An Automated Approach to Bee Identi cation from Wing Venation

Abstract

Colony collapse disorder is killing o the world's honeybees at alarming rates. Since honeybees are the primary pollinators of much of mankind's food supply, we must look to other sources of pollination. There are thousands of di erent species of bees, some of which may be helpful in solving this problem. Yet many of these bees go unnoticed and remain unmonitored. One obstacle that has slowed research of bee population dynamics among melittologists is the di culty of identifying bees down to the species level. Modern pattern recognition techniques using images of the forewing have been used to successfully classify bee genus, species, subspecies, and even gender. The MelittO Biotaxis System (MOBS), is described which is used to collect forewing images of living or deceased specimens. Images are preprocessed using common techniques so that features may be extracted. Using a set of training data, a feature selection algorithm selects a small set of features to use in the classi cation problem. Specimens may be misclassi ed, because they are confused with other known classes, or they belong to an unknown class. Measures of reliability to help detect such occurrences are set forth for a Maximum Likelihood classi er and for a k-nearest neighbor classi er. Reliability measures are shown to be helpful in identifying misclassi ed bees under both scenarios. A dataset of over one thousand bees representing over ve genera and over twenty species of bees was tested with MOBS. Over 90% of specimens were correctly segmented. Species were correctly identi ed in over 90% of the tests.