Abstract
Healthy ecosystems provide indispensable services like pollination, climate regulation, and soil formation. Pollination service provided by insects alone is valued at €153 billion. Bees are one of the major insect pollinators, making them economically important animals, yet their populations are threatened by anthropogenic pressures. Monitoring the distribution and diversity of wild bees becomes daunting as it involves extensive field surveys, sample collection, and traditional taxonomy. Logistic and ethical issues complicate this further. Machine learning (ML) on passively collected visual data can provide a non-invasive, large-scale solution to these problems. However, a significant challenge for ML application is the lack of geographically varying training data for different bee species, especially from species-rich tropical regions. In addition, these algorithms have predominantly been tested on image data collected under controlled conditions inside laboratories. ML models must be trained to perform on field-collected unrefined images of different bee species for rapid yet non-invasive diversity estimation. In this study, to challenge and bolster the existing deep learning networks, we collected 3250 field-captured images of major South Asian social bees ( A. dorsata, A. cerana, A. florea , and Tetragonula spp. ). We did not control these images for lighting and camera perspectives, which posed significant challenges to the models. We benchmarked this image dataset using standard convolution neural network (CNN) models, finding that MobileNet-V2 was the best model, achieving the highest accuracy of 98.4%.