Greenhouse horticulture plays an important role in securing year-round production of fresh and healthy products with a continuous, high quality. Worldwide, the area of protected cultivation is increasing. Production systems for vegetables and flowers in greenhouses must be efficient in the use of natural resources, economically viable, and produce a high quality product. This can be achieved very well in protected cultivation. However, the availability of skilled staff with knowledge on high quality and resource use efficient greenhouse growing is scarce, especially in countries where high technology greenhouses have recently been introduced. Therefore, a step towards more automation in cultivation is required. In the AGROS project, linking to the fundamental NWO funded project Synergia, we work on the concept of Technology for Ecology, which implies that we look at the system (greenhouse, farm, dairy) and investigate what information can be sensed and used in control systems that will secure optimal conditions. In the AGROS use case Horticulture, we work on the realization of the 'Autonomous greenhouse' in which cultivation is controlled remotely via intelligent algorithms, using measurements of crop properties and climate monitored with smart sensors and modelled with mechanistic models.
Modern high-tech greenhouses are equipped with active control of actuators (for example heating, lighting, fertigation) to create the desired greenhouse climate for the crop. These intensive cultivation systems require high use of natural resources such as energy, water, CO¬¬2 and nutrients. Current practice is that growers determine the set points for their greenhouse climate based on the crop status and their experience on how crop growth is affected by the climate. Actuators controlled by a greenhouse climate computer, are operated based on these set points and sensors provide feedback on the realized climate to decide the control on the next horizon. In the AGROS “Towards an autonomous greenhouse”, we make the next step towards control of the greenhouse based on production goals rather than on the climate. To realise this, further knowledge on how the crop responds to changing environmental conditions is collected and used to extend crop growth models. Next, the states of environmental (e.g. temperature, Photosynthetic Active Radiation) and crop parameters (sap flow meter, crop weight) as well as of actuators (screen deployment, light operation) is sensed using smart sensors. This information is further used in the feedback loop of model-based intelligent algorithms that will determine optimal strategic and operational decisions of the next horizon, to realixze production goals.
This project contributes to ST1 Smart Technologies in Agri-Horti-Water-Food, From sensing to decision support and digital twins. In the AGROS project “Towards an autonomous greenhouse”, we develop: 1. Non-destructive sensor technologies that measure physiological and morphological plant traits during the course of a cultivation cycle; 2. Smart collection and processing of information from multiple sensors; 3. Control systems (decision support systems) to control greenhouse actuators based on (i) Mechanistical crop and climate models and on (ii) Artificial intelligence algorithms.
World-wide, protected cultivation is important to provide the population with year-round fresh, healthy food, that fulfills a growing demand. However, greenhouse cultivation faces a number of challenges, regarding sustainability (water scarcity, water quality, emission regulations, reduction of CO2 emissions, limitations to the use of energy and electricity, restrictions to the use of chemical crop protection products and growth regulators) and labor (reduced capacity of staff that is willing to do repetitive actions in greenhouses, demand for highly qualified staff that can deal with the complexity of a greenhouse production system). These topics have become even more urgent in 2020, given the COVID-19 pandemic, which has stressed the need for healthy food, that is produced locally (restrictions on transport), the demand for mechanization and automation to reduce the number of greenhouse staff, and the need for sensor-driven, automated greenhouse control, that can be performed from a distance. That makes the AGROS project very timely and important, since it addresses the topics of a sensor-driven, model based, AI driven greenhouse production system, that is controlled based on goal functions set by the grower, rather than on manually determined greenhouse climate setpoints.
The AGROS project “Towards an autonomous greenhouse” will have as final project output a comparison of a greenhouse cucumber cultivation that is controlled by intelligent algorithms with a greenhouse cultivation manually controlled by growers. In order to achieve this, the following steps have to be made, and the accompanying results have to be achieved:
1. List of plant parameters that are relevant in describing the status of the crop. This list will be form the basis for the measurements that will be required in the experiments, and therefore for the selection of sensors to be used (WP 1.1; year 1).
2. List of sensors that are most suited to quantify the desired plant characteristics at the right interval, location and precision to be used in intelligent control of a greenhouse (WP 1.2; year 1).
3. Inventory of control strategies. A questionnaire is sent to growers, with questions on their control strategies, frequency of adjustments in setpoints, etc. This will be used to prioritize the decisions to be made by the control algorithms (WP 1.3; year 1).
4. Exploration and selection of intelligence algorithms. Potentials of separate or hybrid intelligent algorithms will be explored. The best candidate algorithms will be selected on their ability to satisfy a set of criteria on desired algorithms performance (WP 1.3; year 1).
5. Greenhouse trials with cucumber and chrysanthemum, with sensors for selected plant characteristics, including ground-truth measurements, and the development of a deep learning pipeline to predict plant characteristics based on raw multivariate sensor data. In these experiments, the response of these crops to changes in temperature and light conditions will be established, as well as their time constants (WP 1.1 and 1.2; years 1-3).
6. Analysis of plant responses by crop growth model. The sensor data on the crop responses will be incorporated in a combined greenhouse climate - crop growth model, which will provide training data for intelligent algorithms that will be developed to control the greenhouse autonomously (WP 1.1. and 1.2; years 2-3).
7. Insight in state-of the-art control systems, also in other sectors, and identify options for development (WP 1.4, year 2)
8. Development of intelligent algorithms. The intelligent algorithms will obtain experience in climate control and crop growing in an interactive virtual greenhouse environment, simulated by existing well validated mechanistic climate and crop models. The greenhouse virtual twins generated by the models will assist in learning and improving the behaviour of the algorithms (WP 1.3; years 2-3).
9. Process description. Description of the main processes and definition of performance functions, visualization of a simple user interface (WP 1.4; year 2).
10. Functionality tests of intelligent algorithms. Despite the good performance of the developed algorithms in the virtual environment, deviations from the real-world environment necessitate the adaptation of the best algorithm in an experimental setup. Therefore, the algorithm will be tested in small-scale greenhouses to perform functionality tests. Two greenhouse compartments will be controlled by intelligent algorithms under the same objective (optimize net profit) and the performance will be compared with human operated reference compartment (WP 1.1- 1.3; year 3). Based on those, the algorithms will be refined, and the best performing algorithm(s) will be tested in a validation experiment.
11. Final validation trial. In the final validation trial, the performance of one or more algorithms will be demonstrated (WP 1.1- 1.3; year 4).
12. Communication and AI solutions. System for communication: couple data of different types of sensors, taking into account positions and physical influences, coupling actuators based on performance functions (systematic representation). Insight in AI solutions to optimize performance functions, tested adaptive control system, method to train the system (functioning control system) (WP 1.4; year 3).
13. Greenhouse Artificial Intelligence Accumulator (GAIA, prototype) with graphic user interface, generic applicable module for sensitivity analysis of the controller (GAIA in GUI), and GAIA tested in pilot greenhouse with focus on functionality (WP 1.4; year 4).
1. Consortium meetings, twice per year, with all project partners
2. Work package meetings, discussing the progress per WP, with the partners involved.
3. Stakeholder meetings, with growers as supervision group for the experiments, for the climate inventory, for the overarching view on autonomous greenhouses, with companies outside the consortium and other stakeholders to disseminate.