iFarm together with EORA have completed the development of a new neural network model, allowing growers to determine the weight of a plant by its photo. The autonomous surveillance system for plant monitoring is based on artificial intelligence technologies, computer vision in particular. Cameras fixed in certain places or placed on autonomous drones, continuously monitor the plants in the farm, while the neural network analyses the received images and reports any discrepancies in plant matter growth.
Having a cloud platform in place allows iFarm to save on hiring highly paid specialists (agronomists, agrochemists, specialists on plant protection, engineers, etc.). It contains ready-made recipes for plants as well as clear instructions for growers provided in a form of an interactive list of simple tasks.
Apart from that, the system provides instruments of microclimate management (CO2 level, temperature, humidity, lighting, feeding) and production processes. However, the visual control has until recently remained upon the farm associates, who had to physically walk and keep an eye on thousands square meters of an iFarm.
The neural network already could determine sick plants by photo. Now, together with EORA the company created a new model of a new neural network, making it possible to determine the weight of a plant by photo. Telegram application interfaces were used for testing the system. The data on weight discrepancies came in the form of notifications within the iFarm Growtune system, after which the growth program is adjusted automatically or manually with the help of iFarm specialists.
EORA specialists started with taking controlling photos of plants every minute. At the same time, control plants from identical shelves not equipped with cameras were weighed every three days. The second stage was to remove the distortion on photos, which made it harder to compare the plant sizes. In order to train the neural network, a comparison between a plant on a photo and its mass needed to be made. EORA specialists decided to use the surface of the plant as the main criterion.
During the project, EORA have identified several distinctive features that are important and significant for discrepancies identification:
- Plant surface on a photo (received from cameras)
- Days after the seedling (input in iFarm Growtune)
- Lighting power (input in the crop growing recipe, plus the data received from the controllers)
Thus, the neural network takes in photos and shelf identification, type of lighting (power) and number of days after the seeding. After that, the neural network model defines the plant mass and passes over each plant, entering weight data and the overall plant weight into the system. Also, the expected weight for a certain day is predicted based on the statistics. The data is compared and in case of a discrepancy, the system “raises a flag”.