Upward revisions from USDA on corn yield are in line with TellusLabs' outlook
Our coverage of the US growing season
At TellusLabs, our mission is to build live systems for environmental intelligence at a planetary scale. We seek fast-moving problems where a spaceborne perspective is decisive and stakes are high for markets and society. Behind the scenes, we rely on a powerful, multi-source framework for remote sensing analysis and a human-led approach to machine learning.
TellusLabs yield estimates for the 2017 season: The Kernel web app allows customers to access county-scale visualizations of yield. Source: Kernel; TellusLabs.
For the past two years, we have monitored the US corn and soy growing seasons because we think they pose a difficult but crucial challenge. In 2016, we shared yield estimates daily with 600 beta participants. During that record-breaking harvest, our models showed remarkable earliness and accuracy, particularly for soy.
This year, a cohort of commercial users accessed our US corn and soy insights through Kernel, an immersive web app and API. The application covers the season every day, including yield estimates, weather, plant health, public data, and commentary. This week, we are sharing a recap of our 2017 results publicly, and we are opening up a trial of the product.
Corn in 2017
This year, we produced three models for US corn yield. The TL 2017 model (green below) started the season at 166.2 bu/ac (June 28th coverage). From July, we also produced TL 2016 (blue below) – a model with same structure as the corn model we ran in 2016 (that predicted corn with 1% error), but trained on more recent data. Just prior to finishing model operations (October 3rd) we provided TL 2017 Baseline (purple below), which explored the impact of an alternative approach to long-term trend adjustment (in response to customer requests).
TellusLabs corn yield estimates for 2017 season. Brighter colors for each model indicate first date of introduction into Kernel; dotted orange line is USDA. Source: Kernel; TellusLabs.
As is clear from the figure above, all three models underwent a steep increase of roughly 4 bu/ac over the course of August. Our preferred model, TL 2017 (green) lagged our legacy TL 2016 model (blue), but both arrived above 175 bu/ac by early September. They ended the season at an average of 178.4 bu/ac.
The figures above tell a consistent story: a very strong, potentially historic year for corn. Until Thursday (November 9th), neither the USDA nor the general market shared that view. Trade estimates (via Reuters) were in the 166-170 bu/ac range for all of August-October and the USDA’s own estimates (orange) ranged from 169.5 to 171.8 bu/ac over the same period.
That USDA outlook changed sharply last week with the November 9th Production Report, which featured a revised forecast of 175.4 bu/ac. The market conversation certainly reflects the surprise at this revision. We are pleased that our preferred corn models anticipated the historic strength of this harvest well ahead of both the government and market consensus (since late-August and early-September).
Soy in 2017
TellusLabs produced three models for US soy yield this year, in a similar framework to the corn models described above (in figure below, TL 2017 is purple, TL 2016 is blue, and TL 2017 Baseline is green). We opened the season with a soy outlook of 49.5 bu/ac (June 28th coverage).
TellusLabs soy yield estimates for 2017 season. Brighter colors for each model indicate first date of introduction into Kernel; dotted orange line is USDA. Source: Kernel; TellusLabs.
We’ve seen a much wider range for soy this year than for corn (shaded area above). As was the case for corn, the TL 2017 Baseline variant was the low water mark for our estimates. For a good portion of the season, that TL 2017 Baseline model was the company outlook (meaning that we believed it had the strongest predictive potential of the three). By October, we concluded that conditions did not support such a conservative view on soy, partly thanks to the close agreement emerging between the TL 2016 and TL 2017 models (as with corn). Our final guidance to customers in October was for expectations on the upper end of the range we see above. We are not anticipating significant further movement for US soy this season.
What’s next for us?
The Kernel mission is to provide agricultural intelligence globally. As such, we’ll be building on the success of our Argentina soy beta last year and launching corn and soy forecast coverage for both Argentina and Brazil this season (December - June). By end of 2017, we will also be providing full imagery, weather and plant health metrics for the entire Southern Hemisphere (see below). By Spring 2018 that coverage will span the whole globe and include a new crop (wheat).
TellusLabs Kernel platform. As we prepare for South America coverage, the team at TellusLabs is lighting up weather forecasts for the entire planet. Source: NOAA-GFS; TellusLabs.
We learned a great deal this season on every front: product, engineering, data science and commercial. All of those lessons are helping us deliver better insights to a growing Kernel community.
We urge you to join us as Kernel becomes more global, more immersive and more inclusive. Please sign up here for trial access – if you’ve read this far you deserve it!
About TellusLabsTellusLabs is a Boston-based satellite imagery and machine learning company whose mission is to build live environmental intelligence systems at a planetary scale. In 2016 & 2017, our agricultural intelligence product, Kernel, anticipated US soy and corn yields ahead of the USDA. We've been featured in MIT Tech Review and TechCrunch and have been recognized as a MassChallenge Diamond Award Winner, FIA Innovator of the Year and Singularity University Global Grand Challenge Winner.