AGROS project enters exciting phase: applying autonomous control in cucumbers

Greenhouse horticulture plays an important role in the year-round production of fresh and healthy products. Currently, the horticultural sector is facing a number of challenges. A major one is the limited availability of highly qualified and experienced staff that can oversee all aspects of the production system. Other challenges are the increasing costs of resources, such as electricity, natural gas, CO_2, and fertilisers. The increasing complexity of the production system necessitates the step to a more data-driven approach. The public-private partnership project AGROS aims to realise an 'autonomous greenhouse' in which cultivation is controlled remotely by intelligent algorithms. Sophisticated, non-invasive sensors monitor key crop properties and support management decisions to achieve profitable cultivation under objectified and sustainable criteria.

Two approaches for autonomous greenhouse control

“In the past two years, the AGROS project worked on the building blocks to realise such an autonomously controlled cultivation in cucumbers. We determined a minimum number of plant traits essential in decision-making for crop management and climate control. Together with consortium partners, we selected commercially available sensors and tested them extensively in earlier cucumber trials. For the traits that could not yet be automatically measured, such as the number of newly formed cucumber leaves, we developed vision technology that is currently being validated”, explains Anja Dieleman, AGROS project leader and researcher at WUR business unit Greenhouse Horticulture.

The team selected two approaches for autonomous greenhouse control:

• For the first approach, a Digital Twin is generated by the crop and climate models of the business unit Greenhouse Horticulture. This close-to-real environment Twin determines the ideal control strategy based on the responses of the simulated climate and virtual cucumber crop. The Digital Twin receives near-real-time information on climate and crop from sensors in the greenhouse, and uses that information to “self-calibrate” and improve its control strategy.
• In the second approach, Reinforcement Learning algorithms are trained in mapping actions to observations. By repeatedly simulating trials under different circumstances and controls, the model interprets its environment by evaluating the consequences of different control actions. Desired behaviours are rewarded while undesired outcomes are punished. Once trained on the AGROS objective, the model acts independently and uses the latest available information to evaluate and determine the setpoints of every hour accordingly.

“We have arrived to an exciting phase in the project as we are now applying our approaches to the real crop and evaluating their performance. We planted cucumbers of variety Hi-Power under dynamic LED lighting in three greenhouse compartments at the WUR research facilities in Bleiswijk, The Netherlands. Each compartment is controlled by either the Digital Twin, the Reinforcement Learning algorithm, or a group of crop and irrigation experts from the companies involved in the AGROS project. These experts represent growers’ knowledge and current best practices. The objective for all three compartments is to reach the highest possible net profit. Net profit is determined by the balance between variable costs (electricity, natural gas, CO₂) and benefits (number of harvested cucumbers). We will follow the results and learnings of these controls in the coming months”, says Anja.