Fig 1.
(A) Field deployment of the camera trap and flower platform (35x20 cm) on wooden post. (B) Weatherproof camera trap enclosure with integrated hardware and connected solar panel.
Table 1.
Metrics of the YOLO insect detection models.
Fig 2.
Diagram of the processing pipeline.
HQ frames (1080p or 4K) are downscaled to LQ frames (320x320 pixel), while both HQ and LQ frames run in two parallel streams. The LQ frames are used as input for the insect detection model. The object tracker uses the coordinates from the model output to track detected insects and assign unique tracking IDs. Detected insects are cropped from synchronized HQ frames in real time and saved to the microSD card of the RPi together with relevant metadata (including timestamp, tracking ID, coordinates).
Fig 3.
LQ frame and synced HQ frame with cropped detection.
Downscaled LQ frames (320x320 pixel) are used as model input. Detected insects are cropped from synchronized HQ frames (1080p or 4K) on device.
Fig 4.
Power consumption of the camera trap system.
The two power spikes at ~5W represent the start and end of the recording. Power consumption was measured during a 5 min recording interval while constantly detecting and tracking five insects, and saving the cropped detections together with metadata to the RPi microSD card every second.
Fig 5.
Example images of the 27 classes from the dataset for classification model training.
All images were captured automatically by the camera trap and are shown unedited, but resized to the same dimension. Original pixel values are plotted on the x- and y-axis for each image. Four example images are shown for the class “other”, to account for its high intraclass heterogeneity.
Fig 6.
Normalized confusion matrix for the EfficientNet-B0 insect classification model, validated on the dataset test split.
The cell values show the proportion of images that were classified to a predicted class (x-axis) to the total number of images per true class (y-axis). The model was trained on a custom dataset with 21,000 images (14,686 in train split). Metrics are shown on the dataset test split (2,125 images) for the converted model in ONNX format.
Table 2.
Metrics of the EfficientNet-B0 insect classification model, validated on the dataset test split.
Table 3.
Recording times of the five deployed camera traps and site details of the orchard meadows.
Fig 7.
Evaluation of the insect tracking accuracy in a lab experiment.
Data from 10 replications of 15 min recording intervals with 15 E. balteatus hoverflies placed in a cage with the camera trap and flower platform is shown. Linear regression lines illustrate the effect of 10 different tracking IDs filter methods for post-processing of the captured metadata of each recording interval. With “-min_tracks 1” no tracking IDs are excluded. Dashed line indicates optimal result. True frame visits were manually counted from video recordings of the flower platform.
Fig 8.
Normalized confusion matrix for the EfficientNet-B0 insect classification model, validated on a real-world image dataset.
The cell values show the proportion of images that were classified to a predicted class (x-axis) to the total number of images per true class (y-axis). Metrics are shown on a real-world dataset (97,671 images) for the converted model in ONNX format. All images were classified and subsequently verified and sorted to the correct class in the case of a wrong classification by the model. A dummy image was added for each of the classes “bug_grapho” and “fly_empi”, as no images of both were captured. Results for both classes must be ignored.
Fig 9.
Maximum air and OAK-1/RPi CPU temperatures per day.
Weather data was taken from the nearest weather station (source: Deutscher Wetterdienst).
Fig 10.
Mean humidity per day measured inside the enclosure and at nearest weather station.
Weather data was taken from the nearest weather station (source: Deutscher Wetterdienst).
Fig 11.
Mean PiJuice battery charge level and sum of the sunshine duration per day.
The PiJuice battery was charged by a second battery, connected to a 9W solar panel. Weather data was taken from the nearest weather station (source: Deutscher Wetterdienst).
Fig 12.
Total number of unique tracking IDs for each predicted insect class.
(A–E) Data from camera traps 1–5. (F) Merged data from all camera traps. Data of non-insect classes not shown. All tracking IDs with less than three or more than 1,800 images were removed.
Fig 13.
Time difference to previous tracking ID classified as the same class.
Minimum time difference < 30 s of the previous five tracking IDs classified as the same class is shown. Data of non-insect classes not shown. All tracking IDs with less than three or more than 1,800 images were removed.
Fig 14.
Total number of unique tracking IDs classified as hoverfly and recording time per day.
(A–E) Data from camera traps 1–5. (F) Merged data from all camera traps. Grey lines/areas indicate the recording time per day. Dashed lines indicate a change of the flower platform. All tracking IDs with less than three or more than 1,800 images were removed.
Fig 15.
Number of unique tracking IDs classified as hoverfly per hour and precipitation sum per day.
(A–E) Data from camera traps 1–5. (F) Merged data from all camera traps. Shaded areas indicate days without recordings. All tracking IDs with less than three or more than 1,800 images were removed. Weather data was taken from the nearest weather station (source: Deutscher Wetterdienst).
Fig 16.
Estimated hoverfly activity (number of unique tracking IDs per hour) per time of day.
(A–E) Data from camera traps 1–5. (F) Merged data from all camera traps. Shaded areas indicate hours without recordings. All tracking IDs with less than three or more than 1,800 images were removed.