Preprocessing and normalisation strategies

Sound and sensor data from beehives contain noise and irrelevant signals, which must be cleaned and transformed for analysis. The first step of preprocessing is noise filtering. In Qandour et al. (Reference Qandour, Ahmad, Habibi and Leppard2014), a low-pass Butterworth filter or band-pass filter is applied to restrain frequencies from 0 to 4 kHz, since the targeted sounds in this study usually reside below 1 kHz. The work described in Cejrowski et al. (Reference Cejrowski, Szymanski and Logofatu2020) focuses on the whole colony’s sound, so outlier sound samples where a single bee passes too close to the microphone are removed. Samples are only kept when the frequency of their most significant harmonic component is lower than 600 Hz, or when their root mean square (RMS) level does not deviate by more than 80% of the dataset average. Similarly, noise in images is caused by lighting variations or sensory artefacts. This is processed using Gaussian (Rathore et al., Reference Rathore, Tyagi and Agrawal2023), Bilateral (Alves et al., Reference Alves, M.a, Ventura, Neves, Biron and Junior2020), and Median Filter (Paolillo et al., (Reference Paolillo, Petrini, Casiraghi, De Iorio, Biffani, Pagnacco, Minozzi and Valentini2022). Further, to improve image quality, contrast enhancement is performed using histogram equalisation or adaptive contrast adjustment. To reduce the variability due to recording equipment and environmental conditions, the amplitude of sound signals is normalised to a range between −1 and 1. For images, image pixel values are scaled to 0–1.

Synthetic data generation via augmentation

Data augmentation is a technique employed to expand the training dataset by generating transformed images. Data augmentation reduces overfitting and enhances model learning capacity by enlarging the dataset. For sensor data, it handles missing values with interpolation and imputation, whereas for sound data, it segments the signal into smaller windows (typically 1–5 seconds) or by changing pitch, masking, etc. For image data, it increases data diversity by performing data translation techniques such as flipping, rotating, scaling, etc.

Feature extraction

Feature extraction in analysing sound and image data is used to extract the important information for classification, clustering and predictive modelling. It enables automation, improves accuracy while reducing the data complexity. The different types of features for sound signals are as follows:

• • Time domain features like RMS and zero-crossing rate quantify the vibrational intensity within the hive, indicating colony activity, communication, and swarming behaviour.

• • Frequency domain features like Mel-frequency Cepstral coefficients (MFCCs) (Zgank Reference Zgank2019), spectral centroid, and spectral roll-off describe the frequency distribution of bee sound.

• • Temporal features like autocorrelation is the measure of similarity of a signal with a time-shifted signal. It identifies periodic patterns in bee sound such as queen piping and swarming signals.

Whereas for image data, the features are as follows:

• • Colour-based feature represents the distribution of pixel intensities in RGB or greyscale. It identifies patterns in hive structure, such as honey storage or brood area.

• • Texture based like local binary patterns and grey-level co-occurrence matrix measure texture by analysing pixel intensity relationships. It detects variation in hive structure, such as comb texture or pest damage.

• • Shape based is used for species classification, cell identification, and anomaly detection. It identifies object boundaries like edges, identifies patterns, and shapes to track bee movement and bee count.

Prediction and classification techniques

Prediction and classification enhance beekeepers’ understanding of their hives’ necessities, allowing them to maintain hive health while minimising stress duration. This approach saves time, reduces physical visits, and improves accuracy. ML combines the principles of statistics and computer science to develop algorithms that exhibit enhanced efficiency when exposed to pertinent data, as opposed to being provided with explicit instructions (Jijo and Abdulazeez, Reference Jijo and Abdulazeez2021). The implementation of certain mathematical models and algorithms enables orientation and decision-making processes by analysing data and predicting potential outcomes and effects of a given circumstance. Following are some common techniques:

• • Time-series forecasting: These are used to predict hive states based on historical data. Algorithms like ARIMA (Auto Regressive Integrated Moving Average) (Box et al., Reference Box, Jenkins, Reinsel and Ljung2015), LSTM (Long Short-Term Memory), and Facebook Prophet are particularly effective for modelling and seasonality in sensor data. These methods help in forecasting swarming (Andrijević et al., Reference Andrijević, Urošević, Arsić, Herceg and Savić2022) and resource consumption, enabling proactive hive management.

• • Regression models: Regression techniques are employed to predict continuous outcomes such as honey yields, pest outbreaks or colony strength, based on input features such as environmental conditions and hive metrics. Linear regression model, random forest (RF), eXtreme Gradient Boosting (XGBoost; Brini et al., Reference Brini, Giovannini and Smaniotto2025), and regression tree (Chen and Guestrin, Reference Chen and Guestrin2016) are commonly used algorithms. Kaplan–Meier Estimator estimates survival probabilities over time. These models analyses the relationship between honey production and resource allocation.

• • ML model: Traditional ML algorithms, including support vector machine (SVM; Nolasco and Beltos, Reference Nolasco and Benetos2018; Kulyukin and Mukherjee, Reference Kulyukin, Mukherjee and Amlathe2018), Naïve Bayes (Bjerge et al., Reference Bjerge, Frigaard, Mikkelsen, Nielsen, Misbih and Kryger2019; Schurischuster and Kampel, Reference Schurischuster and Kampel2020), RF, and XGBoost (Chen and Guestrin, Reference Chen and Guestrin2016), are widely used for classification tasks like pest infestation, capped and uncapped cells, and brood identification, etc. (Rathore and Agrawal, Reference Rathore and Agrawal2024). Their interpretability and efficiency make them suitable for real- time applications. Anomaly detection methods like isolation forest and autoencoders are used to identify unusual patterns in hive data like sudden temperature drop, abnormal variations, and mites’ presence.

• • Deep learning: Deep-learning models such as convolutional neural networks (CNNs) and recurrent neural networks excel at image-based tasks like pollen-bearing detection (Sledevič, Reference Sledevič2018), brood pattern, and pollen sac detection (Kulyukin and Mukherjee, Reference Kulyukin, Mukherjee and Amlathe2018). YOLOV3 model (Ngo et al., Reference Ngo, K.c, E.c and T.t2019), UNet (Rathore et al., Reference Rathore, Tyagi and Agrawal2023), Inception V3 (Alves et al., Reference Alves, M.a, Ventura, Neves, Biron and Junior2020), and Mask region-based CNN (R-CNN) are ideal for image data, whereas RNN – and their variants – and LSTM are suitable for sound signal and time-series data. These techniques enable advanced applications, such as identifying bee communication patterns from audio data or predicting colony health from integrated sensor input. Supplementary Table S1 refers to the various deep-learning (DL) and ML techniques reported for different apicultural applications.

Sensor data parameters

Sensor data are specifically used to predict and evaluate colony activity through three major components: (i) swarming analysis, (ii) bee-traffic analysis, and (iii) internal hive-state assessment. Continuous sensing enables the evaluation of environmental conditions, resource dynamics, metabolic activity, and behavioural patterns. Swarming is a particularly critical event for beekeepers because it can reduce colony strength, honey production, and overall profitability. Bee-traffic patterns provide insights into foraging efficiency, pollen and nectar collection, and overall colony health. Internal hive conditions are assessed using several key parameters. Temperature variance indicates thermoregulation efficiency, colony stress, and responses to external temperature fluctuations (Cook et al., Reference Cook, Blackler, McGree and Hauxwell2021). Changes in hive weight reflect honey production, food-resource consumption, and possible swarm departure. Carbon dioxide concentration provides information about hive ventilation and the metabolic activity of the colony. Humidity levels help identify moisture imbalance, inadequate ventilation, and environmental stress, while flight activity, measured over the observation period, supports the prediction of foraging behaviour, traffic intensity, and swarming activity. The mathematical definitions and detailed interpretations of these features are provided in Supplementary Table S2.

Sound data parameters

The sound parameters reveal information about the pest attack, quenlessness, and swarming activity. The recorded hive audio is treated as a real-valued time-series signal and analysed using features such as MFCCs, spectral centroid, zero-crossing rate, RMS energy, temporal entropy, and statistical descriptors including mean, variance, skewness, and kurtosis. These features characterise the spectral distribution, signal intensity, complexity, and temporal variability associated with different colony conditions. Detailed definitions and calculations of the audio features are provided in Supplementary Table S3.

The classification performance was assessed using accuracy, positive and negative predictive values, sensitivity, specificity, F1-score, true-positive rate, and false-positive rate, derived from the numbers of true-positive, true-negative, false-positive, and false-negative predictions. The receiver operating characteristic curve was generated by comparing the true-positive and false-positive rates across different classification thresholds. Its area under the curve (AUC) summarises the overall discriminative performance of the model, with a value closer to 1 indicating superior classification ability. The corresponding performance-metric calculations are presented in Supplementary Table S4.

Colony monitoring and analysis

Table 3 shows the summary of predictive analysis and methods reported in the literature for colony monitoring. To predict colony-level parameters, Braga and Madureira (Reference Braga and Madureira2020) used LSTM regression model to forecast hive temperature 24 hours in advance. Their algorithm predicts accurate 2- and 24-hour forecasts with RMSEs of 0.15 and 0.18, respectively. Robustillo et al. (Reference Robustillo, Pérez and Parra2022) used a time-varying vector autoregression model to predict temperature, humidity, and weight for 1, 3, and 7 days, achieving an absolute error (MAE) of 0.70 for 1-day temperature forecasts and 0.84 for 7-day forecasts, with humidity and weight predictions following suit. Huet et al. used linear regression to forecast interior hive conditions and inform beekeepers of abnormalities. This method kept hives stable and allowed quick environmental stress response. To improve predictive accuracy, Anwar et al. (Reference Anwar, Keating, Cardell-oliver, Datta and Putrino2023) created a DL-based weight prediction model employing a bidirectional LSTM framework and attention-based encoder–decoder structure. The study gathered 57,600 data points on internal and external environmental factors every 30 minutes from eight hives in three locations. The model’s MAE of 0.64 shows DL’s promise for hive monitoring. Audio-based monitoring by Robles-guerrero et al. (Reference Robles-guerrero, Saucedo-anaya, González-ramírez and De la Rosa-vargas2019) predicted hive condition with the presence and absence of a queen bee. The authors calculated MFCC and classification accuracy up to 0.97. Further, Cejrowski et al. (Reference Cejrowski, Szymanski and Logofatu2020) classified bee circadian rhythms into day and night cycles using Fast Fourier Transform (FFT) and SVM models; they identified bee activity patterns with 81% accuracy from 809 data points over 9 days. Additionally, Cejrowski and Szymanski (Reference Cejrowski and Szymanski2022) used a contrastive convolutional variational autoencoder to identify feeding behaviour abnormalities in hives. They found that glucose-supplemented hives behaved differently over 5 days. Alves et al. (Reference Alves, M.a, Ventura, Neves, Biron and Junior2020) and Rathore and Agrawal (Reference Rathore and Agrawal2024) predicted colony strength and productivity using images. They employed DL frameworks like Inception V3, ResNet to classify cells in seven and three different classes, respectively, and achieved up to 98.7% and 94% accuracy.

Table 3.

Prediction analysis and methods used in literature for colony monitoring analysis

Table 3 Long description

Eight studies are listed with columns for references, data type, samples, data feature, analysis, preprocessing method, and performance metric. Data types include temperature and sound, sound alone, sound with temperature pressure and gas, temperature pressure and gas with camera, image, and temperature with pressure. Sample counts reported are 809, 720, 2048, 720, 2204, and 62, with two rows showing no sample value. Data features include sampling rate of 3000 hertz, sampling rate of 24000 hertz in two rows, resolution in two image rows, and units listed as C and grams in two rows, with two rows showing no data feature value. Analyses include machine learning with SVM, machine learning with SVD, machine learning with an autoregressive model plus GAM and DLM, and deep learning with LSTM, FCNN with Unet plus a maturity model, and Unet with InceptionV3 in two rows. Preprocessing methods listed are MFCC with discrete cosine transform, windowing with MFCC, spatial temporal analysis, MFCC with linear regression and LASSO regularization, CHT, CLAHE with CHT, and MICE, with one row showing a dash. Performance metrics reported are accuracy 81.41 percent, area under curve 0.98, mean absolute error 0.64, area under curve 0.97, accuracy 98.7 percent, accuracy 94 percent, and mean absolute error 0.84, with one row showing no performance metric.

DLM, dynamic linear model; GAM, generalised additive model; MICE: Multivariate Imputation by Chained Equations with Random Forest; SVD, singular value decomposition; SVM, support vector machine; CHT, circle Hough Transform; CLAHE, contrast adaptive histogram equalisation; MFCC, Mel frequency Cepstral coefficient.

Swarming activity analysis

Swarming is a critical reproductive behaviour in bee colonies, marked by the departure of the queen and worker bees to establish a new colony. Driven by factors such as colony overcrowding or environmental stress, swarming involves three stages: (i) preparation, (ii) swarming, and (iii) scouting. Early detection of swarming is vital for preventing colony loss and maintaining apiary productivity. Table 4 summarises the prediction analysis and methods reported in the literature for swarming activity. To examine the acoustic and ecological indicators of swarming, Anand et al. (Reference Anand, Raj, Ullas and Srivastava2018) incorporated multiple sensors into a hive and performed power spectrum analysis using a Fast Hartley Transform method with 256 frequency components and an anti-aliasing filter. The study demonstrates that the 400-Hz band activation and significant hive weight loss after swarming were strong indicators of swarming. Hive temperatures fluctuated from 28°C to 32°C before swarming to 34–38°C before lowering again after the swarm left. Zlatkova et al. (Reference Zlatkova, Kokolanski and Tashkovski2020) used the Welch technique to analyse acoustic data from six hives during a month. Their investigation indicated that power variations around 200 Hz preceded swarming, revealing distinct changes in audio patterns before and after swarm departure. Bencsik et al. (Reference Bencsik, Bencsik, Baxter, Lucian, Romieu and Millet2011) assessed swarming behaviour without the use of microphones and instead relied on accelerometers. For lower frequencies, accelerometers give more accurate readings (Bencsik et al., Reference Bencsik, Le Conte, Reyes, Pioz, Whittaker, Crauser, Simon Delso and Newton2015).

Table 4.

Prediction analysis and methods used in literature for swarming activity

Table 4 Long description

Rows list literature references with columns for data type, sample count, data feature, analysis, method, and performance metric. Senger et al. 2024 used temperature data with 60 samples, data feature in degrees Celsius, deep learning analysis, no method listed, and accuracy 85 percent. Andrej et al used sound data with 2443 samples, sampling rate 44100 hertz, deep learning with DNN analysis, MFCC method, and accuracy 94.09 percent. Dimitrios et al. 2022 used sound data with 1435 samples, sampling rate 22400 hertz, deep learning with SVM and UNet analysis, low pass filtering and fast Fourier transform method, and performance metrics of accuracy 95 percent and F1 score 98 percent. Zgank et al. used sound data with 2443 samples, sampling rate 16000 hertz, machine learning with HMM and GMM analysis, LPC and MFCC methods, and accuracy reported separately for GMM at 79.47 and for HMM at 82.43. Zlatkova et al used sound data with no sample count shown, sampling rate 22050 hertz, statistical analysis labeled PSD, Welch method, and performance metric listed as classification. Anand et al used sound, temperature, and pressure data with no sample count shown, data feature listed as sampling rate and C, statistical analysis, Harley transform method, and performance metric listed as classification.

DFA, discriminant function analysis; PCA, principal component analysis; LPC, linear principle coding; GMM: Gaussian mixture model; HMM, Hidden Markov model; MFCC, Mel frequency Cepstral coefficient; PSD, power spectral density.

The shift from traditional statistical models to ML/DL has revolutionised swarming detection, enabling real-time and early predictions: Zgank (Reference Zgank2018) employed Hidden Markov Models and Gaussian Mixture Models with MFCCs and Linear Predictive Coding for feature extraction, achieving 80.4% accuracy. A subsequent Deep Neural Network (DNN) model improved accuracy to 94%, though audio compression reduced performance. Dimitrios et al. (Reference Dimitrios, Bellos, Stefanou, Stergios, Andrikos, Katsantas and Kontogiannis2022) utilised a U-Net CNN with Mel spectrograms, achieving 89% accuracy 10 days before swarming and 95% accuracy 5 days prior. This approach demonstrated the feasibility of early intervention.

Temperature fluctuations serve as a reliable proxy for swarming prediction: Zacepins and Karasha (Reference Zacepins and Karasha2013) observed temperature spikes of 1.5–3.4°C 8–20 minutes before swarming in Latvian hives. A rule-based system using temperature data achieved high precision. Kviesis and Zacepins (Reference Kviesis and Zacepins2015) trained a DNN on 20-hour temperature datasets, detecting nine visually confirmed swarming events and two unnoticed by beekeepers. Senger et al. (Reference Senger, Gruber, Kluss and Johannsen2024) combined temperature, humidity, pressure, and weight data with Dynamic Time Warping, achieving 85% accuracy in predicting swarming 24–48 hours in advance. Vibrational data provide insights into colony dynamics and swarming intentions: Ramsey et al. (Reference Ramsey, Bencsik and Newton2017) attached accelerometers to brood frames and used FFT to create two-/three-dimensional vibrational spectra. Principal component analysis and discriminant function analysis classified swarming colonies with >90% accuracy, predicting events up to 30 days in advance.

Traffic analysis and activity mapping

Traffic volume and patterns reflect foraging efficiency, stress levels, and colony age, offering critical insights for beekeepers. Recent IHMSs leverage computer vision and ML techniques that have significantly improved bee traffic monitoring and foraging analysis (Clarke and Robert, Reference Clarke and Robert2018)]. Table 5 represents the prediction analysis and methods reported in the literature for traffic analysis. Kulyukin and Mukharjee proposed MOG2 motion detection algorithm to track bee exits, followed by a CNN-based classifier to distinguish bees from other objects. Their approach achieved 99.48% accuracy in identifying exiting bees.

Table 5.

Prediction analysis and methods used in literature for traffic analysis

Table 5 Long description

Columns list references, data type, samples, data feature, analysis, method, and performance metric for nine studies. Magnier et al. used image data with resolution as the feature, computer vision analysis, color processing, and reported accuracy of 100 percent. Mukharjee et al. used image data with 2,976 samples, resolution, computer vision analysis, no method listed, and reported DWT of 20.75. Kalyunki et al. appears three times: image data with 167,261 samples, resolution, deep learning analysis, MOG2 method, and accuracy of 99.48 percent; image data with 5,819 samples, resolution, deep learning analysis labeled YOLO, no method listed, and F1 score of 0.43; and image plus temperature data with 13,412 samples, resolution, machine learning analysis labeled RF and SVM, no method listed, and R of 0.74. Andrijevi et al. used sound plus temperature data with 480 samples, sampling rate as the feature, deep learning analysis labeled LSTM, SIFT method, and mean square error of 378.4. Stojni et al. used image data with 1,000 samples, resolution, machine learning analysis, color processing method, and IOU of 0.915. Yang et al. used image data with 1,400 samples, resolution, deep learning analysis labeled FCNN, FFT method, and accuracy of 93 percent. Ngo et al. used sound plus temperature data with 3,500 samples, resolution, deep learning analysis labeled YOLO, Otsu Threshold method, and F1 score of 0.94.

FFT, Fast Fourier transform; SIFT, scale invariant feature transform; MOG, mixture of Gaussian; RF, random forest; LSTM, long short-term memory; YOLO, You Only Live Once; FCNN, Faster CNN.

To further improve traffic volume prediction, Kulyukin et al. tested various YOLO models and compared YOLOv3, YOLOv4-tiny and YOLOv7-tiny on a 5819-image dataset. YOLOv3 outperformed others with F1-score of 0.43, while YOLOv7-tiny proved more energy efficient. Furthermore, Kulyukin and Mukharjee utilised directional particle image velocimetry (DPIV) to predict bee traffic as lateral movement and in-–out traffic. This method categorised bee traffic by calculating the pixel-level movement vectors of individual bees and classifying them by direction relative to the hive entrance. The DPIV technique with human counts scored 20.75 in high-traffic circumstances, whereas the VGG16 method and human counts scored 22.02. This shows that the DPIV technique classifies high-traffic situations better. In addition to image-based tracking, researchers have incorporated environmental factors to predict bee traffic. Kulyukin et al. (Reference Kulyukin, Mukherjee and Amlathe2018) utilised RF and SVM algorithms to predict bee invasion considering environmental conditions along with traffic statistics. The authors used BeePIV system to convert video frames into particle motion frames with uniform backgrounds and computed particle displacement vector fields for each pair of consecutive motion frames. RF performed better than SVM, with R ² values of 0.46–0.74 against 0.39–0.69. Andrijević et al. (Reference Andrijević, Urošević, Arsić, Herceg and Savić2022) further investigated multisensor integration for bee traffic forecasting. They employed ARIMA, Prophet, and LSTM models on sensor data to count in–out movement of bees. The LSTM model achieved the highest accuracy with an RMSE score of 426.4 for outgoing and 378.4 for incoming traffic, demonstrating its potential for predicting bee movement trends. Besides traffic monitoring, various studies have assessed bee pollen collection. Stojnic et al. (Reference Stojnic, Risojevic and Pilipovic2018) classified bee images based on the presence of pollen by combining VLAD with SIFT algorithms. This method classified pollen-carrying bees with an Intersection over Union score of 0.797 and an AUC value of 0.915. Using YOLOv3-tiny and Kalman filtering for hive entrance tracking, Ngo et al. (Reference Ngo, K.c, E.c and T.t2019) developed an automated pollen monitoring system. The model showed a high correlation between pollen collecting rates and environmental parameters, including temperature and light intensity, with an F1-score of 0.94 for pollen detection and R ² score of 0.91 for traffic counting. Yang and Collins (Reference Yang and Collins2019) isolated pollen-bearing bees using HSV-based colour segmentation and Faster R-CNN. The data-augmentation-trained model achieved 93% classification accuracy.

Threat analysis and risk forecasting

Schurischuster and Kampel (Reference Schurischuster and Kampel2020) introduced a DeepLabV3-based semantic segmentation model using ResNet and AlexNet architectures to classify healthy and Varroa-infected bees. Their dataset comprised 13,500 annotated bee images, with the best models achieving over 90% classification accuracy. This approach outperformed traditional methods by accurately segmenting mite-infected pixels. Mrozek et al. (Reference Mrozek, Gȯrny, Wachowicz and Małysiak-mrozek2021) proposed an embedded system for Varroa mite detection using a Raspberry Pi 4 and Coral TPU accelerator. Their two-stage detection framework first identifies individual bees, followed by Varroa mite detection using Google AutoML-trained models. The system, deployed on Amazon Web Services, achieved an F1-score of 0.8, integrating automated beekeeper notifications. Picek et al. (Reference Picek, Novozamsky, Frydrychova, Zitova and Mach2022) presented a smartphone-based mite infestation monitoring system, where images of hive fallout debris undergo thresholding, Region of Interest (RoI) selection, and CNN classification using VarroaNet. Their lightweight model variants (VarroaNet-0.1 and VarroaNet-0.05) achieved infestation detection accuracies of 96.0% (autumn) and 93.8% (winter). However, occlusion errors in RoI preselection remain a limitation. VoudIotis et al. (Reference VoudIotis, Moraiti and Kontogiannis2022) combined Fast R-CNN for bee detection and Hough Transform for mite identification, utilising dual cameras – one monitoring the brood and the other capturing frame-top views. While achieving 88% bee detection accuracy and 77% mite detection accuracy, camera placement above bee space posed risks of brood cooling. Mahajan et al. (Reference Mahajan, Mehta, Miranda, Pinto and Patil2023) applied sound analysis with Mel spectrograms to classify colony strength and compared it against SSD, YOLO, and DETR object detection models. Their YOLOv5s model achieved the highest mAP score of 0.97, demonstrating DL’s efficacy in mite detection. Braga and Madureira (Reference Braga and Madureira2020) developed a decision-support system that used multiple cameras for hive surveillance. A dataset of 2,750 images was collected and processed using the Mask R-CNN model, achieving 85% accuracy in hornet detection. Terada and Satoh, Reference Terada and Satoh2022 implemented a network of Raspberry Pi cameras for real-time detection, capturing 1,400 images from multiple angles. The YOLOv3 model outperformed other DL models, achieving 88.9% accuracy in detecting hornets. De Jong and Lester (Reference De Jong and Lester2023) introduced a portable hornet detection system based on YOLOv5s, processing high-resolution images (3840 × 2160). Accuracy varied depending on camera distance: 96.7% at 40 cm, 93.8% at 55 cm, and 82.5% at 90 cm, demonstrating the importance of optimal camera positioning. Kwon et al. improved YOLOX models with ShuffleNet features, enhancing small-object detection. Their model achieved a mAP of 93.8%, surpassing the vanilla YOLOX model (mAP = 89.6%), and performed inferences 1.34 × faster. O’Shea-Wheller et al., Reference O’Shea-Wheller, Corbett, Osborne, Recker and Kennedy2024. deployed cameras at hornet bait stations, integrating an IoT-based tracking system. Motion detection using ViBe segmentation and training with the YOLOv5 model resulted in F1-scores of 0.98–0.99, proving high reliability.