The Latest VC Craze in GreenTech, Computer Vision, and Deep Learning? A $100B Bet on Corn

Google+ Pinterest LinkedIn Tumblr +

Crop yield protection is a vital part of the agriculture supply chain process. Yield protection insurance is regulated by the U.S. Department of Agriculture (USDA), and every August a corn crop yield forecast is released, which informs the $100B American corn industry at every level of the chain, from grain elevator operators, to ethanol producers, to commodities traders. However, in-depth, multi-hour interviews conducted by NewtonX with a Descartes Labs senior engineer, as well as a former senior-level USDA employee, revealed that costly and manual processes in the corn crop yield forecasting process are costing the industry millions every year.

A new generation of computer vision, predictive analytics, and deep learning is beginning to revolutionize this process, though, promising both cost savings and more accurate yield predictions. This change in the industry will have far-reaching ramifications, both for the economy, businesses, and consumers, who spend an average of over $200 per year per person on corn products.

The insights from this article are sourced from NewtonX surveys, panels, and expert consultations. To gain access to these services visit

Why on-the-ground predictions are being made obsolete

The USDA uses a feet on the ground model where agents survey farms physically to painstakingly create a forecast for the year, and collect survey data from individual farms. Traditionally, forecasts have focused on state and national data, but recently county-level data has gained importance. The production estimates that come out inform marketing and investment decisions, and are also crucial to policymakers who design farm support programs, as well producers who benefit from these farm support programs.

The predictions include forecasts of area, yield, production, value, farm labor usage, land rental rates and values, and stocks for major crops. In 2016, this cost $126M (funded by congress).

Despite the funds poured into the process, however, two issues have arisen with it: the first, is that a trend has developed since the 1990’s of declining survey response rates from growers. The quality of the USDA data depends on a high level of participation, and as numbers have declined, the value of the yearly estimates has also declined. The second problem is that because survey response rates have declined, county-level estimates have led to large discrepancies in farm program payment rates. County-level estimates are of vital importance in agriculture; soil is not the same in different locations, particularly not at the state level, so having country inspections and data is paramount to making accurate predictions.

This has all led to a largely manual, on-the-ground, workflow for corn crop estimates — a costly, and imprecise methodology.

Nanosatellites, computer vision, and deep learning: the future of forecasting

Recently, nanosatellites (satellites roughly the size of a shoebox) have become more affordable and commonplace, opening up new possibilities for startups to use new technology for crop yield prediction. Before nanosatellites, if a company wanted satellite imagery data, it had to turn to government-run programs such as MODIS, which image the globe every seven days or so. But now, with satellite startups becoming increasingly popular (we actually wrote a post on the topic here), getting access to higher resolution, more granular images on a daily basis is affordable and feasible.

Descartes Lab analyzes satellite data from sources like these (a startup called Planet) to measure chlorophyll density, which allows them to get a daily proxy for crop health, on a level that’s granular to the county. Descartes analyzes satellite data of every farm in the country in order to update its corn yield predictions every two days, as opposed to the USDA monthly forecasts. Without using any human or manual labor, the company’s predictive algorithms are reportedly so accurate that they have only a 2.5% average margin of error. Using deep learning, the algorithm can integrate computer vision data and compute a new predictive outcome based on updated crop data.

The combination of regularly updated satellite data and deep learning used in computer vision and predictive algorithms has opened the door to a new era of prediction for crop yield. Where for some, like Descartes Lab, it’s created space for disruptive startups to own an area once dominated by public entities, the biggest ramifications of this new technology will be in production, hedge funds, and crop rates.

The ability to accurately forecast yield for a multi-billion dollar industry through computer vision and deep learning is already having a significant impact on numerous and varied industries at every level of the corn and agriculture chain.


About Author

Germain Chastel is the CEO and Founder of NewtonX.

Comments are closed.