Farming With AI

Farming has been getting “smarter” for some time, with sensors, the Internet of Things, spectrally tuned lighting, robotic machinery, autonomous vehicles, and information networks improving the cultivation and raising of crops and livestock.

Shouldn’t artificial intelligence (AI) also make it into the fields?

It has. Vendors ranging from computing stalwarts such as IBM to lesser-known agricultural and environmental technology specialists including Trapview, Carbon Robotics, CropX, and others are now applying the AI stock in trade of deploying vast amounts of data to train a model to learn, act, or create. In the case of farming, AI is trained on data about weather, climate, insects, soil conditions, feed stores, historic crop yields, and so forth to provide farmers with insights into everything from when to plant and irrigate, and to warn them of impending infestations and natural disasters such as droughts and floods.

In the bigger picture, not only do these insights in principle benefit the business of the farmer, but they also address the issues of global food security, hunger, nutrition, and sustainable production.

“The emergence of rapidly evolving technologies, such as AI, offers agriculture players another powerful tool to meet these challenges head on and unlock greater efficiency and effectiveness throughout their businesses,” noted New York City-based consulting giant McKinsey & Company in an online report on AI in the farming industry. McKinsey cited the “high volumes of unstructured data” as a leading reason why “agriculture is particularly well suited for disruption by AI.” After all, AI is, by its inherent way of working, a massive structured dataset that can deliver information, insights, and reports for farmers.

When it comes to farming, there is perhaps no area of data that is more broadly applicable than weather and climate. Temperature, humidity, precipitation and the like have long been part of the agricultural lexicon, and they now form an important chunk (but by no means the only one) of the AI assistance coming their way.

The Algorithms vs. the Insects

Take, for instance, intelligence about destructive insects that Slovenia’s Trapview provides to farmers looking to ward off the likes of diamondback moths, bollworms, navel orangeworms, spotted-wing drosophila flies, codling moths, and some 50 other villains that can annihilate the likes of cabbage, lettuce, corn, soybeans, citrus, almonds, blueberries, and apples, to name a few.

Trapview gathers historical information about how these targeted insects have fared in the past under different weather conditions and combines it with real-time weather information from farmers’ fields about current temperatures, humidity levels, and other conditions. With these curated datasets, “We can determine, for example, how long it will take for the insects to hatch if you have temperature ranging from, say, 20 to 25 degrees [Celsius],” said Trapview chief technology officer Matej Stefancic.

That would not be possible, however, without the company’s other continual deployment of AI. Trapview, based in Hrusevje, about 40 miles southwest of Slovenian capital Ljubljana, has amassed a collection of about 30 million insect images, which it compares against images of insects it traps in farmers’ fields. It transmits pictures of the trapped arthropods via wireless connections to the cloud. There, using proprietary algorithms in a neural network that it hangs off of a Google TensorFlow 2 framework, Trapview can, in a deep learning process, quickly identify whether the trapped insects are the type the farmer is on the lookout for, looking at distinguishing factors such as color, shape, and size.

“In the past, there might have been 5, 10, 20, 30 different parameters on which I will identify whether that’s what I’m looking at or not,” said Stefancic. “With the introduction of neural networks, that’s changed. Basically, the decision tree now has thousands of parameters, or sometimes tens of thousands or hundreds of thousands of parameters.”

Trapview generally looks for about 60 insect species commonly associated with infestations out of the nearly one million known species of insects. Once it confirms the identity of the pest, it combines that information with the weather and climate data to ascertain the likelihood the insect has laid eggs, and when those eggs might hatch into the larvae (typically caterpillar-like creatures) responsible for crop damage.

With many farms having turned against chemical pesticides in favor of releasing natural insect predators such as ladybugs, the timeliness of that information is critical to growers.

“I want to tell them where, when, and if the pest is your problem,” said Stefancic. “And then you can see the development [of the insect stages] for the next 7 to 10 days. That’s the result of how we combine hardware, biology, AI, statistics, and data management.”

For all of its sophisticated AI, the Trapview process typically begins with something far more rudimentary: sticky paper mounted inside of its trap boxes, which also feature attractants such as sex pheromones and foods.

“That’s our source of raw data: pictures of insects stuck in the trap,” said Stefancic.

Planting Layers of AI

Like with Trapview, weather plays a big role in the AI software that IBM is providing to the farming industry. In another similarity to Trapview, IBM often relies on data drawn not from the language and text world, but typically from images and geospatial sources including, among others, images taken by the European Space Agency’s Sentinel-2 satellites.

The company’s Environmental Intelligence suite, introduced last May by IBM Software for farming and other environmental use, includes two categories of products, both of which make use of AI.

In what IBM refers to as the “data” aspect, the company sometimes applies AI (and sometimes applies non-AI means) to make sense of the huge amounts of data on which it trains its platform to assist farmers in making decisions across crop management such as planting, watering, fertilizing, pest protection, and so forth.

In one such example, IBM Sustainability Accelerator, a pro bono group working in partnership with Buenos Aires-based non-profit Plan21 Foundation and the Costa Rica Institute of Technology, claims to be using such data to help hundreds of smallholder farmers in Costa Rica, Ecuador, Colombia, Chile, and Argentina to improve the yields of crops including coffee, yuca, bananas, and cacao.

The Latin American endeavor also is making use of the other AI aspect of the Environmental Intelligence suite, which uses the data to train an AI model of action. In the Latin American project, that model is an app that farmers use for advice on crop management. IBM has provided a similar set of capabilities to farmers in Malawi, stopping short of providing an AI model, but using AI-managed data about climate, weather, and soil to send farmers advice via text messages.

In the most recent example of an IBM AI model for climate insights, the company in September announced a foundation model that it co-developed with NASA that assists in weather and climate modelling for a wide community of users, not just farmers, who want to ascertain climate likelihoods, risks, and hazards. “Foundation model” is an AI term denoting a base on which to build an application.

In building the foundation model, IBM used AI transformer architecture, but rather than using the Large Language Model associated with AI software such as ChatGPT, it went instead with images.

“We’ve created a foundation model that uses those same technologies, such as transformer technology, but we’re applying it to remote sensing imagery rather than text or code,” said David Blanch, IBM’s director of product management for ESG (Environmental, Social and Governance) and Environmental Intelligence. “It’s a geospatial foundation model. Instead of taking text and producing text, I’m taking in geospatial type data.”

Farmer Talk

The term “geospatial data” is not one you might readily associate with the world of farming, but it and other expressions associated with the AI world are indeed furrowing their way in.

Seattle-based Carbon Robotics claims to be using “AI, deep learning, and computer vision” in its “laser weeder,” a tractor-attachable box powered by Nvidia GPUs that scans soil with high-resolution cameras looking for what the AI identifies as weeds, and then goes after them. Israel’s CropX says on its website that it provides “an easy-to-use yet powerful agronomic farm management system that uses AI and machine learning to connect farm data, real-time conditions, and agronomic knowledge, providing guidance for successful and sustainable farming, while aggregating all agronomic farm data in one place for easy tracking and sharing.”

It would all generally make sense to McKinsey, which estimates that globally, the use of AI could add $100 billion in value to farms by improving yields and cutting costs.

This is not your father’s Massey Ferguson.

Mark Halper is a freelance journalist based near Bristol, U.K. He covers everything from media moguls to subatomic particles.