On-device deep learning for real-time acoustic monitoring of endangered Bombus dahlbomii and invasive congeners

The decline of endemic pollinators, such as the Patagonian bumblebee ( Bombus dahlbomii ), seemingly driven by invasive species, necessitates innovative and scalable monitoring solutions. Traditional methods are often labor-intensive or logistically challenging for long-term, widespread deployment. This study addresses these challenges by developing and evaluating an on-device machine learning (ML) system for real-time acoustic classification of the native B. dahlbomii and the invasive Bombus terrestris . Utilizing a previously established acoustic dataset from Puerto Blest, Argentina, we created a lightweight and optimized convolutional neural network (CNN) architecture for deployment on the resource-constrained Analog Devices MAX78000 microcontroller, integrated within an updated BuzzCam acoustic sensor platform. Through rigorous optimization of data preprocessing techniques and implementation of quantization-aware training (QAT), our final 8-bit quantized model achieved a classification accuracy of 86.1% on a held-out test set for the three-class problem ( B. dahlbomii, B. terrestris , Negative). Deployed on the MAX78000, the system processes a 1-second audio segment with an inference latency of 10.4 ms and consumes only 794 uJ during active classification. The deployed model consists of only 158,144 parameters, with both weights and activation quantized to 8 bit integers. These results demonstrate the feasibility of accurate, low-power, on-device acoustic monitoring for Bombus species. This work paves the way for cost-effective, scalable, and autonomous sensor networks capable of providing near real-time ecological insights, significantly enhancing our ability to understand pollinator dynamics and inform timely conservation strategies for endangered species like B. dahlbomii .