2 minute read

Agriculture is deep into machine learning

Those who don't pay attention to agriculture may assume that farming remains a mechanized, low-tech endeavor. But in reality, nothing could be further from the truth.

We've been directly involved with a number of ag tech vendors and growers who are making serious investments in machine learning. It's fascinating to think about where these investments might take agriculture and the business of creating food.

Let's start with the big picture. Literally. Using appropriately labeled and annotated drone-captured images and video to train machine learning algorithms is making it possible to

  • find and remove rocks from growers' fields
  • identify and mitigate water distribution issues
  • spot and address soil composition and quality issues
  • optimize agricultural space utilization
  • precisely time livestock grazing land decisions

Labeling and annotating high-resolution aerial images of large spaces is a complex challenge. Think about what's involved in teaching a system to find rocks in a field.

It's not as simple as saying "Hmm, ok, I'm looking at this picture of a 30-acre field and I can definitely say there are rocks there. Done, show me the next image..." Rather, all of the thousands of rocks in that 30-acre that are big enough to require removal have to be distinguished from same-sized dirt clods or shadows. And then the rocks have to be individually annotated. Each rock's location has to be precisely captured so that it can be retrieved and removed.

Rather than have a single person spend many, many hours finding all of the rocks in a high-res image, the task should probably be distributed to a bunch of people, who can each work on a small section of the field in parallel. But then you need to know who has done what, for purposes of grading accuracy and efficiency, and you have to reassemble the work of many people back into a coherent view of the field.

Ideally, some machine intelligence should be watching these human annotators and learning how they identify and annotate rocks, so that eventually the annotation roles of human and machine can be flipped.

All of the same considerations and challenges confront agriculture on the ground.

Someday soon there will be implements for orchards that can tell the difference in real time between healthy and diseased plants. That can differentiate among things that should be dosed with pesticide or nutrients. That can determine whether a particular apple is ripe and ready to harvest.

We can provide machines with this kind of intelligence, but it takes massive amount of labeled and annotated data. As with aerial imagery, this kind of labeling and annotation is complex and often subjective. Trained humans are needed to kick off this process. And technology that can learn from these humans - and assume much if not most of the annotation - is required in order to generate the volume of training data required for acceptable implement decision making.

This, obviously, is our business, and we bring these abilities to virtually every industry. But we get a special joy from adding value to agriculture, one of the most essential of industries.

What do you think about the ways growers and their vendors are incorporating machine intelligence into food production? Are there other agricultural use cases you can share? Drop me a line and share your thoughts. 

Request a demo