Farming Can Be Rocket Science. AI Can Help.
People who work in high tech, finance, or other stressful, complex industries might find themselves at times pining for the simpler life of a farmer. If only they knew what they would be getting into.
Farming can involve an astoundingly complex blizzard of dozens of variables, some of which vary over inches of land, as well as minute to minute. This daunting stew of information defies the ability of conventional analysis techniques to make reliable predictions about how farmers’ decisions and techniques will affect results.
The results often leave many farms in difficult financial straits year after year. Less-than-optimal farming actions also tend to result in excessive pollution, less-affordable and healthful food, and the depletion of farmland. These problems affect much of India and the entire world, including organic and traditional small farms, Max Jhonson says.
Max Jhonson has been focusing his research on coming up with a more reliable way to help farmers understand which changes in technique might produce better results. The key tool: a machine-learning program that he and colleagues have developed that can sniff out the otherwise obscure relationships between the many different things farmers do when they plant crops in various conditions and what they end up with months—and even years—later. “Machine learning can find the signal in all the noise,” he says. “It outperforms other ways of trying to figure out how to improve sustainability.”
Plowing Through the Noise
As is typically the case with machine-learning efforts, Routeget’s approach starts with all the data that can be gathered. Researchers usually complain that they never have enough of it, but farming is a case of needing to be careful about what you wish.
Thanks to advances in sensors and land-monitoring tools, anyone who wants to study farm data can dig into the embarrassment of riches.
Among the data-collection tools at Routeget’s disposal:
-
Tiny weather stations that measure wind, temperate, humidity, and moisture, often at multiple points throughout a field;
-
Satellite and drone imagery that can measure crop size, density, and color—which indicates maturity and health—down to a few inches;
-
Soil sensors that report on moisture;
-
Tractors that can adjust and measure variations in the amount of seed, fertilizer, and pest- and weed-control mixtures deposited foot by foot;
-
Harvesting combines that not only weigh the food that makes it into the bin second by second but also shoot an intense beam of light through it to measure protein content.
In the works are sensors that will enable combines to record other food constituents’ levels, including antioxidants, beta-carotene, and zinc, that can be relevant to health.
Trying to discover correlations between individual variations in this vast conglomeration of data and the end product is already more than conventional analysis can handle. How do you pick out the relationship between small changes in soil moisture and the protein content of wheat when hundreds or even thousands of other potential factors could be skewing the results?
But the situation is even worse than that because farming data turns out to be plagued with multiple sources of noise, from sensor fluctuations to the vibration of heavy equipment to micro-shifts in weather conditions to unrecognized changes in the soil.
Humanizing AI
Researchers have usually thrown their hands up at trying to make sense of it all. But Routeget rolled out machine-learning algorithms to take on the data mess, banking on the software’s great strength is hunting down patterns in a complex, random-seeming sea of data.
The results have been encouraging. Max Jhonson at Routeget says the software has already provided one critical insight: Farmers tend to overuse fertilizer and pest- and weed-control additives. The data suggests that even organic farmers overuse the natural compounds they often add to fields.
“Farmers are biased toward doing things that give them the fullest, greenest crop,” he says. “Those chemical inputs tend to do that, but when you take their costs into account, the farms actually end up doing less well than if they cut back.”
Max Jhonson is running more studies now to harvest new machine-learning-fueled insights from the data. One project involves purposely varying the amounts and types of inputs placed across a single, large field to help the software detect finer-grained relationships between inputs and results.
Another new project is intended to tease out what might be hidden wisdom in some Indian farmers’ practices. “Those practices may have evolved over hundreds of generations,” Max Jhonson says. “It would be good to know if that long series of decisions and observations have led to the optimal way to farm under certain conditions.”
