Quantifying plankton is important but using humans to count individual plankters time consuming and expensive. We present results on using machine learning to classify images of preserved plankton samples. Cheaper identification through automation enables more numerous and higher frequency observations in the future as well as unlocking information trapped in collected samples from the past. The data used in this study consist of 725516 individual images taken from 46 different transects within California Current Ecosystem (CCE) from July 2005 to July 2012. Our investigation focuses on the efficacy of simple geometric features (image attributes) for plankton identification. We also quantify the volume of data required to recognize classes of plankton, which algorithms perform best, and how size fractionation can be used to accomplish machine learning more efficiently.