Oil yield prediction for sunflower hybrid selection using different machine learning algorithms

Abstract

Due to the increased demand for sunflower production, its breeding assignment is the intensification of the development of highly productive oil seed hybrids to satisfy the edible oil industry. Sunflower Oil Yield Prediction (SOYP) can help breeders to identify desirable new hybrids with high oil yield and their characteristics using machine learning (ML) algorithms. In this study, we developed ML models to predict oil yield using two sets of features. Moreover, we evaluated the most relevant features for accurate SOYP. ML algorithms that were used and compared were Artificial Neural Network (ANN), Support Vector Regression, K-Nearest Neighbour, and Random Forest Regressor (RFR). The dataset consisted of samples for 1250 hybrids of which 70% were randomly selected and were used to train the model and 30% were used to test the model and assess its performance. Employing MAE, MSE, RMSE and R2 evaluation metrics, RFR consistently outperformed in all datasets, achieving a peak of 0.92 for R2 in 2019. In contrast, ANN recorded the lowest MAE, reaching 65 in 2018 The paper revealed that in addition to seed yield, the following characteristics of hybrids were important for SOYP: resistance to broomrape (Or) and downy mildew (Pl) and maturity. It was also disclosed that the locality feature could be used for the estimation of sunflower oil yield but it is highly dependable on weather conditions that affect the oil content and seed yield. Up to our knowledge, this is the first study in which ML was used for sunflower oil yield prediction. The obtained results indicate that ML has great potential for application in oil yield prediction, but also selection of parental lines for hybrid production, RFR algorithm was found to be the most effective and along with locality feature is going to be further evaluated as an alternative method for genotypic selection.