Epsilon Archive for Student Projects

Abstract

Accurate estimation of winter wheat (Triticum aestivum L.) yield and grain quality can support crop monitoring and agricultural management. This study evaluated the potential of Sentinel-2 imagery and meteorological data to estimate winter wheat grain yield, crude protein concentration and starch concentration at an experimental testbed site in Ultuna, Uppsala, Sweden. Data from four harvest years (2020, 2022, 2023 and 2025) were analysed, comprising 54 observations for grain yield and 48 observations for crude protein and starch concentrations. The experimental plots represented contrasting fertiliser-application treatments, although fertiliser products, nutrient composition and total nutrient inputs were not consistently documented or evaluated across years. Complementary analyses then investigated the effects of fertiliser treatment and year on grain yield and crude protein concentration, as well as associations between selected soil properties using mixed-effects models.
Sentinel-2 spectral bands and vegetation indices were combined with meteorological variables aggregated according to crop development stages. Two predictor-selection approaches were compared: an initial screening based on statistical and agronomic criteria and recursive feature elimination (RFE). Multiple linear regression (MLR), random forest (RF) and support vector machine regression (SVM) were evaluated using leave-one-out cross-validation. Additionally, unmanned aerial vehicle (UAV) imagery collected in 2025 was compared at the plot level with Sentinel-2 imagery using the vegetation index Visible Atmospherically Resistant Index (VARI) to investigate the spatial patterns captured by the two platforms.
Predictors from the vegetative growth stage were particularly important for all three response variables. RFE improved the best predictive performance, although the magnitude of improvement differed among traits. For grain yield, SVM achieved the best performance after RFE (RMSE = 957 kg ha⁻¹; R² = 0.87). For crude protein concentration, RF performed best (RMSE = 0.45 percentage points (pp); R² = 0.84), while SVM showed the best performance for starch concentration (RMSE = 0.67 pp; R² = 0.77). The final SVM grain-yield model was externally validated using 18 independent plot-year observations from the Swedish Infrastructure for Ecosystem Science (SITES) Agroecological Field Experiment (SAFE) at Lönnstorp Research Station in Skåne, Sweden. External performance was substantially lower (RMSE = 1,199 kg ha⁻¹; R² = 0.21), indicating limited transferability beyond the conditions represented at Ultuna. Comparable external grain-quality data were not available.
Mixed-effects analyses showed significant effects of fertiliser treatment and year on grain yield, whereas the Treatment × Year interaction was not significant. In contrast, crude protein concentration showed a significant Treatment × Year interaction, indicating that the response to fertiliser treatment varied among years. UAV imagery collected in 2025 revealed spatial differences in crop development among experimental blocks, highlighting the potential of high-resolution imagery for characterising fine-scale variability within the experimental site.
Overall, Sentinel-2 imagery combined with meteorological data showed potential for estimating winter wheat yield and grain quality within the experimental conditions represented at Ultuna. However, the external validation demonstrated that strong internal predictive performance did not fully transfer to an independent site, highlighting the need for broader calibration and validation across multiple sites and growing seasons.
Keywords: winter wheat, grain yield, grain quality, Sentinel-2, remote sensing, machine learning, weather