Classification and Quantification of Strawberry Fruit Shape

"Classification and Quantification of Strawberry Fruit Shape" is a dataset that includes raw RGB images and binary images of strawberry fruit. These folders contain JPEG images taken from the same experimental units on 2 different harvest dates. Images in each folder are labeled according to the 4 digit plot ID from the field experiment (####_) and the 10 digit individual ID (_##########).

"H1" and "H2" folders contain RGB images of multiple fruits. Each fruit was extracted and binarized to become the images in "H1_indiv" and "H2_indiv".

"H1_indiv" and "H2_indiv" folders contain images of individual fruit. Each fruit is bordered by ten white pixels. There are a total of 6,874 images between these two folders. The images were used then resized and scaled to be the images in "ReSized".

"ReSized" contains 6,874 binary images of individual berries. These images are all square images (1000x1000px) with the object represented by black pixels (0) and background represented with white pixels (1). Each image was scaled so that it would take up the maximum number of pixels in a 1000 x 1000px image and would maintain the aspect ratio.

"Fruit_image_data.csv" contains all of the morphometric features extracted from individual images including intermediate values.

All images title with the form "B##_NA" were discarded prior to any analyses. These images come from the buffer plots, not the experimental units of the study.

"PPKC_Figures.zip" contains all figures (F1-F7) and supplemental figures (S1-S7_ from the manuscript. Captions for the main figures are found in the manuscript. Captions for Supplemental figures are below.

Fig. S1 Results of PPKC against original cluster assignments. Ordered centroids from k = 2 to k = 8. On the left are the unordered assignments from k-means, and the on the right are the order assignments following PPKC. Cluster position indicated on the right [1, 8].

Fig. S2 Optimal Value of k. (A) Total within clusters sum of squares. (B) The inverse of the Adjusted R . (C) Akaike information criterion (AIC). (D) Bayesian information criterion (AIC). All metrics were calculated on a random sample of 3, 437 images (50%). 10 samples were randomly drawn. The vertical dashed line in each plot represents the optimal value of k. Reported metrics are standardized to be between [0, 1].

Fig. S3 Hierarchical clustering and distance between classes on PC1. The relationship between clusters at each value of k is represented as both a dendrogram and as bar plot. The labels on the dendrogram (i.e., V1, V2, V3,..., V10) represent the original cluster assignment from k-means. The barplot to the right of each dendrogram depicts the elements of the eigenvector associated with the largest eigenvalue form PPKC. The labels above each line represent the original cluster assignment.

Fig. S4 BLUPs for 13 selected features. For each plot, the X-axis is the index and the Y-axis is the BLUP value estimated from a linear mixed model. Grey points represent the mean feature value for each individual. Each point is the BLUP for a single genotype.

Fig. S5 Effects of Eigenfruit, Vertical Biomass, and Horizontal Biomass Analyses. (A) Effects of PC [1, 7] from the Eigenfruit analysis on the mean shape (center column). The left column is the mean shape minus 1.5× the standard deviation. Right is the mean shape plus 1.5× the standard deviation. The horizontal axis is the horizontal pixel position. The vertical axis is the vertical pixel position. (B) Effects of PC [1, 3] from the Horizontal Biomass analysis on the mean shape (center column). The left column is the mean shape minus 1.5× the standard deviation. Right is the mean shape plus 1.5× the standard deviation. The horizontal axis is the vertical position from the image (height). The vertical axis is the number of activated pixels (RowSum) at the given vertical position. (C) Effects of PC [1, 3] from the Vertical Biomass analysis on the mean shape (center column). The left column is the mean shape minus 1.5× the standard deviation. Right is the mean shape plus 1.5× the standard deviation. The horizontal axis is the horizontal position from the image (width). The vertical axis is the number of activated pixels (ColSum) at the given horizontal position.

Fig. S6 PPKC with variable sample size. Ordered centroids from k = 2 to k = 5 using different image sets for clustering. For all k = [2, 5], k-means clustering was performed using either 100, 80, 50%, or 20% of the total number of images; 6,874, 5, 500, 3, 437, and 1, 374 respectively. Cluster position indicated on the right [1, 5].

Fig. S7 Comparison of scale and continuous features. (A.) PPKC 4-unit ordinal scale. (B.) Distributions of the selected features with each level of k = 4 from the PPKC 4-unit ordinal scale. The light gray line is cluster 1, the medium gray line is cluster 2, the dark gray line is cluster 3, and the black line is cluster 4.