Gro Yield Model Forum
Welcome to Gro’s Yield Model Forum! Please feel free to post any comments, feedback, or questions here. Note that you can opt to receive email notifications to stay updated on the threads you comment by going into your user settings.
Check out our yield model page to learn more about the model and download our paper: https://gro-intelligence.com/yield-model
26 results found
-
1 vote
Daily updates to our yield model forecasts can be accessed by paying users through our software product, Gro. We however routinely update delayed values on our website here: https://gro-intelligence.com/products/gro-forecasts
-
2 votes
Hi Eric, thanks for your question!
April is already close to harvest, so crops are either mature or at the end of grain-filling. The water demand of crops should be low if not zero, but too much rain may hurt or delay harvesting.
Additionally, no weather variables post-March showed up as significant variables to yield in the Gro Argentine soy model. -
US Yield Model - late season changes
I just listened to the webinar from yesterday, and have a question on the US corn yield model. You mentioned that the last change was in late December, but by that time the crop was already harvested. What changed at that point in the season that caused for a change to your yield estimates?
1 voteHi Mike, thanks for your question.
The major reasons for yield forecasting changes post-season:
1. Our US corn yield forecast is aggregated up from sub-national level forecasts. Yields from US corn belt counties are from the Gro yield model, yields from US non corn belt states are from USDA NASS. Last year our corn belt yield forecasts changed slightly after October. The major changes came from USDA’s forecasts out of corn belt states, although weighted only less than 20% of the national yield aggregation.
2. Usually after the growing season, our county level yield forecasts do not change, unless we added/backfilled new features that were fully tested for increasing our model’s performance, which was the case of precipitation from TRMM last year for US corn model. -
Historical Track Record vs USDA
Do you have available somewhere or could you share a table of your model forecasts around the time of historical USDA forecasts? It would be helpful to see that.
1 voteWe have charts in our yield model paper (which you can download here in the left hand column: https://gro-intelligence.com/us-corn-yield) that show the historical USDA forecasts compared to our model’s 2016 forecasts. For 2017, you can find the charts in our weekly commentaries, also available here: https://gro-intelligence.com/us-corn-yield
-
1 vote
Hi Rafae,
We will be updating our yield model forecast on our website today. However, if you prefer to receive updates faster, you can access daily, automatic updates on our data platform, Gro: https://gro-intelligence.com/features.
-
progressive yield limits
Over the past few weeks your estimate has risen from 156 to 162, then back to 159 & up to 160. Granted that growing conditions can improve and lead to a better yield but corn is determinate type crop. To a considerable extent conditions at a certain stage determine the maximum potential going forward. Ear length & kernel row number is determined in the final vegetative stages, pollination success determines the number of kernels. And while it is true that good weather during kernel fill can compensate for fewer kernels with heavier kernels there is a limit.
Does your model…2 votes -
true yld?
On your home page you show your July & August predictions against "true yield". Is the "true yield" you refer to the USDA's final yield? Or do you have some other number?
2 votesTrue yield refer to USDA final reported yield published the following year.
-
1 vote
We believe that our trend model accounts for a tendency for earlier planting, as well as other multiyear trends that are already in progress. The trend model runs at the beginning of each growing season.
-
1 vote
In our understanding, “phenological growth data” would be data that makes specific reference to growth stages of corn plants. We have not incorporated data of that specific nature in the forecast, but we believe that its predictive value is captured by other features that are in our suite of models.
-
1 vote
One of the features we tested that did not make it to the final model was crop density (plants/acre from NASS). It shows high correlation with crop yields, but not with deviations from our current yield model.
-
1 vote
The corn crop mask used in our current process are five different types of static masks (generated using yearly CDLs from USDA), that are applied across all years. They are static masks with different confidence level of approximation to the truth. For example, we have a low confidence mask that classifies the pixels as corn, which have been corn pixels for at least a year in CDLs.
However, for our Argentine soybean model we generated our own algorithm since crop mask data doesn’t exist there, and we will soon look at applying it to the US corn model as well. -
1 vote
We have a widely disparate population of users including representatives of the following sectors:
Academia
Hedge funds
Physical hedgers/trade houses
Government
Agricultural lenders
Physical suppliers/logistics companies
Insurance providersThe breadth of our universe of data is very wide and constantly growing. We can’t possibly imagine all the uses that our audience is putting it to, but three examples that have come to our attention are:
-Early warning of local crop failure to policymakers for prepositioning of emergency stocks and food shortage countermeasures
-Provision of economic sourcing alternatives during a period of volatile trade policy
-Avoiding repetitious entry of obscure and inconveniently formatted statistics for use in financial models
-
1 vote
The satellite images we use are primarily from MODIS, specifically, for LST we use the MODIS MOD11 product which has a spatial resolution of 1km and is collected daily. As for NDVI, we actually use a product from GIMMS which combines both MODIS platforms (MOD09 from the Terra satellite and MYD09 from the Aqua satellite) and this product has a spatial resolution of 250m with a temporal resolution of 8 days.
-
1 vote
Yes for major crops in United states. CDL has crop specific masks covering corn, soybeans, wheat, and so on. For regions where there is no available crop specific mask, we are able to create the yearly masks with NDVI imagery through geospatial processing.
-
1 vote
In order to predict the current year’s corn yield, we need the current year’s corn mask, which is not available from USDA until the following year. Instead of using a current year corn mask from USDA (which does not exist), we created a set of static masks from yearly CDL, namely low confidence and high confidence ones.
-
1 vote
For the most recent forecast (as of June 13 with a backtest) at the county level, 52.06% of counties are predicted within 15.8 bushel/acre. At the national level, the mean error rate is 5.21%.
-
1 vote
State and county level results will be made available to users of the Gro platform, but are not on our public website.
-
1 vote
We intend to use any data that’s available and useful. We made heavy use of ground-based data in the US model because it’s the prototype for all of the other models that we are building in areas with poor ground-based data. We believe that satellite data alone will allow us to successfully model most crops around the world now that we have the US results.
-
What is the effect of higher volatility due to climate change?
Your models will get better with time as it "learns", but climate change makes it that the past is not a good predictor of the future.
1 voteClimate change occurs slowly enough that the refitting of the model with each year’s new data captures the effect pretty well. In a case of discontinuous rapid climate change, our model would suffer from poor results due to its historical bias.
-
1 vote
Clearly, forecasts with high confidence have to wait for greenness and other data that only arrives after the crop is growing. Nevertheless, our model does have statistical value as early as the beginning of May.
- Don't see your idea?