Yield prediction at the field level is crucial for optimizing agricultural productivity and ensuring food security. This study analyzes the yield variability of maize, sunflower, and winter wheat across 481 agricultural fields in two regions of Ukraine (Kyiv and Cherkasy) over a three-year period (2020–2022). The objective was to explore the influence of environmental factors on crop yield predictions using satellite and weather data, sowing dates, and field area as predictors in a machine learning model. The study employed Random Forest model. Satellite data from Sentinel-2, including NDVI and LAI values, were used to assess crop conditions during the growing season. For each investigated year during the April-September period, focusing solely on the NDVI and LAI values for each month. Weather data, especially precipitation, was also examined but found to have limited predictive power due to the coarser spatial resolution of the gridded data (6.5 km), which cannot fully account for the local variations within each grid cell. As a result, despite the strong correlation between precipitation and yield at a broader scale (regional), weather data alone were not sufficient to accurately predict yield variability at the field level. The results showed that maize had the highest yield variability, while sunflower and winter wheat exhibited more stable yields. For maize, the model demonstrated strong predictive performance, with an R-squared of 0.8 and an RMSE of 1.5 t/ha. The most significant predictors were vegetation indices in August and sowing date. The normalized RMSE for maize was 20%. For sunflower, the model exhibited moderate accuracy, with an R-squared of 0.4 and an RMSE of 0.9 t/ha. Key predictors included the average LAI in May and July. However, the model’s predictive power was limited, resulting in a normalized RMSE of 23%. Winter wheat showed similar performance to sunflower, with an R-squared of 0.35 and an RMSE of 0.9 t/ha. Due to higher average yields, the normalized RMSE for winter wheat was 15%. Overall, the study demonstrates varying levels of model accuracy across different crops, with maize achieving the best predictive performance. The results also emphasize the need for additional factors, such as soil properties, microclimates, and detailed field management practices, to improve predictive models at the field level.
Funding: This research received no external funding.
Data Availability Statement: Not applicable.
Acknowledgments: The authors would like to express their sincere gratitude to the Earth Observing System Data Analytics company (eosda.com) for support. We are also grateful to reviewers and editors for their valuable comments, recommendations, and attention to the work.
