Predicting Crop Yield from Pre-Harvest Satellite Imagery

Ingvar Baranin
Creating automatic and accurate forecasts of crop yield using satellite imagery has been the subject of hundreds of studies across decades, starting from the first Earth observation satellite which focused specifically on land monitoring, launched by the United States Landsat program. While these forecasts can help agricultural workers allocate their resources and provide timely estimates of production, the data required to facilitate the training of analogous prediction models is not easily attainable. Thus, no such large-scale model has been made for agricultural environments pertaining to Estonia.

In this work, we utilize the field data provided by eAgronom to consider the prediction of field-scale crop yield, with focus on winter wheat fields in Estonia and Poland from 2018 to 2022, both as a pre-harvest estimate and post-harvest data validation. Through creating a data pipeline that combines Sentinel-2 satellite time series data with other field metadata, we train and compare various learning methods to ultimately report the first state-of-the-art Estonian crop yield model. The resulting pipeline and model benefits both eAgronom, by providing a multipurpose product, and farmers, by improved decision-making.
Graduation Thesis language
Graduation Thesis type
Master - Data Science
Tiit Sepp
Defence year