEA/D-based multi-objective optimization to improve tomato production while reducing the carbon impact of supplemental LED lighting. The CNN-ELM model was trained using key environmental variables, including photosynthetic photon flux density (PPFD), red-to-blue light ratio, canopy temperature, CO₂ concentration, and relative humidity. Within the experimental conditions, the model achieved high predictive accuracy, with an R² of 0.976 and an RMSE of 0.712 µmol m^-2 s^-1. Using these predictions, the MOEA/D algorithm generated Pareto-optimal lighting strategies, which were ranked through entropy-weighted TOPSIS and implemented via cloud-based control connected to a LoRa wireless sensor network and pulse-width-modulated LED drivers. The system was evaluated during a 110-day tomato cultivation trial and compared with single-parameter control and ambient-condition treatments. Results showed a 38.4% reduction in LED-related carbon emissions, a 22.6% increase in net photosynthetic rate, and a 31.7% improvement in harvestable yield relative to ambient conditions. Physiological analyses further indicated enhanced photosynthetic performance, radiation-use efficiency, and light utilization. Overall, the findings demonstrate that data-driven, closed-loop lighting management can simultaneously enhance productivity and reduce greenhouse gas emissions in controlled-environment agriculture when applied within the validated operational domain.