A continuing trend in precision apiculture is to use computer vision methods to quantify characteristics of bee traffic in managed colonies at the hive’s entrance. Since traffic at the hive’s entrance is a contributing factor to the hive’s productivity and health, we assessed the potential of three open-source convolutional network models, YOLOv3, YOLOv4-tiny, and YOLOv7-tiny, to quantify omnidirectional traffic in videos from on-hive video loggers on regular, unmodified one- and two-super Langstroth hives and compared their accuracies, energy efficacies, and operational energy footprints. We trained and tested the models with a 70/30 split on a dataset of 23,173 flying bees manually labeled in 5819 images from 10 randomly selected videos and manually evaluated the trained models on 3600 images from 120 randomly selected videos from different apiaries, years, and queen races. We designed a new energy efficacy metric as a ratio of performance units per energy unit required to make a model operational in a continuous hive monitoring data pipeline. In terms of accuracy, YOLOv3 was first, YOLOv7-tiny—second, and YOLOv4-tiny—third. All models underestimated the true amount of traffic due to false negatives. YOLOv3 was the only model with no false positives, but had the lowest energy efficacy and highest operational energy footprint in a deployed hive monitoring data pipeline. YOLOv7-tiny had the highest energy efficacy and the lowest operational energy footprint in the same pipeline. Consequently, YOLOv7-tiny is a model worth considering for training on larger bee datasets if a primary objective is the discovery of non-invasive computer vision models of traffic quantification with higher energy efficacies and lower operational energy footprints.

From June to October, 2022, we recorded the weight, the internal temperature, and the hive entrance video traffic of ten managed honey bee (Apis mellifera) colonies at a research apiary of the Carl Hayden Bee Research Center in Tucson, AZ, USA. The weight and temperature were recorded every five minutes around the clock. The 30 s videos were recorded every five minutes daily from 7:00 to 20:55. We curated the collected data into a dataset of 758,703 records (280,760–weight; 322,570–temperature; 155,373–video). A principal objective of Part I of our investigation was to use the curated dataset to investigate the discrete univariate time series forecasting of hive weight, in-hive temperature, and hive entrance traffic with shallow artificial, convolutional, and long short-term memory networks and to compare their predictive performance with traditional autoregressive integrated moving average models. We trained and tested all models with a 70/30 train/test split. We varied the intake and the predicted horizon of each model from 6 to 24 hourly means. Each artificial, convolutional, and long short-term memory network was trained for 500 epochs. We evaluated 24,840 trained models on the test data with the mean squared error. The autoregressive integrated moving average models performed on par with their machine learning counterparts, and all model types were able to predict falling, rising, and unchanging trends over all predicted horizons. We made the curated dataset public for replication.

Since bee traffic is a contributing factor to hive health and electromagnetic radiation has a growing presence in the urban milieu, we investigate ambient electromagnetic radiation as a predictor of bee traffic in the hive’s vicinity in an urban environment. To that end, we built two multi-sensor stations and deployed them for four and a half months at a private apiary in Logan, UT, USA. to record ambient weather and electromagnetic radiation. We placed two non-invasive video loggers on two hives at the apiary to extract omnidirectional bee motion counts from videos. The time-aligned datasets were used to evaluate 200 linear and 3,703,200 non-linear (random forest and support vector machine) regressors to predict bee motion counts from time, weather, and electromagnetic radiation. In all regressors, electromagnetic radiation was as good a predictor of traffic as weather. Both weather and electromagnetic radiation were better predictors than time. On the 13,412 time-aligned weather, electromagnetic radiation, and bee traffic records, random forest regressors had higher maximum $R^2$ scores and resulted in more energy efficient parameterized grid searches. Both types of regressors were numerically stable.

In precision apiculture, the portable digital camera is a cost-effective sensor for capturing hive images or videos used to quantify different colony variables. Openly accessible, well-annotated, interoperable cell-level image datasets are still the exception rather than the norm. This shortage constitutes a major barrier to AI-driven approaches aimed at automating image-based comb analysis. In this article, we present FAIRHiveFrames-1K, a publicly available dataset of 1265 annotated hive frame images (1920 × 1080 PNG) designed to facilitate research in AI-intensive image-based comb analysis automation. The dataset, derived from a 2013–2022 U.S. Department of Agriculture–Agricultural Research Service multi-sensor research reservoir, includes 124,669 annotated regions of interest for seven biologically meaningful categories consistent with comb analysis literature and standard hive inspection protocols. FAIRHiveFrames-1K is curated according to FAIR principles (Findable, Accessible, Interoperable, Reusable) and distributed under CC-BY 4.0 with standard annotation formats, fixed training and validation splits, and reproducible benchmarking artifacts. To establish preliminary baseline performance, we iteratively tuned four YOLO architectures (YOLOv8n, YOLOv8s, YOLOv11n, YOLOv11s) under a shared tuning protocol over the period of dataset growth.

We explored the stationarity, trend, and seasonality of the hive weight and in-hive temperature of ten managed honey bee (Apis mellifera) colonies at a research apiary of the Carl Hayden Bee Research Center in Tucson, Arizona, USA. The hives were monitored with electronic scales and in-hive temperature sensors from June to October 2022. The weight and temperature were recorded every five minutes around the clock. The collected data were curated into 2160 timestamped weight and 2160 timestamped temperature observations. We performed a systematic autoregressive integrated moving average (ARIMA) time series analysis to answer three fundamental questions: (a) Does seasonality matter in the ARIMA forecasting of hive weight and in-hive temperature? (b) To what extent do the best forecasters of one hive generalize to other hives? and (c) Which time series type (i.e., hive weight or in-hive temperature) is better predictable? Our principal findings were as follows: (1) The hive weight and in-hive temperature series were not white noise, were not normally distributed, and, for most hives, were not difference- or trend-stationary; (2) Seasonality matters, in that seasonal ARIMA (SARIMA) forecasters outperformed their ARIMA counterparts on the curated dataset; (3) The best hive weight and in-hive temperature forecasters of the ten monitored colonies appeared to be colony-specific; (4) The accuracy of the hive weight forecasts was consistently higher than that of the in-hive temperature forecasts; (5) The weight and temperature forecasts exhibited common qualitative patterns.