Analysis of coca leaf shape

COMPARE_CNN_VS_LDA: Using the CULTIVATED1ST dataset, this analysis compares leave-one-out cross-validation (LOOCV) approaches for linear discriminant analysis (LDA) and convolutional neural networks (CNNs) to predict cultigen and type from leaf shape. For LDA, only Procrustean coordinates are used, and for CNNs both a shape mask of the leaf and Euler characteristic transform (ECT) are used. Synthetic leaves are generated using the synthetic minority oversampling technique (SMOTE). For each fold (which is the number of samples), the SMOTE synthetic data generation creates equal class numbers, matching the most abundant class. Using the true vs predicted identities for all samples/folds/models, permutation 5000 times of the predicted column against the true is used to calcualte a p-value for each class and overall. The CNN approach worked better for type models and the LDA worked better for variety models. Given the compute time for the CNN approach in the LOOCV context, LDA was chosen going forward.COCA_LDA: This analysis tests six different configurations to predict various classes from Procrustean pseudo-landmarks. Three different class types are predicted: dataset, type, and variety. LOOCV sampling is performed in two ways: by sample and by plant. A total of six model configurations of each combination of the above were performed. When using LOOCV sampling by plant, when a sample is selected, all other leaves from the same plant are also removed (except for the PLOWMAN dataset). A csv with model performance results and calculated p values by permutation, confusion matrix values, and confusion matrix plot are saved for each model configuration. Overall summary figures showing confusion matrices by count and normalized are saved as well.WILDSPECIES: The same as above, but using only LOOCV by sample (not plant) predicting type for cultivated leaves and species for wild species data. All leaf data has no petiole for this analysis, to be consistent with the wild species data.