The AGROS project | Works at an independent greenhouse

Greenhouse horticulture plays a key role in the production of fresh fruit and vegetables for an ever-growing population. In recent decades, greenhouses have grown in size and cultivation has become increasingly complex.

Growers today must balance production with the use of energy, water and nutrients. On top of it all, there is a lack of skilled workers who can oversee all the complex processes in a greenhouse. A grower must determine the right settings for all parameters at all times: heating, ventilation, dehumidification, shade, artificial light, crop management, monitoring for pests, predators and diseases or precision spraying are just a few of the decisions to be made. A well-educated and highly experienced grower can oversee most aspects of such a system. However, we will soon have too few of these highly trained growers worldwide.

Autonomous greenhouse

“What we need is an independent greenhouse,” says Anja Dieleman, AGROS project manager and researcher at WUR Horticulture. “In such a greenhouse, the crop is central, and cultivation is regulated based on predetermined goals. For that, you need extensive knowledge of the physiology of the crop, but also precise sensors that can measure relevant crop characteristics and intelligent algorithms to be able to control the greenhouse autonomously.”
To supply an ever-growing population with fresh fruit and vegetables in the future, new production systems must be developed that can be operated by less experienced growers and even non-agricultural parties, anywhere in the world.

In the AGROS project ‘Towards an autonomous greenhouse’, researchers from WUR Greenhouse Horticulture are working together with business partners to realize fully automatic cultivation in greenhouses. Anja: “We have taken steps from data collection with sensors to data-driven support for cultivation and development of intelligent algorithms. We hope to soon be able to apply them and further improve sustainable production systems for healthy and fresh foods.”
The research team’s vision for the future is an independent greenhouse, where the expertise of an experienced grower is replaced by artificial intelligence. With new model-based control algorithms, the conditions in the greenhouse can be adjusted autonomously to achieve the cultivation goals.

Autonomous greenhouse challenge

In 2021/22, WUR organized the 3rd Autonomous Greenhouse Challenge. To stimulate the AI ​​community’s participation, an Online Challenge was organized this time. The first and second editions of the Autonomous Greenhouse Challenge have shown that artificial intelligence can potentially be superior to human intelligence and thus potentially master indoor agriculture in the future.

The goal of the third challenge is fully automated control, and this edition’s crop was lettuce. Grown completely independently! The challenge was divided into two parts. In part A – the computer vision challenge – teams were given access to a number of lettuce plants. The images were taken with a RealSense camera under defined conditions and contain images of individual lettuce plants of different varieties at different stages of growth, grown under different growing conditions. Each image is associated with information about the plant’s ground-truth characteristics, such as plant diameter, plant height, plant fresh weight, plant dry weight, and leaf area.

The teams use approximately 300 images in batches to develop a computer vision algorithm during the preparation phase. This algorithm had to be able to estimate the plant characteristics of a series of approximately 50 unseen lettuce plant images provided during the online challenge, under time and memory constraints. The computer vision algorithms had to detect the plant parameters described above.
In Part B – the machine learning challenge – teams were given access to a practically simple greenhouse climate and lettuce production model. This simulator consisted of a specific set of outdoor climate conditions, a specific greenhouse type, and given greenhouse actuators (ventilation, heating, lighting, shielding). It had to be provided with a number of climate set points (ventilation strategy, heating strategy, lighting strategy, shielding strategy per time step) as input. The input climate setpoints activate the available virtual actuators which control the indoor climate of the greenhouse. The realized indoor climate parameters are given as feedback value.

The teams had to develop machine learning algorithms to provide the simulator with the optimized operating parameters with the objective of maximizing the net profit. During the preparation phase, the teams were able to interact with the algorithm development simulator. During the Online Challenge, this algorithm had to be suitable for growing a virtual crop in a virtual greenhouse under changing circumstances (eg different weather conditions, different greenhouse types, different greenhouse types, different lettuce varieties) and limited time pressure.

The teams had to grow lettuce with a target weight of 250 grams. Quality was also assessed. If the plants were too small, had leaf tip burns or other deformities, they were classified as lower price grade B or even unmarketable grade C. If the plants were too large, teams wasted resources. The last team finished their harvest on June 17th. Resource consumption (e.g. heat energy, electricity, CO2) was measured during the growth period and operating costs were calculated. The fixed costs depended on the covering of the greenhouse space and the use of various installations (e.g. artificial light). The net profit was determined from these figures.
In a first cultivation cycle in February/March this year, each team was able to test their algorithm and procedure. The real challenge was another crop cycle in May/June. Teams could no longer access their algorithm after the experiment started, but had to request permission if they needed to make urgent changes (bug fixes) to their algorithms. Admission was charged and expenses were deducted from the net profit. The winning team only accessed the virtual machine once to fix a minor bug.

Team ‘Koala’ from the USA won the Online Challenge. The team built an algorithm that achieved a virtual net profit of €8.68 per m2 and cultivation period. In addition, their algorithm was able to recognize lettuce images with high accuracy (total error = 0.094) and estimate the correct growth parameters of lettuce plants. Team Koala’s ambition is to advance greenhouse horticulture with intelligent automation technology that is scalable to farms, crops and even more broadly to process-based manufacturing industries.

Autonomous greenhouses can ensure that more people are fed vitamin- and mineral-rich products. In addition, these techniques contribute to increasing food security and a higher production volume of healthy vegetables, using fewer resources such as energy. Its potential has been successfully demonstrated in previous editions of the Autonomous Greenhouse Challenge.

Leave a Comment