Multivariate Machine Learning Approach for Size and Shape Prediction of Sunflower Seeds

Abstract

Sunflower constitutes an important source of protein, mineral, vitamin, fatty acid, and offer a balanced source of amino acids. Machine learning is mostly performed for the prediction of descriptive attributes in the quality evaluation of foods. In this study physical attributes of two different sunflower varieties (Metinbey and İnegöl Alası) were determined and algorithms were applied for size and shape prediction of these varieties. In addition, five different machine learning predictors were used as Multilayer Perceptron (MLP), Gaussian Processes (GP), Random Forest (RF), k-Nearest Neighbors (kNN), and Support Vector Regression (SVR). The prediction of surface area, volume, geometric mean diameter, aspect ratio, elongation, and shape index were based on the main physical attributes. İnegöl Alası variety had the greatest physical attributes. The seed length, width and thickness were obtained from İnegöl Alası variety as 23.89, 8.80 and 4.15 mm and from Metinbey as 17.88, 6.20 and 4.01 mm. All varieties were determined as significant in terms of the selected attributes as reported by Pillai Trace and Wilks’ Lambda (p<0.01). In the Wilks’ Lambda statistics, unexplained of the similarities or differences among the groups was 12.30%. Present findings revealed that MLP and SVR algorithms had the greatest correlation coefficients for all predicted attributes. In the study, the best predicted attributes were geometric mean diameter with an R value of 0.9989 (SVR), followed by volume and elongation with an R value of 0.9988 (MLP). Present findings revealed that MLP and SVR algorithms could potentially be used for size and shape prediction of sunflower varieties.