Wageningen, Netherlands – September 2023
In a significant stride toward the realization of fully autonomous greenhouse cultivation, the recently completed AGROS validation trial has demonstrated the efficacy of cutting-edge technologies in managing a cucumber crop. This pioneering endeavor, led by Anja Dieleman, project leader and researcher at Wageningen University & Research’s business unit, Greenhouse Horticulture, showcased the potential of Artificial Intelligence (AI) algorithms and Digital Twins in greenhouse farming.
Over the past two years, the AGROS project meticulously laid the groundwork for autonomous cucumber cultivation. Essential plant traits guiding crop management and climate control were identified, and suitable sensors were chosen to monitor these traits. Two distinct approaches were employed to achieve autonomous control: a mechanistic model-based Digital Twin and a Reinforcement Learning (RL) machine learning algorithm.
The validation trial commenced earlier this year, featuring these two approaches alongside a grower’s conventional practices as a reference. Real cucumber crops were cultivated at the WUR research facilities in Bleiswijk, the Netherlands, where the primary objective was to maximize net profit by optimizing the balance between variable costs (electricity, natural gas, CO2) and benefits (harvested cucumbers based on fruit weight).
Success in Traditional Grower Compartment:
In the traditional “growers’ compartment,” crop management and climate control adhered to existing grower knowledge and best practices. A predefined cultivation and irrigation strategy aimed to achieve balanced crop growth and production. The strategy was centered on the crop’s assimilate demand, which was met through sunlight, additional LED lighting, and temperature and CO2 supply. The irrigation strategy resulted in robust early growth, and fruit pruning was determined based on the number of newly formed leaves and expected light levels. This strategy, focusing on vigorous crop growth, demonstrated a high degree of reliability, closely matching the predicted yield. The net profit in this compartment ultimately proved to be the highest.
Digital Twin’s Precision and Optimization:
The second compartment of the trial was controlled by a Digital Twin generated using a combined crop and climate model. This sophisticated twin determined the optimal control strategy by considering real-time data from climate sensors, manual crop measurements, and simulations. It continuously self-calibrated, improving its control strategy over time. Although the number of harvested cucumbers was slightly lower than in the growers’ compartment, the Digital Twin excelled in optimizing electricity usage costs, achieving the lowest price per kWh used. This approach represents a significant step toward the practical application of autonomous cultivation in commercial greenhouse practices.
Reinforcement Learning Breakthrough:
In the third greenhouse compartment, a Reinforcement Learning (RL) algorithm, trained on virtual data sets of cucumber crops and climate conditions, took the reins. This marked one of the first instances of a fully autonomous greenhouse controlled by an RL algorithm. While the RL-controlled climate differed considerably from the other compartments, the cucumber crop demonstrated resilience to temperature fluctuations, yielding good fruit production. Notably, this RL model controlled actuators like lighting, screens, CO2 concentration, and heating but did not manage irrigation and fruit pruning.
Anja Dieleman highlights the next frontier in autonomous greenhouse control, which involves integrating sensors into the control loop for continuous, automated, and objective data collection. Achieving this objective will rely on robust sensors supported by soft sensor-based solutions, further advancing the development of autonomous greenhouse cultivation.
The AGROS validation trial stands as a pivotal achievement in greenhouse agriculture, bringing us one step closer to realizing fully autonomous and sustainable farming practices.