BeeMachine v1.0 is an image classifier that can ID images of 36 of the most common North American bumble bee species. BeeMachine is built on a convolutional neural network modified from the Inception v3 model, utilizing a data set of ~90,000 (80/20 train/test) bumble bee images. Images were gathered from BumbleBeeWatch, iNaturalist, BugGuide, and my personal images. Bombus classification follows Williams et al. 2016 (ISBN: 978-0691152226). Overall accuracy on the test data is > 91% (> 97% top-3) but this varies by species depending on the number of training images and their degree of morphological variability. For example, Bombus fraternus has a consistent appearance across its range and among castes so it can be reliably classified even though it is trained on few images. B. sandersoni, on the other hand is more variable, so more training images should help improve its classification accuracy. More model performance metrics here.
Visit BeeMachine.org. For the most accurate results, crop your image as close to the bee as possible to reduce extraneous pixels. Then upload your cropped .jpg or .png images < 1MB.
The model outputs the top 3 most likely species and associated probabilities. Based on assessment of the test data, the correct ID is given by the top prediction over 91% of the time and is in the top 3 predictions over 97% of the time.
BeeMachine is in the early stages of development. I will be frequently updating the model for greater accuracy, including more bee species – not just bumble bees – and broadening the geographic coverage. The user interface will be updated to automatically locate the bee within your image so you won’t have to manually crop it before uploading. I will also add a “not-a-bee” class; the model currently forces any image (e.g., a honey bee, a shoe, or random noise) into one of the 36 bumble bee classes. A mobile app and a Git repository are incoming. Contact me (email@example.com) or leave a note below if you have suggestions, comments, or would like to contribute images (especially of poorly sampled species!).
Collaborators & Acknowledgements
Claudio Gratton, University of Wisconsin – Madison
Richard Wang, University of Kansas
This project was developed with data supplied by the Xerces Society for Invertebrate Conservation and Bumble Bee Watch (credit), iNaturalist (credit), BugGuide (credit), and my personal image collection. I am grateful for the volunteer participants in these programs that gather the images and share their taxonomic expertise.