One of the major uses of machine learning has been in the agricultural sector. Applications converge with drones determining weather, if water and machinery is accessible in certain locations and if pests or nutrient deficiencies in the soil will make it difficult to grow crops. All of these pieces are put into an adaptive algorithm which self-corrects and “learns” where farming will be most successful.
Almonds are a tricky crop to grow, they require a lot of water and can only thrive in a very specific climate. Central California’s climate it ideal for them, and the region grows most of the almonds that are consumed in the United States. However, the area relies on sophisticated and expensive irrigation since most of the state’s water is located in the northern region. It’s difficult for farmers to get the right range of soil moisture for growth and operators can’t simply turn on irrigation systems like a faucet, they must plan ahead to ensure availability and compliance with governmental regulations. It’s difficult to predict the amount of water needed at a given time due to temperature variations, historical data and data from sensors at a current time. Too often, almond farmers have produced sub-optimal results or lost crops due to human error.
Machine learning can be used to solve this problem. Z-Farm, an almond cooperative, partnered with ThingWorx Machine Learning to run on top of the Internet of Things (IoT) databases detecting patterns, abnormalities and to predict the need for water and automating the solution from start to finish. Within the first day, ThingWorx discovered that farmers were over-watering their crops on the hottest days due to solar radiation penetrating the soil rather than the soil moisture itself. This alone is saving the company water and money while yielding higher-quality crops.
Beyond planning and growing, machine learning can be used for sorting crops too, saving farmers time. Japan doesn’t have a standardized classification model for cucumbers and Makoto Koike, a former embedded systems designer in the automotive industry, has helped his parents come up with an automated way to sort their harvest into nine categories — something that can take up to nine hours per day when done manually. Koike uses Raspberry Pi 3 as the main controller to take photos of the cucumber and then sends the data to Google’s TensorFlow to determine whether or not the image is a cucumber. The image is the forwarded to a larger TensorFlow network running on a Linux server for further classification. The system works well, but can still be improved. Koike reported a 95 percent success rate with image recognition but only a 70 percent success rate with real use cases.