Samples were collected and preserved during the austral summer-autumn transition in 2007-2008 and in addition to routine McMurdo Dry Valley Long Term Ecological Research (LTER) core limnological sampling in November and December of 2008, 2010, and 2011.

Samples preserved in 1% Lugol’s solution and at 5 °C were homogenized on a shaker table (150 RPM for 5 mins) before subsampling 150 mL and settling for 24 hrs. The supernatant was then aspirated off to concentrate the sample by 5-fold. 1-5 mL of homogenized subsample was imaged on FlowCam model VS-IV with a 10X lens and a 100 µl flow cell. Images were captured using AutoImage Mode at 18-20 frames sec^-1 at a flow rate ranging 0.15-0.20 mL min^-1. The fluid volume that was imaged ranged from 5-20%. Files containing the imaged plankton from each sample were then used to sort and build statistical filters based on particle morphologies.

The morphotypes of the imaged plankton were sorted and classified using the Visual Spreadsheet Software (version 4.17.14). From a select number of samples that covered a range of times and depths throughout the record, images of like-morphotypes, such as a specific taxon or class of phytoplankton, were manually grouped into taxon- and morph-based libraries. For a total of 27 libraries that covered the range of morphotypes found in the plankton, we built statistical filters based on the specific geometries (e.g., length, width, aspect ratio) of the given particles in each library. The statistical filters were then built into a classification template of seven classes based on broad morphologic characterizations of the plankton. This classification template, based on statistical filters generated by particle attributes, was then applied to all the samples of interest to automatically sort and classify the image-based data.

We chose to assign taxonomic identifiers to only the classes in the classification template with high enough image resolutions for sufficient identification. This condition was met only for the haptorian ciliates of the genera Askenasia and Monodinium because these taxa are known to occur in these lakes and likely not to be confused with other protists. Whereas we assigned generic morphologic-based names to morphologically similar groups with overlapping attributes (e.g. round coccoidal or filamentous) that did not classify well at the FlowCam resolution used (i.e. 10x magnification). Moreover, taxonomic delineation between small bacteria and chlorophytes was unachievable with this method.  To overcome this challenge, we classified small coccoidal algae consisting of mostly small chlorophytes, cryptophytes, chrysophytes and some small ciliates as little green round things (LGRT) of small (<5 µm), medium (5-15 µm) and large (>15 µm) size classes based on their cell diameters. Libraries of filamentous cyanobacteria (mostly comprised of Oscillatoria sp. and Phormidium sp.) were classified as either short (<25 µm) medium (25-75 µm) or long (>75 µm) filaments based on lengths. A separate class with filters based on detritus, sediment, and air-bubbles was used to control for unwanted image data. Using our classification template, we were able to automatically compute plankton density (cells mL-1) and biovolume for each of the seven classes.