BeeMachine was developed by Brian Spiesman in collaboration with:
Claudio Gratton, University of Wisconsin – Madison
Rich Hatfield, The Xerces Society for Invertebrate Conservation
William Hsu, Kansas State University
Sarena Jepsen, The Xerces Society for Invertebrate Conservation
Brian McCornack, Kansas State University
Krushi Patel, University of Kansas
Richard Wang, University of Kansas
BeeMachine was funded by USDA NIFA and developed with data primarily from the Xerces Society and Bumble Bee Watch, iNaturalist, and BugGuide. I am grateful for the volunteer participants in these programs that share their images and taxonomic expertise.
The current version of BeeMachine identifies 36 of the most common North American bumble bee species. Overall test accuracy is more than 91% (> 97% top-3) but this varies by species depending on the number of training images and their level of morphological variability (see fig below). BeeMachine uses a convolutional neural network, modified from InceptionV3, and was trained on over 90,000 bumble bee images. Bombus taxonomy follows Williams et al. 2016.
36 Bombus species currently classified by BeeMachine:
Bombus affinis, B. appositus, B. auricomus, B. bifarius, B. bimaculatus, B. borealis, B. caliginosus, B. centralis, B. citrinus, B. crotchii, B. cryptarum, B. fervidus, B. flavidus, B. flavifrons, B. fraternus, B. frigidus, B. griseocollis, B. huntii, B. impatiens, B. insularis, B. melanopygus, B. mixtus, B. morrisoni, B. nevadensis, B. occidentalis, B. pensylvanicus, B. perplexus, B. rufocinctus, B. sandersoni, B. sitkensis, B. sylvicola, B. ternarius, B. terricola, B. vagans, B. vandykei, and B. vosnesenskii
Neural networks learn their own set of features in images to differentiate species. Each point in this visualization represents a test image. Images with similar learned features are close together, forming clusters that correspond well with species. Thus learned features parallel real life qualities that we use to differentiate species.
Error rates are generally lower for species that were trained on more images. We can use this information to target particular species for acquiring more images to improve the classification model.