Using machine learning to find links between plant performance and greenhouse water quality

A virtually emission-free greenhouse sector by 2027: this is the ambitious goal, requiring the minimisation of the negative effects of fertilisers and pesticides on groundwater and surface water. At the same time, dry summers are a headache for greenhouse horticulturalists, who use rainwater for irrigation. In the light of these developments, there is an increasing need to recirculate water in the greenhouse.

Preliminary phase

In a preliminary phase, the earlier TKI project ‘Waterkwaliteit snel in beeld (Water Quality Quickly in View) ’ made a start on using sensors to measure new parameters that give an indication of the quality of recirculated irrigation water. The results were promising. The next phase included the recently completed project ‘Automatische aansturing van waterkwaliteit voor goed recirculatiewater glastuinbouw (Automatic control of water quality for good recirculation water in greenhouse horticulture’. The aim was to develop a sensor- and data-driven model for water quality that uses clear control parameters and a user-friendly dashboard to help horticulturalists to optimise water, climate and plant health in a targeted way.

Black box

“The greenhouse water system is effectively a black box,” summarises water quality expert Van den Broeke. “Because reusing water also means returning any problem substances to the greenhouse. But there is no clear picture of which substances they are. Laboratory tests are being conducted of water samples to look at critical plant nutrients, which are also added in the form of fertilisers. But micropollutants, including possible pathogens, can accumulate. And substances from the plants themselves are also not monitored. The thinking behind this project was to use existing sensor technology in order to extract information from the greenhouse water system that allows horticulturalists to control water treatment at an early stage. By combining different sensors, we wanted to work towards trigger values for a range of parameters that indicate when intervention is required. The thinking was that horticulturalists could get to work once there was a dashboard presenting an overall picture.”

Monitoring plant performance

In addition to water quality, sensors monitoring the condition of plants were also used. Van den Broeke: “In general, the focus with plant performance is on factors such as yield. But it can take weeks or months before you get that information. We were looking for something you can monitor in the plant with the same response time as with water quality. There are sensors that can do this by measuring photosynthesis. We also used sensors that monitor a plant’s electrical signals, which are an indicator of possible stress. In that way, we were able to determine the quantities we can use to establish a picture of plant performance associated with differences in water quality.”

Biggest challenge

A series of sensors were then installed in the greenhouse at Tomato World – het innovatiecentrum van de Nederlandse glastuinbouw – to collect data about water quality and plant performance, as well as the greenhouse climate in terms of humidity, CO₂ levels and light intensity. The aim was to use machine learning models based on those data to look for a predictive value in the relationship between water quality and plant health. “With machine learning, you use computer tools to find relationships in a dataset that you can’t find yourself,” explains Van den Broeke. “So it’s not about describing how a system works, it’s about optimising a prediction. To collect all those data, we used as many as a hundred sensors located in different areas of the greenhouse. The biggest challenge was to get everything working. We spent eighteen months installing the sensors and getting them operational. The sensors have to be properly calibrated and also cleaned regularly. In addition, interfaces had to be developed to connect the data flows to a central platform. At the outset of the project, our thinking was the more sensors, the better. Machine learning is the technology for transforming a mound of information into something useful. But we really underestimated that process. And so the final report includes numerous recommendations about what to watch out for when working with so many sensors.”

Operational model

Greenhouse measurements were taken at other horticulture businesses in addition to Tomato World. On the basis of all the combined data, we managed to establish a working model that can predict the response of plants to changes in water quality within a three- to eight-hour time frame. “This result does not mean that plant sensors are obsolete now,” emphasises Van den Broeke. “Currently, we only have information about when the plant is doing well: the tests were all conducted in a real-world situation and so no risks involving plant health were allowed. That meant we couldn’t pester the plants with poor water quality to the point where they died. I believe this project has shown the potential of machine learning for this application. We have got closer to a model with trigger values that will allow horticulturalists to operate on the basis of water quality. We also know which parameters are important here, such as electrical conductivity, light intensity and oxygen levels. At the end of the project, the consortium decided that it would be a shame not to keep going. And TKI is an excellent environment to take this innovative research forward.”

Taking horticulture to the next level

There are also positive responses from the field about the project, despite the outcome being different than expected. Rob Weerdenburg, who is a horticulturalist with the Van Geest company, explains how sensors were installed in tomato plants in their greenhouses to measure plant performance. “We believe it is important to be at the forefront of sustainability. Because that is the future. When people want to try something in that area, we always say: feel free to call. We enjoy working on taking horticulture to the next level.” Whenever the progress of the project was discussed, Weerdenburg was there as much as possible to provide real-world input. “At times, I did feel that there was a need to keep the researchers grounded. And I sometimes thought the approach needed more structure. The report on the final results also took a long time. For me, the bird had already flown by then.”

Different perspective

Weerdenburg says he had thought the project would give him a clearer picture of how to control water quality in the greenhouse. That was not the case. “Even so, I’m happy we were involved,” he says. “That intensive focus on the sensors has raised another question for us. What do we really know about the microbiology of our water? And what are the implications for our plants? It is possible that, as horticulturists, we have not yet gone as far down the road with sensor technology as the researchers. We first want to know the extent to which the fungi and bacteria in our recirculated water are good or bad for our plants. I think establishing a different perspective of this kind is much more important than blindly adopting something. At times, the results looked disappointing but, in the end, we learned a lot from them.”

Alliance partners

The ‘Automatic control of water quality for good recirculation water in greenhouse horticulture’ project was possible thanks to the following partners: Agrona, Glastuinbouw Nederland, Kennis in je Kas, KWR, LetsGrow, Normec Groen Agro Control, Plantum, Sendot, Stichting Control in Food & Flowers, STOWA, Tomatoworld, Vivent Biosignals and it was an alliance between the Horticulture & Basic Materials Top Sector and the Water and Maritime Top Sector.