Statistical filtering to aid in the classification of phytoplankton: The effects of image library size and phytoplankton shape

The demand for image classification methods has increased due to
technological advancements that enable more intensive phytoplankton
monitoring. Regardless of whether the methods are based on statistical or
machine learning algorithms, algal taxa may be misidentified in
taxonomically diverse samples, in which phytoplankton morphology and image
traits can be variable. We evaluated the statistical filtering performance
for two approaches to image library development, which we applied
independently to seven commonly occurring algal shapes. To do so, we used
the statistical filter in the image processing software of an imaging flow
cytometer (FlowCAM) and previously classified samples. One statistical
filtering approach used a small selection of images (5-15 images of a
target taxon) from the same sample being filtered (i.e., intrinsic), and
the other used a larger selection of images (30-80 images of a target
taxon) compiled from different samples. Filter accuracy, precision, and
recall varied with the type of image library, image library size, and
target taxon. The largest image libraries offered high recall (>
90%) but low accuracy and precision with both image library building
approaches. For the largest image libraries, accuracy and precision were
higher for the intrinsic method (>90% and 72-97%) than the compiled
method (>40% and 10-20% for most taxa, respectively). Statistical
filtering performance was higher for larger, solitary-celled taxa with
relatively uniform features (e.g., Gyrosigma) compared to small-celled
colonial species with more complex or variable shapes (e.g., mucilaginous
colonial cyanobacteria, and Scenedesmus). Results indicate that
statistical filtering can be used to augment manual sample
classification.