Innovating farming using Computer Vision and Neural Networks.
The ongoing COVID-19 pandemic is endangering the health and livelihoods of millions of people in Asia and MENA regions. Disruption of food supply is a major challenge resulting from COVID-19. It has been estimated that the number of people facing a food crisis will grow from 135 million to 265 million by the end of this year.
Labor for harvesting, processing, transport, and distribution has been threatened by restriction on movement and management of risks of the spread of the virus. New challenges require new methods and innovations of collecting and processing data. Prominent farmtech startup has approached Vostok AI team to create customer Computer Vision & Neural network solution to eliminate intensive manual labor and automatically control the growth of the plants on their farms.
Challenges with Existing Solution
Our company was approached by farmtech and foodtech startup that builds vertical software-controlled farms for growing herbs, vegetables and berries. Company stuck with existing solution and couldn’t scale their farm production without significant increase of manpower to do basic manual labor tasks, such as growth data collection and monitoring.
They asked Vostok AI team to fine tune their software that will allow them to:
Monitor plant’s growth 24x7;
Highlight when growth of one plant is at slower pace than others;
Automatically regulate the microclimate inside the greenhouse;
As a result, have better growth rate using same or less manpower.
Solution.
Our team was very excited to start working on this project, as food safety and foodtech becoming increasingly important aspects of our life during COVID-19.
We have divided project work in few key steps:
Get the data of the average weight/size of the plant for each day of growth;
Create a model that predicts weight depending on the day of the growth;
“Feed” our computer vision model with 5,000 images of plants;
Trained the model to separate each individual plant;
Taught a neural network to predict the weight of each plant;
Finally, launched a Telegram bot to estimate weight by photo for quick test.
Results.
Once project was completed, we got 90%+ accurate model (prediction vs real growth of the plant). This is how upgraded with AI process looks now:
Stationary camera take picture of plants on the shelves;
Images are sent to the AI system;
The neural network compares the parameters of the plant vs forecast;
Neural network indicates if any parameters are beyond forecast;
Vertical Farm system regulates the micro climate in the greenhouse.
As a results of this project, plants need less pesticides and fertilizers. The cost of the salad was also reduced by 15-20%, due to automation of labor work.
For future development we plan to add drone filming to have faster coverage of the plants.