Automatic Identification of Diatom Morphology using Deep Learning

This paper proposes a method to automatically identify diatom frustules using nine morphological categories. A total of 7092 images from NIWA and ADIAC with related taxa data were used to create training and test sets. Different augmentations and image processing methods were used on the training set to see if this would increase accuracy. Several CNNs were trained over a total of 50 epochs and the highest accuracy model was saved based on the validation set. Resnet-50 produced the highest accuracy of 94%, which is not as accurate as a similar study that achieved 99%, although this was for a slightly different classification problem.