Bombus affinis, commonly known as the rusty-patched bumble bee, is a critically endangered species endemic to North America. Accurate identification of this bee is crucial for monitoring its presence and abundance, and in turn conservation efforts. Here, we design computer vision techniques to identify the rusty-patched bumble bee from images uploaded by citizen scientists on the iNaturalist platform. Our original dataset consists of 200 images of the rusty-patched bumble bee, and 200 images of other bees across six genera. After image augmentation (mirror, rotate) to yield a final dataset of 3, 200 images, we fine-tuned an EfficientNetV2B0 AI model to classify the rusty-patched bumble bee, and were able to achieve accuracies of 90% and 92% for color and grayscale images. However, motivated by a unique V-shaped black band that is present on the thorax of the rusty-patched bumble bee, we implemented an anatomically inspired learning framework wherein we automatically extract and classify only the thorax pixels from an image. Our resulting thorax-only model yielded classification accuracies of 94−95% and sensitivity (recall) values of 95−99%, across color and grayscale versions of manually and AI-cropped thorax images (the improved results presumably coming from removing other sources of noise). Deployment of these techniques could be useful in the automatic and real-time identification of rusty-patched bumble bees, hence aiding their identification and conservation. Our models can be easily extended to other insects, and to more anatomically inspired learning methods for classifying insects in nature.