Fig. 3

Confusion matrices of a resnet-50 transfer learning classifier. Note that this model is simultaneously classifying species and specimen fragmentation state. Panels a and b show species classification confusion matrices while panels c and d show genera classification confusion matrices, where a and c are unnormalized and b and d are normalized. Normalization scales values across each ground-truth label (i.e., row) such that they sum to 1; thus, the color saturation represents the fraction of that true label that was classified for each predicted label (where greater saturation indicates more images in the category). Confusion matrices for species classification shows that only extremely rare species are heavily misclassified (typically as a non-foram object). Panels e and f show unnormalized and normalized confusion matrices for fragmentation state, respectively. In this use case, the classifier tends to misclassify the fragmentation state for fragmented shells, but not for complete shells. This standard, unoptimized transfer learning classification approach has validation accuracies of 80.6% (species), 85.6% (genus), and 76.1% (fragmentation state).