Research library

DeepBees – Building and Scaling Convolutional Neuronal Nets For Fast and Large-scale Visual Monitoring of Bee Hives

The decline of bee populations is a global trend and a severe threat to the ecosystem and to pollinator-dependent industries. Factor analysis and preventive measures are based on snapshot information; information about the health state of a hive is infrequently acquired and remains labor-intensive and costly. In this paper, the authors describe a system that enables near-time, scalable, and cost-efficient monitoring of beehives using computer vision and deep learning. The pipeline consists of four major components: (1) off-grid hardware at the hive gate capturing in/out streams of bees; (2) on-edge inference for bee localization and tracking of single entities; (3) a cloud infrastructure for device and data management with near-time sampling; and (4) a cloud-hosted deep convolutional multi-net (DeepBees) inferring entity-based health insights from individual bee images. DeepBees uses a shared encoder with task-specific decoders for genus identification, pollen classification, pose estimation, and unsupervised representation learning. The overall system was deployed by apic.ai and monitored 49 beehives in Karlsruhe, Germany.

Download PDF

Publication details

Authors
Julian Marstaller, Frederic Tausch, Simon Stock
Organizations
🇩🇪 Karlsruhe Institute of Technology🇩🇪 apic.ai
Year
2019
Type
Conference

Relevancy to Gratheon

DeepBees represents the commercial state of the art that Gratheon benchmarks against. The four-stage pipeline (edge capture → on-device detection → cloud transfer → cloud inference) is the same architecture Gratheon uses for its Entrance Observer product. The multi-task MobileNetV2-based DeepBees CNN — simultaneously covering pollen classification, genus identification, and pose estimation — demonstrates that a single lightweight backbone can cover all the individual-bee signals Gratheon wants to expose in its web app. The apic.ai deployment scale (49 hives, continuous data) also sets the operational target for Gratheon's precision-beekeeping platform.