Seed Architectural Phenes Prediction and Variety Classification of Dry Beans (Phaseolus vulgaris) Using Machine Learning Algorithms

Abstract

Proper identification and categorization of seeds at an earlier stage of the cultivation process is an imperative procedure that contributes to better crop quality and higher production yield. As a strategy to supplement this procedure, integration of computer vision approach and machine learning algorithms including gaussian process regression (GPR), decision trees for regression (RT) and classification (CT), support vector machine regression (SVMR), k-nearest neighbors (KNN), linear discriminant analysis (LDA) classifier, and Naïve Bayes (NB) classifier are explored in this study to predict the extended morphological features (solidity, roundness, compactness) and variety classification of dry bean (Phaseolus vulgaris L.). A total of 13,611 image samples were used. CIELab color channel thresholding was applied in segmenting bean pixels and region properties for extracting the morphological features (bean biomass area, perimeter, major and minor axis lengths, convex area, eccentricity, extent, equivalent diameter, and axis length proportionality, shape factors, roundness, solidity, compactness). Based on RMSE and MAE performances, the optimized GPR is the most reliable model for predicting seed solidity, and regression tree for both seed roundness and compactness. Classification models with seven morphological predictors (LDA7, KNN7, CT7, NB <inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">7</inf> ) exhibited sensitive classification performance, all having accuracies greater than 90%. Further, KNN7 bested out other models with 93.69% accuracy, 93.64% precision, 93.66% specificity, and 93.69% f1-score. The developed machine learning models are innovative approaches in the seed variety classification and phenotyping of dry bean seeds.