The global population is increasing rapidly together with the demand for healthy fresh food [1
]. The greenhouse industry can play an important role providing fresh food, such as fruits and vegetables being high in vitamins and minerals. Greenhouses allow a high crop production per area combined with a high water use efficiency per unit of produce [2
]. Worldwide, the area of greenhouse production is increasing [3
]. However, the greenhouse industry encounters difficulties finding enough skilled labor to manage crop production [4
]. A crop manager must have a high level of knowledge and experience in order to control crop growth. As farms become larger, monitoring all details of the various greenhouse compartments becomes more demanding. Moreover, resources (water, fossil energy) are becoming scarcer, which causes an urgent need for maximum resource efficiency.
A greenhouse protects the crop from outside influences, such as rain, wind, low temperatures, or pests. A modern high-tech greenhouse is equipped with active control of actuators (e.g., heating, lighting, irrigation) in order to create a favorable growing climate. Of course, this comes at the cost of resource consumption (e.g., fuel, electricity, water). A grower determines the climate and irrigation strategy and defines the setpoints for all climate and irrigation parameters. Actuators are operated based on the setpoints, and sensors give feedback on measured data for the control loop. Automated greenhouse climate control algorithms have already been developed decades ago [5
]. Today, modern high-tech greenhouses are equipped with process computers, which are able to control greenhouse actuators based on the setpoints manually set by the grower.
In order to add more automated control, various greenhouse climate and crop models have been developed. An overview of today’s greenhouse climate models is given in a previous study [13
]. An overview of greenhouse crop models and modelling approaches are given in other studies [14
]. Dynamic greenhouse climate models and dynamic crop models have been used to determine setpoints automatically and take over the decision of the grower. If climate and crop simulation models [15
] are combined and connected to the sensors and actuators of a greenhouse, greenhouse climate and crop growth can be controlled by automated algorithms. Such experiments have been conducted successfully with tomato [17
] and sweet pepper in The Netherlands [18
]. In this experiment, outside weather conditions and weather forecasts were used for climate simulations. Crop growth simulations were carried out along with the cropping cycle to predict further crop growth and development for different sets of setpoints. The optimum set was then applied in the greenhouse automatically. The computations were repeated every day, and in this way, crops were grown with an optimum control strategy. Other experiments with tomato have previously been conducted [20
Another way to take over parts of the decisions of a grower is to use machine learning algorithms for greenhouse climate control [21
]. Diverse methods have been applied in research, such as K-algorithms [22
], Bayesian networks [23
], support vector machines regression [24
], neural networks [27
], reinforcement learning [35
], or genetic algorithms [36
]. However, to our knowledge machine learning has not been used yet to control climate and irrigation and make crop management decisions for growing a greenhouse crop autonomously during a longer period with yield levels comparable to commercial practice.
On the other hand, the use of artificial intelligence (AI) has reached major breakthroughs in several areas of daily life and society, such as medical applications [39
], autonomous cars [40
], or robotics [41
]. AI algorithms have been shown to outperform humans in complex decisions, e.g., checkers [42
], chess [43
], and go [44
]. It is obvious to use AI also for agricultural purposes [45
In order to combine the use of modern AI algorithms and greenhouse climate, irrigation, and crop growth control, in 2018 an international challenge on “autonomous greenhouses” was conducted at the high-tech research greenhouses of Wageningen University and Research in cooperation with five multi-disciplinary international teams. The challenge aimed at combining horticultural expertise with AI to make breakthroughs in fresh food production with fewer resources. The experiment was set-up with the goal of benchmarking the use of state-of-the-art AI algorithms for cucumber production. In the experiment existing commercial greenhouse equipment (actuators), standard sensors for measurement and control, and a standard commercial process computer were combined with the latest AI technology in order to maximize net profit and minimize resource use, while controlling greenhouse crop growing remotely. The goal of this paper is to describe the results obtained by teams concerning net profit and resource use, to analyze differences in climate and crop growing strategies used, and to investigate which lessons can be learned from the results for the future, in terms of optimizing crop yields and net profit.
2. Materials and Methods
2.1. Greenhouse Compartments and Actuators
Six identical greenhouse compartments were available for the cucumber growing experiment. Each compartment was equipped with standard actuators, also available in commercial high-tech greenhouses (Figure 1
). Two pipe heating systems, a rail pipe heating on the floor, and a pipe heating on crop height (peak capacity 180 and 30 W/m2
respectively), were available, both controllable by different setpoints. Continuous roof ventilation (ventilation area of 0.3 m2
opening per m2
greenhouse, equipped with anti-thrips netting), two types of inside moveable screens (LUXOUS 1547 D FR energy screen and OBSCURA 9950 FR W light blocking screen, Ludvig Svenssion, Sweden), a high-pressure-sodium artificial lighting system (capacity of 187 μmol/m2
/s), a fogging system (maximum capacity of 330 g/m2
/h), and CO2
supply (maximum capacity 15 g/m2
/h) were available. Plants were grown in rockwool substrate cubes and placed on rockwool substrate slabs; the plant-substrate system was then located on hanging gutters. Irrigation water and nutrients were supplied with drippers operated by a valve. The surplus of the nutrient solution (drain) was recollected in the hanging gutter in a closed loop system.
2.2. Sensors and Remote Control
Five teams (Sonoma, iGrow, deep_greens, The Croperators, AiCU) were able to control the operation of all actuators remotely based on their own AI algorithm. A sixth greenhouse compartment was controlled by Dutch growers and served as a reference (growers = reference). Competing teams used their own AI algorithms to determine the climate and irrigation control setpoints, such as minimum rail pipe temperature (°C), minimum crop pipe temperature (°C), heating temperature (°C), ventilation temperature (°C), minimum ventilation opening (%), humidity deficit setpoint (g/m3
), energy screen position (0–100%), blackout screen position (0–100%), artificial illumination (0% or 100%), CO2
concentration (ppm), and time between last and next irrigation turn (min). Setpoints were sent via a digital interface (LetsGrow.com) to a central climate process computer (IISI, Hoogendoorn, The Netherlands), which then operated the actuators accordingly (Figure 2
). A nutrient solution for fertigation was prepared by a central fertigation computer and then stored in a buffer tank per compartment before being provided to the crop with drippers. The composition, concentration (EC), and pH of the nutrient solution was determined by the teams. Based on detailed chemical analysis of the drain water, provided every fortnight, the teams could send requests to change the composition, EC, and pH of the nutrient solution.
Standard sensors continuously measured data, such as cumulative outside global radiation (J/cm2
/d), outside photosynthetically active radiation PAR (μmol/m2
/s), air temperature outside (°C), outside relative humidity (%), wind speed (m/s), outside global radiation forecast (W/m2
), outside air temperature forecast (°C), outside relative humidity forecast (%), wind speed forecast (m/s), air temperature inside (°C), air humidity deficit inside (g/m3
), heating pipe temperature (°C), heating power used (W/m2
) for both heating systems, lamp status (on/off), CO2
dosage (on/off), screen position (%) of both screens, irrigation supply (l/m2
), drain (l/m2
), drain EC (dS/m), and drain pH (−). The following data was calculated from the measured data: inside PAR sum (mol/m2
), heating energy used (kWh/m2
), electricity used (kWh/m2
), water consumption (l/m2
), and was provided to the teams as well. Measurements and calculations were sent back to the teams via a digital interface (Figure 2
). Both, setpoints for control of actuators and measurements were exchanged at a 5-min-interval.
Teams were allowed to install additional sensors at the start of the experiment. They chose different types of sensors, such as RGB cameras, thermal cameras, sensors for net radiation, root zone sensors, crop and substrate weight, stem diameter, crop sap flow meters, and wireless temperature-humidity-light sensor networks. One team chose to rely on the standard greenhouse sensors only (iGrow).
2.3. Crop and Crop Parameters
Cucumbers seedlings cv. “Hi-Power” were sown on 20 July 2018, in rockwool cubes and were transplanted to the greenhouse compartments on 14 August 2018, at the start of the experiment. The crop was grown in a high-wire growing system. Plant density and stem density had to be chosen by the teams before the start, resulting in values between 2.6 and 3.6 stems/m2
(iGrow, 2.6 stems/m2
; deep_greens, 2.6 stems/m2
; AiCU, 3.6 stems/m2
; Sonoma, 3.3 stems/m2
; The Croperators, 3.2 stems/m2
). The reference was 2.5 stems/m2
. The first harvest was on 6 September 2018, and the last harvest was scheduled for all teams on 7 December 2018. Based on this last harvest date, the date of topping (removal of head of the crop) had to be chosen by the teams and differed from 19–28 November 2018. The reference was topped on 9 November 2018. Crop development and harvest is shown in Figure A1
, Figure A2
, Figure A3
and Figure A4
Teams sent weekly instruction for fruit and leaf pruning in the top of the canopy to the greenhouse workers. Fruit pruning strategies ranged from a stable procedure of 50% fruit removal for the whole cropping period to a more variable strategy of removing alternately 50% and 67% of the fruits. With respect to leaf pruning, the majority of the teams decided for on pruning (0%) or on pruning a small fraction of leaves (33%). One team used a deviating strategy of removing 50% of the leaves throughout the whole cropping. As a standard procedure applied to all crops, greenhouse workers removed leaves below last harvested fruits. Three harvest quality categories were distinguished (A: >375 g and no defects, B: 300–374 g or defects e.g., shape, color, others, C: <300 g per fruit). Harvest data such as number and weight of fruits (#/m2
per quality category A–C) were measured manually by the workers. Crop related parameters such as stem elongation (cm per week), fruit growth period (d per fruit), leaf formation rate (# per stem per week), and cumulative number of leaves (# per stem) were also measured. Instructions by teams and data measured by workers were exchanged via the digital interface (Figure 2
2.4. AI Algorithms
Each competing team developed their own AI algorithms, which varied between supervised, unsupervised, and reinforcement machine learning (Dynamic Regression, Deep Reinforcement Learning DRL, Deep Deterministic Policy Gradient DDPG, Generative Adversarial Networks GAN, Convolutional Neural Networks CNN, Recurrent Neural Networks RNN).
In order to use AI techniques, training data is essential. Since training data with a wide variation for the described application are scarce, an artificial training data set was created. The use of artificial training data sets has been shown to be very useful in other applications earlier [46
]. In this experiment artificial training data was created using the broadly validated dynamic greenhouse climate model KASPRO [16
] and the cucumber crop model INTKAM [15
] that was modified for a high-wire cucumber crop. The artificial dataset was provided to the teams before the start of the experiment.
2.5. Performance Criteria
Teams’ performance was evaluated based on three criteria.
Sustainability: 20% of the total score of a team was given for sustainability. The following aspects were calculated based on measured data: Energy use efficiency (MJ/kg cucumber), CO2 dosage (kg/kg cucumber), water use efficiency (m3/kg cucumber), pesticide usage as registered (mL/cucumber).
Net profit: 50% of the total score of a team was given for the net profit. Net profit was calculated based on the following obtained data: number of fruits harvested x price per fruit and category. The prices varied per week during the cultivation period and were determined by a jury at the start of the experiment, and reflected an average seasonal trend. The prices varied between €0.30 and €0.40 per fruit (class A). B-class fruits had a 15% lower value, and C-class fruits had no value. The prices were revealed on a weekly basis during the growing cycle in order to mimic price uncertainty, typical for agricultural products in practice. Also, the initial costs for the young plants (costs of a young plant x number of young plants placed in the compartment) and costs of the substrate were considered. This way, the teams had to weigh the faster initial growth of a high stem density crop against the higher initial costs. Other greenhouse equipment used was identical, and therefore not considered in the calculation of the net profit. The fact that capital costs equal for all teams were left out of consideration explains the high values of net profit shown in the results. Resource use of electricity, heating, CO2, water, nutrients, chemical and biological pesticides, and labor were measured during the experiment per greenhouse compartment, and assigned to the teams. Multiplied with the given price, costs were calculated and communicated with teams weekly during the ongoing experiment.
AI algorithm: 30% of the total score of teams was given by a jury based on novelty of the AI algorithm with respect to the overall scientific community, novelty with respect to application on the horticultural domain (novelty), capacity to operate autonomously at a distance without manual interventions (functionality), capacity to operate without too many additional sensors or information (robustness), and easiness of implementation on a large scale (scalability).
2.6. Analysis and Interpretation of Results
The AI-based operation of the different greenhouse compartments by different teams resulted in different cropping, climate, and irrigation strategies, and different yields and resource use efficiencies. In order to properly analyze and compare the different approaches, a combination of a dynamic greenhouse climate model KASPRO [16
] and a cucumber crop simulation model INTKAM [15
], which was modified for a high-wire cucumber crop, was used. The combined model assumes adequate supply of water and nutrients and does not simulate the presence and effects of pests and diseases.
The KASPRO model computes the greenhouse climate as a function of outside weather conditions and greenhouse climate control settings. The model processes these settings by a control algorithm comparable to the ones used. The analysis was by commercial greenhouse climate computers. The model takes full account of the limitations of real greenhouses, which means, for instance, that a CO2 dosing setpoint of 800 ppm is simply not met in sunny periods when the vents are wide open to carry off the heat excess. This is caused by the limitations in maximal supply rate, just like in real greenhouses.
The greenhouse climate, as computed by KASPRO [16
], is then fed to INTKAM [15
], which computes the daily gross photosynthesis from the sum of hourly photosynthesis-rates. The hourly values are the result of light-intensity, temperature, CO2
-concentration, and relative air humidity in combination with the dynamically-simulated crop architecture (in particular leaf area index). After subtracting maintenance costs, the daily amount of assimilates is partitioned over the growing organs (roots, stem, leaves, and fruits) on the basis of their relative potential growth rates. Next, dry matter fraction and fresh organ weights are computed, and finally the harvest moment of individual fruits is determined on the basis of, amongst others, fruit weight [15
With the availability of these models, the contribution of the variation seen in the control strategies of the different teams in the final production could be determined. The variations applied referred to both the observed cropping and climate strategies to interpret and understand the results of the different AI-based operations and identify which additional improvements could have been made.
In a first step, the combined model was used to calculate the cucumber yield of each of the compartments, while using the actually applied crop density, fruit and leaf pruning strategy, and the realized lighting and climate (temperature and CO2) setpoints in that compartment as model inputs. The calculated model output was the predicted fresh yield (kg/m2, #fruits/m2) per greenhouse compartment, which could then be compared with the realized yield in the same greenhouse compartment to validate the models.
In a next step, for each greenhouse compartment, model calculations were carried out applying the cropping strategy of other teams or the reference in order to predict the changes in yield while maintaining the original lighting and climate strategy. In another step, model calculations were carried out for each greenhouse compartment, maintaining the original cropping and climate strategy but applying the lighting strategy of the other teams. In another step, calculations were made for each greenhouse compartment applying the climate strategy (CO2) of each of the other teams, while maintaining the original cropping and lighting strategy. For joint comprehensibility, interactions of cropping, lighting, and climate strategies were not calculated. The simulations of the swapping strategies represent the yield retrieved prior to topping, to eliminate the effect of early topping dates selected by some of the teams.
In Figure 3
, the cumulative cucumber production per team in the different greenhouse compartments during the experimental period is given. From the beginning, one team (Sonoma) had the highest production and was able to continue this, as is shown by the highest curve slope. This team obtained this harvest with a high daily light integral (Figure 4
), as they assumed that with a higher daily light integral a greater harvest could be obtained. Therefore, team Sonoma focused its AI algorithms on this particular aspect. The algorithm allowed them to obtain a high daily light integral by maximizing the amount of artificial light (Figure 5
b), while optimizing other defining factors, such as temperature and CO2
. Hence, they realized the highest yield and they were also able to maintain high light use efficiency of the crop for a long period (Figure 6
Another team (The Croperators) increased the daily light integral after a short period at the beginning of October (Figure 4
and Figure 5
b), however, this did not lead to a higher light use efficiency in October (Figure 6
), since at the same time they maintained a low CO2
concentration. In addition, they opted for a crop pruning strategy that resulted in insufficient assimilates to sustain the growth of all fruits, and is probably the cause of approximately 30% aborted fruits (Figure 7
). These results show that a high daily light integral and a high CO2
concentration are important production factors, which is related to their effect on the photosynthesis rate [47
] for cucumber production, together with balanced crop management (stem density and pruning strategy).
Growers (reference) started with a relatively low harvest. They allowed lower daily light integrals at the beginning (Figure 4
), and therefore only used low levels of artificial light (Figure 5
b), with the philosophy that this approach would prepare the crop better for the approaching autumn season. They were able to balance supply of, and demand for, assimilates with their crop pruning strategy, which resulted in the lowest amount of aborted fruits (Figure 7
) and a high light use efficiency (Figure 6
). In fact, the manual growers were able to realize the highest light use efficiency during almost the whole cropping cycle. Minimizing fruit abortion is an important objective for cucumber growers, and this strategy was clearly and successfully applied by the manual growers.
and Figure 9
show the CO2
concentration and the CO2
dosage realized per team, respectively. All teams started with relatively low CO2
concentration, due to loss of CO2
due to high ventilation rates. Most teams increased CO2
concentration from mid-October onwards. From mid-November towards the end of the experiment, most teams lowered the CO2
dosage. Team Sonoma increased the CO2
concentration continuously during the total cropping period. The Croperators suddenly doubled the dosage and concentration towards the end of the crop (Figure 9
) and were able to catch up with their harvest with that strategy (Figure 3
). Team deep_greens had the highest total dosage, which did not, however, lead to high concentrations, due to an unfavorable ventilation strategy (data not shown).
Fruit growth duration is the time between flowering and harvest of a cucumber fruit, and varied between as little as 11 days to as much as 24 days. The overall average fruit growth duration was 17.3 days, but varied notably between the different teams. Deep_greens had the shortest (13.4 days) and AiCU the longest (21.6 days) average fruit growth duration. This correlates with the average greenhouse temperature during the fruit growth period (Figure 10
). Figure 10
shows the average greenhouse temperature to which fruits were exposed during growth.
shows the amount of heating energy. Team deep_greens show a very different strategy from the others because their algorithm decided to create a very warm air temperature, probably to shorten the fruit development time. This resulted in a high resource use for heating (Table 1
). Together with high daily light sums, mainly from artificial light (Figure 5
b), while blocking natural light (Figure 5
a), especially at the beginning, they were able to have a good harvest during the first weeks, but at the cost of high resource use on electricity (Table 1
). Unfortunately, in October, technical problems (connection of AI remote control) and extremely low irrigation during several days led to a dip in harvest, from which they were not able to catch-up again.
Team AiCU applied relatively low temperatures (Figure 10
) and a low daily light integral (Figure 4
), while maintaining the highest fruit density (number/m2
), which together explain the extremely high fruit abortion rate (Figure 7
) and low light use efficiencies (Figure 6
). The Croperators, due to the detection of small fungal disease spots in their compartment in November, were advised to lower the humidity levels. In order to reduce relative air humidity, they deactivated the misting system (which was intensively used in the second-half of October; data not shown) and started ventilating by opening both lee- and wind-side vents. Setting the minimum temperature of the pipe rail system to 40 °C allowed for maintenance of the air greenhouse temperature (night = 19 °C and day = 25 °C) and prevented it from falling, while ventilating. For this reason, the strategy led to a steep increase of the heating energy demand (Figure 11
) but not to such an evident increase in the air temperature in the same timeframe.
In Figure 12
the course of irrigation supply is given. Notable is the relatively high amount of irrigation supply of team The Croperators, and also the relatively high supply of the reference growers. In this experiment, high drain did not result in high water use because drain water was captured and fed back to the irrigation water, while taking the nutrients from the drain water into account when refilling the irrigation water buffer tank. High or low drain might affect the root quality, since it influences the oxygen availability. However, in this experiment such effects were not analyzed in detail.
In Table 1
the sustainability factors obtained during the growing experiment are presented for each team. In general, team Sonoma was able to realize the lowest resource use for CO2
, heating, and water per kg cucumbers produced (class A + class B). Only on electricity use for artificial light did they realize average values. The reference team of growers obtained the lowest usage of electricity. In total, deep_greens obtained the highest resource use for heating, electricity, and CO2
), which together with low production also led to a low net profit (Table 2
). In Table 2
, costs, income, and net profit are shown for each team. Highest net profit was obtained by team Sonoma, whose AI strategy resulted in a better performance than the manual growers. Sonoma won the challenge.
In order to analyze the different cropping, lighting, and climate control strategies, we used a combination of a dynamic climate and crop model. The validation of model predicted yield per greenhouse compartment and realized yield per team, and thus the greenhouse compartment is shown in Figure A5
. A good agreement was found for all teams and compartments, except for deep_greens. The reasons for the poor agreement of predicted and realized yield for deep_greens were technical problems (connection of AI remote control) and extremely low irrigation over several days, which was not simulated by the combined model. The data of deep_greens is, therefore, not considered further in the analysis.
In Figure 13
and Figure 14
, the results of the strategy analysis on cucumber yield and net profit are shown, respectively. The winning team, Sonoma, could have improved their yield by applying the cropping and climate (temperature and CO2
) strategy of AiCU, whereas AiCU could have improved their yield, and thus net profit, by applying the lighting strategy of Sonoma. From that we conclude that a dense crop with a high number of fruits per m2
is only effective when combined with a high light integral and high levels of temperature and CO2
In summary, most teams were able to obtain a good production and low resource use and reach a net profit close to the performance of manual growers, or even better. All data of the growing experiment is published under doi 10.4121/uuid:e4987a7b-04dd-4c89-9b18-883aad30ba9a (Supplementary Materials
The first successful benchmark experiment on remote control of greenhouse cucumber production with the use of state-of-the-art artificial intelligence algorithms has been carried out. We showed that AI algorithms can compete with experienced manual growers and can even outperform them.
However, more developments towards robustness, scalability, and generalization of algorithms will be essential. For the development of new AI algorithms, a large and complete set of training data with a wide range of variations will be needed. We have shown that artificial training data do provide a first step when real training data is lacking. However, current artificial data do not cover all aspects (e.g., pest and diseases) and provide only a limited description of the actual crop management.
To make the step towards a real “autonomous greenhouse”, it would be required to automate crop registrations and obtain automated data on pest and diseases. Next to that, it might be helpful to explore how manual labor in the greenhouses can be automated or robotized.
In the meantime, AI has affirmed that a combination of high amount of light, temperature, and CO2 at the right moments during the production process are essential to obtain high profits in an autumn cucumber crop cycle. Analyses showed that this strategy would have paid off even more at high plant densities. Manual growers can apply the lessons learned directly to their daily practice, even without AI.
AI-assisted or -managed greenhouse production can potentially improve crop production in locations where knowledge is limited, possibly also in greenhouses operated by highly skilled personnel. The technical make-up of the greenhouse has to meet certain minimum criteria to enable this. For example, sensor and communication facilities have to be in place, the greenhouse construction and installations should enable interventions, and fertigation and crop protections should meet minimum standards. Further development will be needed to make AI a full alternative for the top grower-skills that are nowadays required for (near)-optimal greenhouse production.