Zone-Aware Greenhouse Control and Multitask Crop Yield Prediction Using ConvLSTM and EMMYP-Net

In the new age of greenhouse management, smart space adaptive systems are required to allow environmental control and crop analytics. This study presents an end-to-end deep learning architecture for spatiotemporal prediction and control in agriculture, which enables seamless integration between the different components of a decision loop. A ConvLSTM model predicts zone-specific microclimate variations, and a Dueling DQN agent selects the optimal actions for irrigation, ventilation, and fertilization according to energy demand, emission prediction, and soil moisture balance. The proposed EMMYP-Net (Enhanced Multimodal Multitask Yield Prediction Network) involves a CNN-BiLSTM-attention architecture that combines visual canopy data with multi-sensor sequences to co-classify growth stages and estimate yield. Experimental tests in a 1,200 m² four-zone greenhouse showed remarkable improvements, as ConvLSTM decreased RMSE by 45±2.3% compared to ARIMA and by 31±1.8% compared to LSTM. EMMYP-Net achieved an accuracy of 96.0% for classification, as well as an R² of 0.912±0.007 in predicting yields. This process-integrated approach enhanced resource sustainability by achieving savings of 19.3% in energy, 16.5% in water, and 15.2% in fertilizers relative to a conventional system. Combining predictive control and crop intelligence offers a scalable basis for sustainable data-driven greenhouse management. The key novelty of this work lies in the seamless integration of ConvLSTM-based spatiotemporal forecasting with the EMMYP-Net multimodal crop analytics within a unified reinforcement-learning-driven decision loop, enabling both predictive control and biological feedback in real time.