TinyML On-Device Detection of Hive Anomalies and Varroa Mites

What makes TinyML different from ordinary hive analytics

TinyML refers to running machine learning inference directly on extremely resource-constrained hardware — microcontrollers or low-power single-board computers with only kilobytes to a few megabytes of memory and no reliable continuous power supply — rather than sending data to a cloud server for processing. For remote apiary sensors, this distinction matters enormously: a model running locally can make a decision and trigger an alert instantly, using a tiny fraction of the power and bandwidth that streaming raw data to the cloud for the same decision would require, and it keeps working even during connectivity outages that would otherwise leave a cloud-dependent system blind.

The tradeoff is that TinyML models must be small and efficient enough to run on hardware with a fraction of the compute power of even a modest smartphone. This generally means using compact model architectures and quantizing them down to 8-bit integer precision, which shrinks both memory footprint and computation time substantially compared to a full-precision model, usually at a small, carefully managed cost to accuracy that is acceptable for the binary or few-class decisions these systems typically need to make.

Temperature and humidity anomaly detection

A colony's internal temperature and humidity follow a fairly narrow, well-regulated band when things are going well — bees actively thermoregulate the brood nest and manage moisture through fanning and, in extreme heat, evaporative cooling. Deviations from that expected band can indicate a range of problems: queenlessness disrupting normal brood-nest thermoregulation, disease reducing the workforce available to regulate conditions, ventilation blockages, or external weather stress exceeding the colony's buffering capacity.

A practical on-device pipeline samples temperature and humidity every one to five minutes into a small non-volatile memory ring buffer, then extracts features such as sliding-window statistics and simple frequency-domain characteristics that capture whether readings are behaving normally or drifting into unusual territory. A lightweight model — often a one-class support vector machine or a small autoencoder trained only on normal operating conditions — scores each window for how anomalous it looks relative to what it learned as typical, and that score passes through a hysteresis-based threshold before triggering a low-power radio alert, avoiding the false-alarm flapping that a naive fixed threshold would produce on borderline readings.

On-device varroa mite detection from images

Varroa destructor remains one of the most consequential threats to managed honeybee colonies, and catching an infestation early meaningfully improves treatment outcomes compared to discovering it only once damage is severe. On-device vision systems aim to automate part of this detection by analyzing images of brood cells for visible mite presence, using a camera positioned to capture comb imagery either during routine inspection or through a semi-permanent in-hive mount.

Building a workable detector starts with a labelled dataset of brood cell images balanced between mite-positive and clean examples, since an imbalanced dataset dominated by clean cells — which is the natural state of most real hives most of the time — will bias a naively trained model toward under-detecting the rarer positive cases that actually matter. A compact architecture such as MobileNet, further optimized for hardware with a dedicated edge accelerator where available, is then quantized to INT8 precision and deployed through a runtime like TensorFlow Lite or ONNX Runtime suited to microcontrollers or small single-board computers.

Because false negatives on a genuine parasite threat carry real consequences, tuning the precision-recall tradeoff deliberately toward higher sensitivity is usually the right call for this specific application, even if it means somewhat more false positives requiring a beekeeper to double-check. Periodic retraining as more labelled field data accumulates helps the model keep pace with variation in comb appearance, lighting conditions, and mite presentation across different apiaries and seasons.

Powering these systems in the field

Both applications need to run for extended periods on limited power in a remote apiary setting, which shapes hardware choices throughout. Duty cycling — waking briefly to sample and process, then returning to a deep sleep state — is the single most effective power-saving technique available, since the compute and radio components draw far more current than the sensor itself during idle periods. Solar-charged battery systems sized appropriately for the local climate's worst-case winter sunlight availability are the common power solution for installations expected to run unattended for a full year or more, and both the audio, image, and environmental sensing applications discussed here typically transmit their compact alert payloads over the same low-power wide-area links, such as LoRaWAN or NB-IoT, used elsewhere in a smart apiary deployment.