Crop Yield Prediction

EOSDA team of data scientists and engineers has developed effective techniques for crop yield estimation using remote sensing and machine learning models. We’re relying on earth observation data retrieved from satellites to cover areas ranging from individual farms to regions.

satellite collecting yield data
Numbers

Crop Yield Prediction in Numbers

Accuracy

up to 95%
Accuracy of yield estimated depends on the quality of statistical data and can vary from 85% to 95%.

Forecasts

Up to 3 months ahead
Current season yield forecasts up to 3 months in advance.

Crop types

100 +
Yield predicted for over 100 crop types.

Project speed

up to 14 days
We’ll produce a 95% accurate yield forecast in two weeks or less, depending on the complexity of the project.

Entries per crop

0 to 100 fields
WOFOST yield estimation model requires no data at all.

Data Sources

10 +
We make sure the forecasts are based on the most comprehensive data analysis.

Watch video about our yield prediction solution

Global food security depends on the efficiency of food management practices, such as yield prediction, which allows farmers to raise crops in a more sustainable way. EOSDA’s yield forecast solution – developed based on the latest technological advances in machine learning and geospatial analytics – provides farmers, agri holdings, food security companies and other decision-makers with crucial data needed for sustainable and profitable crop production.

Benefits

Yield Estimation Benefits

  • Increased speed of decision-making related to harvesting, storing, and transporting operations.
  • Data on crop profitability in your area of interest based on yield estimation.
  • Opportunity to strengthen global food security by introducing crop yield forecasting to developing countries - helping them to prevent famine, boost local economies, and implement sustainable agricultural practices.
information for yield forecast
  • Improved understanding of the agricultural market and better-informed decisions on the management of stocks, imports, and exports, in accordance with CAP and other similar policies.
  • A much better understanding of cumulative effects of hostile field conditions (pests, diseases, nutrient deficiencies, and others) on crop development.
crop yield simulations
Methodology

Our Approach

For maximum efficiency and accuracy of crop yield forecasting, we fuse two different types of yield prediction models - biophysical and statistical. This "hybrid" approach allows us to take on more complex projects.

Biophysical yield prediction model

  • Collect data (weather parameters, soil analysis, crop state, phenological data, etc.).
  • Calibrate the model and carry out the LAI assimilation to ensure accuracy of a crop yield forecast in the absence of statistical data and to increase the variability of values.
  • Simulate the biological productivity parameters (TAGP, WSO, relative soil moisture, total water consumption, and others) to estimate yield.
  • Update the data once every 14 days to increase the accuracy. This has to do with weather updates.

Statistical yield prediction model

  • Collecting data to create a crop yield prediction dataset and combining it with possible predictors (rainfall, temperature, humidity, soil type, and others).
  • Picking the right ML model for the project - e.g. Linear regression, Random Forest, LightGBM, XGBoost, CatBoost, to name a few.
  • Adjusting the model to answer the specific needs of the project in question for best results.
Model fusion stage
The fusion stage is necessary if we want to achieve the highest possible accuracy of 95%. We fuse the biophysical yield prediction model with the statistical model described above.
EOSDA Crop Modeling + LAI Assimilation
  • Ensemble of model scenarios

  • Asquisition 1

  • Asquisition 2

  • Asquisition 3Selection of most likely scenario and re-initalization of the modelled ststem state with the scenario

  • Asquisition 4Observations of LAI

  • Asquisition 5

  • Harvest

Applying LAI assimilation allowed us to achieve the 95% accuracy in 30% of the fields. For the fields marked in red, the accuracy of less than 80% was achieved, while the crop yield forecast accuracy for the green-marked fields exceeded the 80% mark.

EOSDA Crop Modeling accuracy
EOSDA Crop Modeling
EOSDA Crop Modeling + LAI assimilation accuracy
EOSDA Crop Modeling + LAI Assimilation
Use Cases

Our Success Stories

Yield forecast for a large agroholding in Ukraine

In 2020, we implemented a yield prediction project for 6 major crops: Winter Barley, Winter Rapeseed, Winter/Spring Wheat, Sunflower, Soy, and Maize.
Two different reports were generated:
  • 45 days prior to the harvest
  • 2 weeks prior to the harvest.
  • less than 80% accuracy
  • more than 80% accuracy
WOFOST Yield prediction
WOFOST Yield prediction
WOFOST (inputs/outputs) + LAI (Sentinel-2)
EOSDA Yield prediction Machine Learning Model WOFOST (inputs/outputs) + LAI (Sentinel-2)
Accuracy (Wofost) Accuracy (Wofost + Lai)
Maize 0.75 0.91
Soy 0.78 0.86
Sunflower 0.71 0.88
Winter barley 0.53 0.82
Winter wheat 0.75 0.92
By improving the model with LAI assimilation, which was developed by the team, we managed to increase the accuracy of yield estimation in 30% of the fields compared to the traditional WOFOST approach.
The table below shows the correlation between the accuracy of yield estimation, the target crop and the number of fields. For example, the predicted yield of winter barley was more than 90% accurate for 52 fields.
Number of Fields
Crop / Accuracy <70% 70-75% 75-80% 80-85% 85-90% >90%
Winter barley 27 7 5 22 23 52
Winter wheat 33 17 19 21 19 102
Winter rapeseed 26 6 20 14 27 22
Sunflower 12 11 12 14 19 22
Soy 28 22 29 58 37 86

Crop yield forecasting for a Canadian insurance company

Goal: Reliable predicted yield data on every customer to reduce insurance risks.

Input data: Over 100 fields on 20 farms.

Task 1.
Estimating average yield for 6 major crop types growing in every field on all 20 farms and comparing it against the actual yield report.
yield estimation model vs actual yield by crop type given in %
  • Canola
    > 98,03
  • Corn
    > 87, 59
  • Peas
    > 76,25
  • Soybean
    > 95,94
  • Sunflower
    > 98,21
  • Wheat
    > 98,63
Crop Modelled Yield Actual Yield
Canola, lbm/ac 41,81 41,00
Corn, BPA 123,65 110,00
Peas, q/ac 30,94 25,00
Soybean, BPA 22,89 22,00
Sunflower, lbm/ac 1767,73 1800,00
Wheat, BPA 53,72 53,00
Grand Total 95,6 94,47
Task 2.
Estimating yield 14 days prior to the 2020 harvest.
yield estimation model vs actual yield by crop type given in %
  • Canola
    > 96,96
  • Corn
    > 91,69
  • Oats
    > 99,98
  • Rye Fall
    > 85,85
  • Sunflowers Confects
    > 85,36
  • Sunflower Oils
    > 98,06
  • Wheat
    > 94,95
Crop Modelled Yield Actual Yield (Farm 4)
Canola 40,19 39,00
Corn 119,14 110,00
Oats 125,03 125,00
Rye Fall 64,39 75,00
Sunflowers Confects 2063,60 1800,00
Sunflower Oils 1834,19 1800,00
Wheat 61,73 65,00
Grand Total 584,34 528,00
Task 3.
Providing the client with the crop yield forecasting data to enable more efficient planning of crop rotation and, as a result, significantly reduce insurance risks.
The graph shows the predicted yield for target crops in 3 selected fields in bushel/ha.
Field name Canola Corn Soybean Sunflower Oils Wheat
SE-2-6-28-W1 58,68 194,33 41,45 2208,85 72,49
SW-36-7-28-W1 30,91 169,49 14,42 1146,91 46,24
W-34-5-27-W1 38,77 151,58 24,71 1476,83 59,82
The Jackknife resampling technique was used. Namely, by systematically omitting every observation from a dataset we calculated the estimate and then discovered the average of the calculations. To exclude climatic and technological factors, we only used the data for the past 6 years.

The harvest period for the target crops in Canada usually lasts from August till September. Knowing this, we were able to forecast yield two months before the harvest, achieving an accuracy of over 82%. The accuracy was steadily increasing as the harvest was approaching until it reached 90% just two weeks prior to the harvest, as had been expected.