In order to boost the participation of the AI community, an Online Challenge will be organized. The goal will be to invite the AI community interested in AI for horticulture and to motivate them in participating in the Hackathon and Greenhouse Growing Challenge.
Recruitment of AI teams and scouting of talents
We believe the following expertise from AI community is needed towards fully autonomous crop control: Machine learning skills and Computer vision skills. Machine learning skills will be tested in the interaction with a lettuce growing simulator. The simulator will consist of a simple greenhouse climate and crop production model that will be provided. Computer Vision skills will be tested on real lettuce images. A series of annotated images will be provided as training dataset.
Online Challenge for AI experts
The first part of the 3rd edition of the Autonomous Greenhouse Challenge takes place from 1 June to 14 July as open Online Challenge, aiming at testing machine learning and computer vision skills of participants of the AI community.
In Part A - the computer vision challenge - teams will get access to a series of lettuce plants. Images are taken with a RealSense camera under defined conditions and contain images of individual lettuce plants of different varieties in different growth stages and grown in different growing conditions. Each image is connected with information on the ground truth plant traits, such as plant diameter, plant height, plant fresh weight, plant dry weight, and leaf area. Teams use ca. 300 images provided in batches to develop a computer vision algorithm during the preparation phase. This algorithm will have to be able to estimate the plant traits of a series of ca. 50 unseen lettuce plant images provided during the Online Challenge under limited time and memory constraints. The computer vision algorithms have to detect the plant parameters described above.
In Part B – the machine learning challenge - teams will get access to a virtual simple greenhouse climate and lettuce production model (simple simulator). The simple simulator consists of a given set of outside climate conditions, a given greenhouse type and given greenhouse actuators (ventilation, heating, lighting, screening). It needs to be provided with a series of climate setpoints (ventilation strategy, heating strategy, lighting strategy, screening strategy per timestep) as inputs. The input climate setpoints will activate the available virtual actuators, which will control the inside greenhouse climate. The realised inside climate parameters will be provided as a feed back value.
Since the crop growth in the simulator is determined by the realised greenhouse climate, also the crop growth parameters (fresh weight, height, diameter) over time will be provided as output. Teams will have to develop machine learning algorithms to feed the simple simulator with the optimised control parameters in order to maximise net profit. During the preparation phase teams can interact with the simple simulator for algorithm development. During the Online Challenge this algorithm should be suitable to control the growth of a virtual crop in a virtual greenhouse under changed conditions (e.g. other weather conditions, different greenhouse type, different lettuce type) and limited time constraints.
Teams and results
In total 46 teams from 24 countries, with 286 participants took part in the Online Challenge. Teams had different backgrounds and were experts from start-ups and companies, students and researchers from universities, research centers.
Team 'Koala' from the U.S. won the Online Challenge. They achieved a total of 85 (out of 90) points and were first in the machine learning challenge (with 45 points) and sixth in the computer vision challenge (with 40 points). The team built an algorithm that realized a virtual net profit of €8.68 per m2 and cultivation period in a simulator that produced virtual lettuce plants in a virtual greenhouse. In addition, their computer vision algorithm was able to recognize lettuce images with a high accuracy (total error=0.094) and estimated the correct growth parameters of lettuce plants.
Team captain of team 'Koala' is Kenneth Tran, he also led the winning team 'Sonoma' in the first edition of the Autonomous Greenhouse Challenge in 2019. Congratulations to him and all team members (Neil Mattson, Minh Duong, Hanh Bui, Tim Shelford and Michael Eaton) with this first place. The motto of team 'Koala' - "I have all the Koalifications" - seemed apt. Team 'Koala' will receive a wild card to participate in the next phase of the Autonomous Greenhouse Challenge, the greenhouse experiment that will start early 2022. Participating teams will then each have a greenhouse department to their disposal at the business unit Horticulture of WUR in Bleiswijk to grow lettuce fully autonomously in reality.
Second was team 'CVA' from South-Korea led by Hee Kyung Ryoo, third was team 'IUACAAS.ICANnettuce' form China led by Xiao Yang. More information about all teams and all results can be found on our website www.autonomousgreenhouses.com.
Lettuce images were made available by WUR and will also be made publicly accessible after the Autonomous Greenhouse Challenge. The simulator used for a virtual lettuce greenhouse and cultivation has also been developed by WUR and will be made accessible again to participating teams in the next phase of the Autonomous Greenhouse Challenge.
Here are the results of the Online Challenge: