一种基于SC-StackingPSO-MI的油菜籽粒表观特征筛选方法OA

[Objectives]To extract effectively the key phenologic traits of rapeseeds and serve the rapid screening of excellent rapeseed germplasm resources,based on spectral clustering(SC),stacking particle swarm optimization(StackingPSO)and mutual information(MI),a method of rapeseeds key phenotypic trait character screening SC-StackingPSO-MI was proposed.[Methods]After the determination of the optimal spectral clustering of rapeseed grains,this paper designed the Stacking ensemble learning model based on the logistic regression(LR)algorithm combining with k-nearest neighbor(KNN),random forest(RF),and support vector machine(SVM)to evaluate the fitness function in the particle swarm optimization(PSO)for the screening of key apparent trait features of rapeseed grains by combining the MI between the features and their clustering results.Furthermore,clustering analysis of rapeseed grains,the comparison of the classification accuracies of rapeseed grains under different classifiers and different feature selection algorithms based on the selected key features were carried out.[Results]Under the binary clustering condition,the SC-StackingPSO-MI feature selection algorithm could effectively extract the feature indicators with high statistical variance from the apparent features of rapeseeds.The Calinski-Harabasz(CH)indices with the optimized features increased from 42.89 to 83.90,which resulted in a significant improvement in the clustering performance.The overall classification test performance of rapeseeds screening features from SC-StackingPSO-MI under different classifiers was compared based on binary clustering labels.Compared with the full feature combination,three classification models such as KNN,SVM and LR all showed varying degrees of improvement in testing accuracy apart from RF model.That proved the feature selection results had good robustness.Compared to the feature selection methods of LASSO,SVM-RFE and XGBoost under binary clustering labels,the SVM model's test accuracy(ACC)and Kappa values were 97.22%and 93.88%,respectively,indicating the higher reliability.[Conclusions]The feature screening method proposed in this paper had demonstrated good superiority in the selection of key apparent trait features of rapeseed grains,and could provide relevant technical references for the selection of key apparent trait characteristics in the screening of excellent germplasm resources of rapeseed grains and other crops.