What question did this study set out to answer?

To evaluate the predictive potential of NDVI metrics for accurately estimating winter oilseed rape yield across multiple fields in Lithuania.

April 8, 2026Open Access

Yield Prediction in Winter Oilseed Rape Based on Multi-Temporal NDVI and Modelling Approaches

Key Points

To evaluate the predictive potential of NDVI metrics for accurately estimating winter oilseed rape yield across multiple fields in Lithuania.
Collected NDVI data from September 2023 to May 2024 across various fields.
Applied four modelling approaches: OLS regression, Random Forest, XGBoost, and Deep Neural Network.
Used random and spatially independent validation schemes to assess model performance.
Conducted feature importance analysis focusing on seasonal NDVI data.
Under random sampling, Random Forest and XGBoost models achieved RMSE of approximately 0.85 t ha−1 and R2 around 0.55.
Model performance declined under spatial validation, showing near-zero R2 values for tree-based models.
Deep Neural Network showed better generalisation with RMSE of 1.09 t ha−1 and R2 of 0.28.
Spring NDVI, particularly from March to May, was identified as the key predictor of yield, while autumn NDVI was less effective.

Abstract

Accurate prediction of winter oilseed rape yield is essential for optimising crop management and improving production efficiency. However, the reliability of commonly reported model performance remains uncertain due to the widespread use of random validation strategies. This study evaluated the predictive potential of multi-temporal Normalised Difference Vegetation Index (NDVI) metrics collected between September 2023 and May 2024 for yield estimation across multiple Lithuanian fields, while explicitly addressing spatial generalisation. The analytical dataset comprised dry yield (t ha−1), monthly NDVI, and field identifiers, and underwent quality control, including outlier removal. Four modelling approaches were compared: ordinary least squares (OLS) regression, Random Forest (RF), Extreme Gradient Boosting (XGBoost), and a Deep Neural Network (DNN). Model performance was assessed using both random (80/20) and a spatially independent field-wise (GroupSplit) validation schemes designed to assess model transferability to previously unseen fields, further extended by repeated group-based resampling to quantify variability in model generalisation. Under random sampling, RF and XGBoost achieved the highest accuracy (RMSE ≈ 0.85 t ha−1, R2 ≈ 0.55). However, under spatially independent validation, predictive performance declined markedly for all models, with tree-based ensembles showing near-zero R2 values, indicating limited transferability to unseen fields. In contrast, the DNN demonstrated more consistent generalisation (RMSE = 1.09 t ha−1, R2 = 0.28). Repeated field-wise validation confirmed that performance estimates based on random splits substantially overestimate true predictive capability. Feature importance analyses consistently identified spring NDVI, particularly from March to May, as the dominant predictor of yield, whereas autumn NDVI showed weaker and less consistent relationships with yield. These findings demonstrate that a large portion of the predictive skill reported in NDVI-based yield modelling may arise from spatial information leakage rather than transferable crop-environment relationships. By explicitly quantifying the gap between random and spatial validation, this study provides a more realistic benchmark for model performance and highlights the necessity of spatially robust evaluation frameworks for operational yield prediction in precision agriculture.

Perguntar à IA

Bookmark

View Full Paper