When AI and agriculture collide, opportunity risesby Peter Dorfman Creative Services
By the middle of this century, the world’s farmers will feed almost 10 billion consumers. Mechanization, satellite surveillance, genetic engineering, meteorological advances, soil sensor and galloping automation all have more than tripled global agricultural production since 1960. That’s good news – in the aggregate. For the individual farmer, the picture is, well, it’s complicated.
The modern farm is data-rich but resource-constrained. The average size of a U.S. farm, in cropland acreage, has doubled since the 1980s, driven by scale economies in investment in technology, seed, chemicals and fertilizer, and the trend toward larger contracts for crop sales. But investment is dicey. Large farms produce more but are subject to boom and bust cycles. Overproduction in the last several years has put downward pressure on grain prices. Margins are thin, and farmers are seeking solutions that can make them more competitive while reining in production costs – especially now, as they wait for prices to go higher.
One start-up, Visual Farms LLC , based in Fort Collins, CO., is literally betting the farm on artificial intelligence (AI). And they are putting a stake in the ground to prove it.
The company, (founded in 2008 BY Margaret Maizel, a long time Ag information expert) is conducting field trials in three states to demonstrate that its proprietary AI solution can substantially boost a farm’s grain yields by optimizing seed selection and field management, with data-driven advice tailored to the individual field level. The result, the company believes, will be consistently greater revenue per acre at lower input cost. “We have an audacious goal: to have a Machine Learning system on every farm. This is key because every farm is different and every farmer needs to be able to manage the growing amounts of data that is specific to his own farm”, she says. New AI technologies are beginning to make that possible.”
Preliminary field tests in 2016 were so encouraging that Visual Farms has elected to kick off a 2017 field test with an audacious announcement: In early August, the company will publish its advance projections for the October yields for its test sites in three states across the US Corn Belt. “In last year’s field test in Indiana, our advance projection, in bushels per acre, was within 2% of the actual yield,” says Jim Hall, the data scientist who designed the algorithm that drives Visual Farms’ solution. “We’re betting we can do at least that well in our new trials.”
AI is largely about recognizing and analyzing patterns in complex data. Intuition will suggest that crop yield must be a function of rainfall, the length of the growing season, temperatures, soil quality, insect and microbial parasites, seed genetics and other such factors. All of these are parameters that vary from season to season but have some consistency in any given region and for each crop.
The data set required to make useful inferences is enormous. For a human farmer, decision-making is more art than science. But for today’s AI, predicting a particular farm’s yield is a fairly straightforward application, Hall says. And yield prediction is just the beginning.
The 2016 trial was conducted at Fairholme Farms, a 1,850-acre grain operation in Lewisville, IN. Visual Farms approached Kim Drackett, a co-owner of the farm, with a proposal to use its AI model to help Fairholme Farms to select more productive seed hybrids from the tens of thousands of commercially available seed types, based on the soil chemistry, climate and moisture conditions specific not just to Lewisville but to Drackett’s own fields.
Drackett was sufficiently intrigued to fund the trial himself. He’s now an investor in Visual Farms. And he’s committed Fairholme to participate in the 2017 trials.
“We are going to announce the yield projections from four farming areas,” says Ron Olson, a consultant the Visual Farms. “In addition to Fairholme, we have two locations in Illinois. We are working on three individual farms, six individual fields — two fields on each farm. We also will test a four-county area in Iowa, making county-level projections.”
“Visual Farms is going to provide yield estimates at several points during the growing season and then right before harvest,” Kim Drackett says. “When we go through the field with a yield monitor, we’ll know whether they’ve achieved what they think their algorithm can do.”
A Unique Data Resource
Several things differentiate Visual Farms‘ technology from those of other companies trying to apply AI to agriculture problems. One is the sheer volume of data. The company has compiled 10 years’ yield test data from across the US from public sources. “A given seed variety will be tested side-by-side with other varieties, for a given location and the same conditions,” Hall explains. “Our initial data include corn, soybeans, and wheat. Over the last 10 years, there have been about 25,000 different varieties of corn sold. Some of the data have been published by universities and some by seed companies or from farmers pooling their results. We have the data to make about 20 million comparisons among corn seeds, where one variety is being compared to another variety in the same location, under the same climatic conditions. That’s an order of magnitude more data than our next competitor would have.”
The company also has compiled a history of all of the seeds that have been sold commercially in the last 10 years, and their genetic traits. And Virtual Farms has built what it believes is the nation’s largest agri-climate database. “A plant takes in carbon dioxide from the air, nutrients and moisture from the soil, and energy from sunlight — and each plant is adapted for certain temperature range,” Hall says. “Corn is one of the most researched plants of the world, and there are very detailed models for it. If you have accurate information for how much water and nutrients are in the soil, and how much solar radiation is hitting the plant, and what the temperature range is for that field, you can predict very accurately what that plant is going to do — how many carbohydrates it can make and so forth. The challenge is to calibrate these models geographically, so that you know the soil and climate data relate, with high resolution, to a particular patch of ground.”
The goal for Visual Farms is to provide tools and services to help the individual grower make more informed decisions to maximize profit. Those decisions will include which seed varieties are optimal for the location and seasonal conditions, benchmarked against the yields from neighboring farms; when to plant; what fertilizers to apply and when; and how to harvest and market the crop.
The initial product will be a thin client, cloud-based application — a hybrid seed selection tool that enables a farmer to choose the best hybrids for a specific farm.
“If you came to our farm before we select seat hybrids for the year, you would see a pile of seed catalogs and test plot data and university trial results,” Drackett notes. “It’s extremely difficult to get your arms around. Visual Farms can draw on an incredible range of test plot data. They can say, ‘you’ve chosen to use hybrid X, but given your soils and your geographic location and your cultural practices, hybrid Y can give you 10% higher yield, or 8% higher yield with better grain quality, or 8% higher yield with lower moisture content at harvest.’ All of those things have economic implications.”
Beyond Seed Selection
Smarter seed selection is tremendously valuable, but there are implications beyond that. The tools will enable the farmer to create a model that he can use to maximize yields during the growing season. In effect, the farmer can virtually plant the crop before physically planting the crop. The simulations are reminiscent of computer games popular in the 1990s, simulating the development of cities, civilizations and farms, and allowing players to tinker with the conditions and observe the changing outcomes. “It’s the same thing, but it’s real, with real data and better models,” Hall suggests.
The farm will be able to project yields for 2018, and use those projections to do early budgeting and planning. Once the crop is in the ground, in April, the farmer can take measurements of actual rainfall and sunlight and so on, and tweak their models as the operation tracks the progress of the crop over the season. The farm will be able to project the cost-effectiveness of putting down extra nitrogen or a fungicide, using the AI model to make those decisions.
“Potentially,” Kim Drackett suggests, “the model could give us an alert that on field 14, which the system knows was planted with a certain seed variety on such and such date, now would be a good time to go and scout it for insects, based on the growth phase for that field. Helping to target field scouting to identify problems before they become economic issues would be very significant.”
The “hairy audacious goal” for Visual Farms is to put machine learning tools on every farm in the United States. For the end user farms, however, the goal is much more pragmatic.
“Producing food and fiber with fewer land resources — fewer chemicals, less fertilizer — that’s the objective, for the human race, really,” Drackett says. “Right now all of this is a blend of art and science. We are trying to push the science farther and make it less of a guessing game.”