Software evaluation
To evaluate the software and measure the precision of the analyses and
generated outputs, we recorded a total of 23 videos from 15 nesting
units during two consecutive days using the nesting units as described
in Fig. 1. All nesting units were placed in large flight cages (54 m2)
that contained sufficient floral sources for offspring provisioning by
nesting female Osmia bicornis (sown purple tansy, buckwheat
and/or field mustard). A total of 24 females marked with the above
described 24 unique digit-color ID tags were released into each of these
flight cages and videos were recorded after initiation of nesting. Each
video was recorded between 9 AM and 3 PM when flight activity was high
for 2 to 4 hours. Cameras were placed with a distance of 1 m from the
nesting unit with frontal view (camera placed at same height as nesting
unit). From the recorded videos 8 randomly selected ones were used to
train the software to this experimental setup while the remaining 15
videos were used to measure precision. Precision was assessed by
visually checking 180 randomly selected events (12 events per video) for
their correctness using the visualization option of the software
(see above). Only events that were used for the generation of output csv
files (after error correction) were inspected as described above.
For the comparability of bee health under different environmental
conditions (e.g., different field sites with variable habitat quality or
flight cages with/without pesticide application) a similar precision
across videos is required. We therefore tested if precision varied
between videos in our set of evaluated videos. Therefore, a generalized
linear model with a binomial distribution was used. The correctness of
the detected event (correct or wrong) was included as response variable
and the video as explanatory variable. We further fitted a regression to
test if the proportion of females that can be assigned to a nest cavity
depends on the video recording time. The analysis was done in R 4.1. (R
Core Team, 2021).