Abstract

We build a new dataset from the results of a pilot experiment on crop yield prediction, resulting from a cooperation between agricultural statisticians and the French national center for space studies (CNES) [1] [2]. We trained an XGBoost model to predict yield from a sample survey, on features from Sentinel 2 and meteorological data. We then use the model to build a dataset of predicted yield for all France at the plot level. Crops with an out-of-sample R² at plot level above 0.8 are selected, including wheat, maize and rapeseed : the model is best on the most common crops.
The features are monthly averages, minimum and maximum of radiometric indices (ndvi, ndwi, etc.) computed from calibrated and corrected Sentinel 2 imagery. Additional features include geographic data (coordinates, slope, orientation, altitude) and meteorological data (monthly indicators such as growing degree days, cumulative rain, etc.). The model is trained on five years (2017-2021) of the individual data from a yearly statistical survey on agricultural yield for major crops, “Terres labourables” (“arable lands”), with a representative sample of about 17000 farmers each year, with high geographical coverage. Yields are known only at the farm scale, not the field scale, and survey respondents exhibit a rounding behaviour in their responses. However, we can leverage a bigger training sample than related literature [3] thanks to the direct access to yield survey individual data. As a result, the limits of the training data might impose a ceiling on measured model accuracy not far above our current, good, metrics. Further progress might need more detailed training data.
We report various performance metrics on same-year test sets and following-year yield. As an economic forecast tool, the model has good results at the plot level but the aggregated results are not as good as our (classic) forecasting surveys, probably because the training data has too little historical depth for the model to efficiently learn year-to-year fluctuation. In same-year extrapolation however, the model is stronger at plot level prediction and very good on aggregate values. We choose to focus on same-year extrapolation.
The new dataset has three main applications:
1- Easier data dissemination. The original survey is protected for statistical confidentiality and data access is very restricted. Data are publicly disseminated only at an aggregate level (crop type and NUTS3 averages). Plot level confidential data is difficult to anonymize because the geometry and coordinates of a given plot can lead to re-identification. Our dataset contains approximated values for the yields of each survey respondent, inferred from a model not trained on this respondent’s answers. This allows for the dissemination of detailed, plot-level but noisy yield without the restriction imposed on the original, confidential survey data. Targeted users are researchers in agronomics and agricultural statistics.
2- Extrapolation for all France at plot level. We apply the model to all plots for selected crops in France, expanding the size of the dataset by an order of magnitude over the original sample survey dataset. This will allow for new use cases leveraging the exhaustivity of the data. In particular, it is difficult to match individual data from different sample surveys when the samples are not coordinated, because the sample rates compound and the matched set is very small. With an exhaustive extrapolation, it becomes possible to match predicted yield to other data from any source, including for small samples, low frequency phenomena or local analysis.
3- Auxiliary data for sampling and estimation. The model is trained on a sample survey. In return, the results can feed the sampling strategy and improve the survey’s accuracy. The comparison between the aggregated predicted yield in sampled plots and in the overall universe of plots informs on the quality of the sampling. The predicted yield can be used for post-stratification, ensuring the weighted sample is representative of the overall predicted yield, or directly for stratified sampling. Finally, sampling can target the units with the higher prediction error to focus on the information most complementary to the satellite model.
Finally, we discuss our goal to put this work into production: retraining models on new years of satellite and survey data, producing new yearly dataset of predicted yield at plot level, and improving the statistical survey accuracy.
[1] INGLADA, Jordi. Exploitation de l'enquête TERLAB pour l'estimation du rendement des cultures à la parcelle à partir de séries temporelles Sentinel-2. Phd thesis, CESBIO, 2020.
[2] BABET Damien, BALLET Bertrand, QUEYRUT Olivier, DUTHOIT Thomas, CHAPUIS Thierry, NARRAU Antoine, Prediction of crop yields at field scale from earth observation data, NTTS 2025 poster, 2025
[3] ZHAO, Yan, POTGIETER, Andries B., ZHANG, Miao, et al. Predicting wheat yield at the field scale by combining high-resolution Sentinel-2 satellite imagery and crop modelling. Remote Sensing, 2020, vol. 12, no 6, p. 1024.