Major companies are bringing together new machine learning algorithms, better and cheaper sensors, and increased computing power in hopes of addressing growing global demand for food and agriculture's diminishing labor force. Alphabet's X and John Deere, startups and universities are looking to AI-based agriculture to address these problems. But farming presents hard problems for AI that, if solved, could ultimately help it be deployed in more structured places.

Machine learning is used to analyze data collected from farmers' fields, satellites and drones and inform decisions about planting and fertilizing, to spot disease, and to try to predict crop yields.

On the ground: AI-enabled equipment is on the market and under development.
- John Deere combines come with an option that uses machine learning to coordinate their spouts with grain-collecting carts to minimize spillage. Another model uses it to assess the quality of grain going into the bin— though still with human oversight.
- Fruits and vegetables are more difficult and labor-intensive crops to harvest. Harvest CROO Robotics is working on a strawberry picker that recognizes and picks ripe fruit — with limited success. "We really take for granted how good humans are at performing fine manipulation tasks. Our bodies provide us with a rich set of data about our environment, and our brains synergistically fuse this sight, touch, smell, and sound with prior experience in a way we're still struggling to understand," Carnegie Mellon University's Tim Mueller-Sim says.
- Chowdhary developed and is testing a robot that can move through rows of plants and use computer vision to measure their height and stem width.

The big field test for AI, though, is whether it can abandon following a script and be trained to adapt to a dirty, messy and uncertain life on the farm. "If you can deploy it in an unstructured environment, it will work in a more structured one," says Mueller-Sim.

Data: Success in computer vision has largely come from deep learning, an AI technique that relies on data with detailed labels and tags. "The challenge is we don't necessarily have that supervised data for ag," Chowdhary says.

But it's surmountable, Mueller-Sim says, pointing to soon-to-be published work on generative adversarial networks (GANs) that, trained on 40–60 images, can detect features of sorghum, grape vines and cannabis flowers in the field.

Uncertainty: Robots struggle with change. Soil texture, glare, clouds and other variables can all interfere with movement and computer vision, particularly as it tries to move towards driverless tech.

"We're trying to put robots in environments that aren’t meant for them," Chowdhary says. "Cars are meant to be on roads and even that is hard to make autonomous."

Driverless tech is making components for automation more accessible, but agriculture has unique challenges. It's off-road, and the machines are much bigger. "The winners in driverless tech aren’t necessarily the winners in ag," says Alex Purdy, head of John Deere Labs.

Variation: Another challenge for AI in agriculture — and more broadly — is developing it to be accurate within and across different fields.

"Not every plant looks the same, even when planted side by side. So when it comes to the task of manipulating a robot arm for harvesting fruit or cutting a vine, the challenge is working in an unstructured environment," says Carnegie Mellon's Abhisesh Silwal, who is working on robots to prune grapes and harvest apples.

One possibility is develop algorithms that account for the differences between where they were trained and where they are operating. "That sounds okay qualitatively but to get it into the mathematics we still have a lot to do," Chowdhary says.

Adoption: Farmers often rely on generations of knowledge about their land. So, new — and often expensive — technologies will have to prove their worth.

Source: https://www.axios.com