Crop Steering in Smart Greenhouses via Plant RL
Crop Steering in Smart Greenhouses: Controlling Vegetative and Generative Growth with Reinforcement Learning
A smart greenhouse shows its real value when climate control goes beyond merely maintaining temperature and humidity and becomes a decision-making system for guiding the plant’s growth trajectory. In greenhouse production, the plant constantly allocates resources among leaf and stem development, flower and fruit formation, quality improvement, water use, and energy consumption. This is exactly where Crop Steering becomes meaningful: the deliberate guidance of the plant’s source–sink balance by adjusting temperature, light, carbon dioxide, humidity, fertigation, and canopy operations. The issue is not simply making the plant grow more; it is directing growth at the right time, toward the right organs, and with the right economic quality.
The importance of this subject for food security and greenhouse economics begins with a simple observation. A grower does not merely want to produce more biomass; they need marketable yield, predictable harvest timing, stable quality, and net profit. In studies related to Wageningen’s Autonomous Greenhouse Challenge, fruit load has been used as an indicator for monitoring the balance between vegetative and generative growth, because the number and mass of developing fruits show how much of the plant’s photosynthetic capacity is being allocated to economic production. This perspective turns Crop Steering from a simple climate-control recipe into a physiological and economic problem.
– Silke Hemming, Head of the Greenhouse Technology Scientific Team at Wageningen University: “The Autonomous Greenhouse Challenge shows that autonomous cultivation is not just theory; it works in practice.”
The connection between Crop Steering and reinforcement learning comes from the multivariable and delayed nature of the greenhouse system. At the operational level, a smart greenhouse is a nonlinear, multi-scale system: climate, radiation, and photosynthesis respond quickly, while cluster development, fruit load, quality, yield, and profit appear with delays of several days or even weeks. Such a system is not well suited to simple moment-by-moment decision-making, because any short-term change in light, temperature, or irrigation may reveal its economic impact much later. For this reason, predictive control, digital twins, and reinforcement learning can evaluate decisions across a longer time horizon instead of relying on isolated reactions.
How Does Crop Steering Turn the Balance Between Vegetative and Generative Growth into a Control Problem?
In plant physiology, vegetative growth refers to leaves, stems, canopy area, and photosynthetic capacity, while generative growth is measured through flowers, clusters, fruits, ripening, and marketable quality. Source–sink balance refers to the relationship between carbohydrate production in the leaves and its consumption by growing organs, especially fruit. If the plant becomes too vegetative, the canopy expands, but the economic fruit load may not develop sufficiently. If generative pressure exceeds source capacity, quality, size, ripening, and growth stability may all be at risk.
The controllable levers in a smart greenhouse include a set of climate, root-zone, and canopy decisions. Air temperature, artificial lighting, screens, ventilation, humidity, carbon dioxide dosing, irrigation scheduling, EC and pH of the nutrient solution, stem density, leaf pruning, and fruit pruning can each shift the ratio between vegetative and generative growth. These levers do not operate independently. For example, increasing temperature can change the pace of cluster formation, while also affecting water use, transpiration, and heating costs. Therefore, successful Crop Steering requires the simultaneous integration of multiple variables, not the isolated adjustment of a single setpoint.
– Fruit Load as a Practical Indicator of Source–Sink Balance
Fruit load is an important indicator for Crop Steering because it connects climate and nutrition decisions to the plant’s actual condition. In WUR data, plant load, or fruit load, is associated with the number of developing fruits per square meter and has been used to assess the balance between vegetative and generative growth. What distinguishes this indicator from purely climate-based variables is that it is closer to the biological outcome of decisions. If an algorithm only observes temperature, humidity, and light, it may control the environment without fully understanding the plant’s status; fruit load helps close part of that gap.
Temperature is a clear example of how a climate variable can influence the generative pathway. In WUR’s sensitivity analysis, changing the temperature from 19 to 24 degrees Celsius was associated with a change in cluster formation rate from 1.1 clusters per week to 1.47 clusters per week. This data shows that temperature is not only a factor of climate comfort; it also changes the speed at which the plant moves toward generative production. For this reason, temperature control in Crop Steering must be interpreted together with fruit load status, expected quality, and energy cost.
Why Does Reinforcement Learning in Smart Greenhouses Require Plant-Centered Data?
Reinforcement learning becomes meaningful for Crop Steering when the greenhouse problem is translated into state, action, reward, and transition. The system state should at least reflect the internal climate, weather forecast, crop status, root-zone condition, product price, energy price, and safety constraints. Control actions may include turning HPS and LED lighting on or off, adjusting LED channel intensity, positioning energy and blackout screens, setting the minimum heating-pipe temperature, adjusting ventilation openings, setting humidity targets, controlling carbon dioxide concentration, and determining irrigation intervals. The reward should not be limited to raw yield, because the economic objective of greenhouse production is net profit along with quality and efficient resource use.
In WUR’s cherry tomato challenge, data exchange and setpoint updates were performed every 5 minutes, while daily indicators such as daily PAR, heating energy, electricity consumption, carbon dioxide dosing, and water use were calculated. This data architecture matters for predictive control and reinforcement learning because the algorithm must both execute short-term decisions and evaluate their cumulative consequences at the end of the day and the end of the production cycle. Light metrics must also be reported through PAR or DLI, not through a general perception of brightness. In WUR data, daily PAR was reported in mol/m²/day and LED intensities in µmol/m²/s, and this level of unit precision makes comparison and optimization possible.
– Silke Hemming and colleagues, researchers at Wageningen University: “For autonomous optimization, automatic registration of crop-related data will be essential in the future.”
The data required for Crop Steering does not come only from climate sensors. The WUR case introduces a set of data for this type of control, including air temperature, absolute humidity, carbon dioxide, PAR, lamp status, screen position, ventilation, heating, irrigation, drain-water EC and pH, slab EC and pH, slab temperature, plant weight, stem diameter, sap flow, leaf temperature, and RGB and thermal images. This dataset moves the control loop from a climate-centered approach toward a plant-centered approach. Without such data, the algorithm may tune the environment precisely while detecting the plant’s actual response with delay or error.
– The Economic Reward Must Measure Quality and Resource Use at the Same Time
In reinforcement learning, the reward functions as the language that translates the grower’s objectives into algorithmic decisions. In Crop Steering, if the reward focuses only on increasing kilograms of product, the algorithm may ignore quality, harvest timing, electricity, heat, water, carbon dioxide, or fertilizer use. In the WUR challenge, the main objective function was net profit, and the revenue model was linked to product quality; tomato price was associated with Brix, quality, and season timing. Therefore, an appropriate reward should treat the product not only as harvested mass, but as a commodity with quality, price, and production cost.
This perspective is also visible in fertigation and quality metrics. EC measured in dS/m, pH, slab EC, slab pH, slab temperature, liters per square meter of irrigation, and liters per square meter of drainage are essential for understanding the root-zone environment. On the other hand, TSS and Brix are closer to fruit quality and economic reward. In the cherry tomato experiment, quality was assessed using Brix and the WUR taste model, with taste reported on a 0-to-100 scale. This connection between fertigation, quality, and price turns Crop Steering from simple growth control into the economic management of the crop.
What Do WUR Data from Cucumber and Cherry Tomato Show About Autonomous Control?
WUR’s Autonomous Greenhouse Challenges are among the clearest experimental examples for understanding the potential and limitations of algorithmic Crop Steering. In the first challenge, each team remotely controlled a modern 96-square-meter greenhouse compartment for four months of cucumber production. The equipment included heating, ventilation, screens, lighting, fogging, carbon dioxide, water, and nutrients. The key result was that one AI team recorded 6 percent higher Class A production and 17 percent higher net profit compared with the human reference. However, this result belongs to a controlled research environment and should be interpreted cautiously before being generalized to commercial production.
– Silke Hemming and colleagues, researchers at Wageningen University: “One team was able to outperform the manually managed reference cultivation.”
The second challenge, focused on cherry tomatoes, provided a more detailed picture of the economics of Crop Steering. The experiment was conducted in six 96-square-meter compartments, with five AI teams competing alongside one human reference. The production period lasted six months, and the objective was defined as maximizing net profit through the integration of climate, irrigation, crop, and quality. All AI teams achieved higher net profit than the human reference, with net profit ranging from 3.10 to 6.86 euros per square meter. Class A fruit production ranged from 12.9 to 14.4 kilograms per square meter.
The economic details of the cherry tomato challenge show that the central indicator in Crop Steering is net profit, not raw yield. The top-performing team in the cherry tomato challenge recorded revenue of 37.22 euros per square meter, costs of 26.07 euros per square meter, and net profit of 6.86 euros per square meter. The human reference recorded revenue of 35.56 euros per square meter, costs of 29.38 euros per square meter, and net profit of 3.10 euros per square meter. The important difference in this comparison was not merely higher revenue, but the combined balance of revenue, cost, quality, and resource use; this combination is crucial for defining a reinforcement learning reward.
The resource analysis of the best-performing cherry tomato team also shows that Crop Steering must be evaluated together with water, energy, carbon dioxide, and nutrient consumption. Resource use for the best team included 12.9 MJ/kg of heat, 18.7 kWh/kg of electricity, 0.63 kg/kg of carbon dioxide, 25.0 L/kg of water, and 83.0 g/kg of nutrients. These figures are useful for reward design and strategy comparison, but they should not be directly generalized to other climates, markets, or cost structures. Their value lies in demonstrating the evaluation architecture: production, quality, and resource use must be viewed within a single decision framework.
Technical Limitations of Algorithmic Crop Steering in Quality and Fertigation
The WUR data is not only a success story; scientific and operational limitations are visible in the same dataset. One of the main risks in greenhouse reinforcement learning is the gap between the high density of climate data and the low density of crop data. Internal climate can be recorded every few minutes, but crop status, fruit load, quality, dry matter, Brix, and physiological response are not always available at the same frequency. This gap makes the machine learning model dependent on the volume, diversity, and quality of training data, and when data is insufficient, its decisions may become unstable or overfitted.
– S. C. Mári and colleagues, authors of the Sensors article on dwarf tomato: “For an optimal strategy, the control algorithm must account for plant response through validated sensors, data, and models.”
Fertigation is one example of this complexity. In the cherry tomato challenge, the relationship between drain-water EC and Brix was not reliable for the cultivar used. Average Brix was reported as 8.7 and fruit dry matter as 9.0 percent, but the authors concluded that existing knowledge was not sufficient for autonomous fertigation optimization. This point is highly important for RL design, because the algorithm cannot guarantee a quality reward based merely on a simple correlation between EC and fruit sweetness. Product quality is the result of a network of light, temperature, water, nutrients, fruit load, and growth stage.
WUR’s sensitivity analysis on carbon dioxide also reflects this multi-factor logic. A sharp reduction in carbon dioxide could reduce production by up to 0.72 kilograms per square meter, but increasing it beyond the strategies already applied produced only a small additional benefit. This limitation was related to ventilation, plant uptake, and carbon dioxide cost. Therefore, a higher setpoint does not always mean a better outcome; Crop Steering must identify the point where plant response, resource cost, and technical constraints are aligned.
Artificial lighting also showed a stronger effect on net profit than carbon dioxide and temperature in WUR’s analysis, but even here, decision timing mattered. Adding 2 hours of daily lighting around the tenth week after planting simulated approximately 0.1 euros per square meter of additional net profit, whereas near the end of the season its effect was small or negative. This data shows that reinforcement learning must understand the growth stage and base its decisions on the crop’s biological timing. More lighting is not an optimal strategy at all times, because crop response and energy cost change throughout the season.
Cybersecurity and Data Governance in RL-Based Smart Greenhouses
When remote climate control, API exchange, connected sensors, and reinforcement learning algorithms enter the greenhouse, the food production system becomes a digital operational environment. The climate computer, actuators, sensors, communication network, and algorithmic decision layer become part of an industrial control system. Access errors, data manipulation, or disruption of setpoints can have direct consequences for the crop, energy costs, and production safety. Therefore, cybersecurity in Crop Steering is not a peripheral issue; it is part of the control architecture and risk management framework.
– International Society of Automation, ISA standards organization: “The ISA/IEC 62443 standards define requirements and processes for the security of industrial control systems.”
The IEC/ISA 62443 framework for industrial automation and control systems provides an appropriate language for classifying these risks. In a smart greenhouse, secure design must cover development, operation, remote access, software updates, network segmentation, and permission management. This requirement becomes even more critical when an RL algorithm makes decisions based on five-minute data and sends setpoints to the process computer. The more automated control becomes, the more important decision traceability, model versioning, and human review capabilities become.
Data governance is important not only for security, but also for market access and auditing. The GLOBALG.A.P. IFA Smart Version 6.0 document for fruit and vegetables defines documentation through identification, date, pagination, sufficient detail, periodic review, assignment to staff, and version control. In a greenhouse based on Crop Steering, this logic extends to the audit trail of climate, root-zone, crop, quality, and algorithmic decision data. If the product is intended for professional and export-oriented supply chains, data is not merely a control tool; it becomes part of market trust, auditing, and production traceability.
Localizing Crop Steering in Iranian Greenhouses Through Data-Driven Pilots
The growth trend of greenhouse production in Iran shows that there is infrastructural capacity for a gradual transition toward smart systems. Published academic data on Iran shows that greenhouse cultivation area increased from 3,380 hectares in 2002 to 6,630 hectares in 2012, and reached 15,678 hectares by the end of 2019. This historical trend alone does not prove readiness for reinforcement learning, but it does show that greenhouse production has become one of the country’s important paths toward controlled-environment agriculture. For Iran, the decision point is not a direct leap into fully autonomous control; the more logical path is to build measurable and auditable pilots.
For Iran, localizing Crop Steering should begin with evidence-based analysis of global experience and the need for domestic data. The quantitative impact of this technology on profit, water use, energy consumption, or export quality under local conditions becomes reliable only when climate, root-zone, energy, water, quality, and sales data are recorded systematically. The WUR architecture shows that a 500- to 2,000-square-meter pilot with five-minute data logging and daily indicator calculation can be a suitable starting point for testing predictive control and RL. Such a pilot must define its objective from the beginning: increasing yield, stabilizing quality, reducing resource use, or maximizing net profit.
In designing an Iranian pilot, crop selection and quality indicators play a decisive role. The WUR cherry tomato experience showed that the combination of Brix, taste, season timing, and price can make the economic reward quality-oriented. For products whose domestic or export markets are sensitive to uniform quality, size, ripening, or Brix, Crop Steering is not just a tool for producing more; it becomes a tool for standardizing quality. However, each crop must have its own price and quality model, and the algorithm’s reward should not be transferred from one crop to another without testing.
Iran’s implementation path must also account for energy risk, data quality, operator training, and control security. Decisions such as increasing artificial lighting, changing temperature, or dosing carbon dioxide can reduce profit if they are made without a cost model and operational constraints. At the same time, if sensors are not calibrated or crop data is recorded with delay and error, the reinforcement learning algorithm may appear advanced while resting on a weak decision foundation. Therefore, the real competitive advantage does not come from installing a few scattered sensors, but from building a complete chain of data, modeling, control, auditing, and human feedback.
A Practical Summary of Crop Steering for Smart Greenhouse Investment
Crop Steering in a smart greenhouse is not a single technique; it is a method for translating plant physiology into controllable economic decisions. WUR data shows that algorithms have been able to outperform human references in controlled research environments, but the same data also shows that model quality, crop-data density, reward definition, and technical limitations are decisive. Vegetative and generative growth cannot be guided through a fixed recipe, because the plant’s response to temperature, light, carbon dioxide, fertigation, and fruit load changes throughout the season. A good algorithm must see these changes across the right time horizon.
For investment, the decision criterion should not be the promise of full automation. A more precise criterion is whether the project can measure net profit, product quality, energy use, water use, carbon dioxide use, nutrient use, and operational risk within an auditable data system. The WUR cherry tomato experience shows that differences between strategies become meaningful when revenue, cost, and quality are evaluated simultaneously. Within such a framework, reinforcement learning can become an optimization tool, but only when it relies on reliable sensors, stable data, reviewable models, and a secure architecture.
The realistic path for Iran runs through limited, crop-specific, data-driven pilots. Instead of making early claims about broad profitability, these pilots must show which crop, which climate, which sensor combination, and which reward function work under local conditions. If five-minute data logging, daily resource metrics, quality indicators, and algorithmic decision traceability are designed from the outset, Crop Steering can become a shared language among growers, investors, data specialists, and market managers. This technology reaches maturity when plant growth, resource consumption, and net profit are placed within a single, reliable decision loop.
