A multi-model deep learning approach to address prediction imbalances in smart greenhouses

Abstract

The creation of smart greenhouses is playing a crucial role in paving the way toward precision agriculture characterized by enhanced efficiency. Integral to these greenhouses are decision-support systems that leverage sophisticated forecasting algorithms to predict a range of parameters. However, these predictors often employ a single model approach for forecasting all variables of interest, leading to imbalanced predictions where some variables are accurately predicted while others do not. Such inconsistencies can undermine the overall reliability of the decision-support systems. Addressing this challenge, this paper proposes an approach that harnesses the potential of multiple deep-learning models operating concurrently to predict a broad array of environmental parameters within a smart and operational greenhouse. Each model is specifically tailored to concentrate on a distinct subset of target variables, thus ensuring that the overall accuracy of the prediction is optimized. The effectiveness of this approach has been evaluated in a real-world greenhouse setting. The results indicate a substantial improvement, exhibiting more than an 8% enhancement in the Mean Absolute Percentage Error (MAPE) compared to a single-model alternative, particularly in predicting specific environmental variables, confirming the potential for more reliable and precise agricultural decision-support systems.