The model was first trained using 3,000 images of the plants and their height. They then re-trained it a month later on 6,000 images and the results were significantly improved. The model’s prediction for how high the plant is in the photo is now very similar to the human measurements for this. Using the model’s prediction for the height of the plant, and then combining this with other data points relating to the plant, Svegro now receives an estimated final height for the plant on the date of planned harvest. They also have dashboards with the results, alongside measurements of carbon dioxide, temperature, light and conductivity levels in the soil.
Mikael Huss, Senior Data Scientist at Codon, explains that the choice of Peltarion for setting up the model was obvious when they learnt how fast it was. ‘We tried both doing it through Keras and through the platform, but setting it up in Peltarion was so quick that it became the obvious choice for the project.’
Lisa Lindström, head of cultivation at Svegro, says that it has been remarkable to see the types of conclusions that the consultants from Codon have been able to draw based on just the data. ‘Because of the Corona pandemic, they have never actually been to the site, but just by looking at the data they have been able to work out things about the greenhouse that only employees that have worked here for many years know about.’ She adds that even though this part of the project is only the first step, it has already made a big difference in how they plan their work. ‘The work the people at Codon have done with the Peltarion platform enables us to predict the heights of each plant based on a photo, and thereby removes a big chunk of our manual work here at Svegro. We’re very excited about taking this to the next level.’’