Prediction of oil yield in sunflower using deep learning regression algorithm under normal and drought stress conditions.

Abstract

Accurate prediction of sunflower oil yield is essential for crop management and breeding decisions, particularly under water-limited environments. In this study, 100 pure oilseed sunflower lines were evaluated under normal and drought stress conditions across two consecutive growing seasons (2014–2016), and all morphological and physiological variables were measured directly in the field. Input variables were selected based on their previously demonstrated influence on grain yield and oil-related traits, while the complete two-year dataset was used for model training, validation, and testing following data augmentation, standardization, and 7-fold cross-validation. The predictive performance of multiple linear regression (MLR) and deep learning regression (DLR) models was compared using different combinations of input variables. The DLR model consistently demonstrated superior performance, achieving higher accuracy and lower prediction errors across all scenarios. The best results were obtained under drought stress with 11 input variables, where the DLR model achieved (R2 = 0.98, RMSE = 0.4, MSE = 0.16, MAE = 0.20 (train), R2 = 0.96, RMSE = 0.55, MSE = 0.31, MAE = 0.34 (test)), along with markedly lower RMSE and MAE values than MLR. Even when fewer variables were used, the DLR model maintained strong predictive ability, highlighting its capacity to learn complex nonlinear relationships and generalize from limited data. Overall, the findings underscore the potential of DLR as a robust predictive tool for estimating oil yield under contrasting environmental conditions and provide practical implications for genotype selection, harvest planning, and sunflower breeding strategies.

Plant phenotyping relevance

深層学習回帰によるヒマワリ油収量という植物形質の推定手法を開発し、複数モデルの性能比較と交差検証を行っており、予測・検証が研究の中心である。

Code and data availability

The article describes field-measured morpho-physiological traits of 100 sunflower lines and deep learning regression modeling, but contains no public data deposit, repository, or author code URL. Supplementary materials are only DOCX files (e.g., Table S1 line list) without explicit dataset or code content. No paper-­‐

No evidence-backed public reproduction asset is currently recorded.