Fig. 2: Example plots of the temporal sequence of three movement metrics
(median daily time spent at the nest, median of the daily maximum
distance from the nest, and maximum time away from nest) for three adult
female red kite individuals that failed to raise fledglings (left), bred
successfully (centre), and exhibited behaviour that resulted in
uncertain classification (right).
The overall breeding propensity of the tracked red kite population in
our study area in Switzerland was 82.2% when inferred from tracking
data (compared to 82.2% observed), and overall breeding success
inferred from the tracking data was 70.9% (compared to 68.9%
observed).
Transferring the models to populations in
Germany
We tested our models with red kite tracking data from Germany that
encompassed 163 individual seasons from a total of 65 tracked
individuals, some of which were tracked over several years (range 1 -
7). After down-sampling the data to 60-min intervals and filtering the
data to the seasonal phenology in Germany, we were able to predict
reproductive parameters for 99 individual seasons, which contained 98
seasons with home range, 89 nesting attempts, and 55 successful broods.
The process to prepare the data, predict the three responses of home
range, breeding propensity and breeding success, and summarise the data
took 130 seconds.
Our home range model predicted that 95 seasons had home range behaviour
- an accuracy of 94.9% (Table 1). The single season without a home
range was not correctly predicted, and had a higher predicted
probability than four other seasons with home range behaviour.
Compared to the 89 individual seasons that initiated a nesting attempt,
our nesting model predicted 87 seasons to breed - an accuracy of 89.9%
(Table 1). Out of the 10 individual seasons without a nesting attempt, 4
seasons were predicted to have a nesting probability of
>0.5. Conversely, of the 89 actual nesting attempts 6 had a
predicted nesting probability of <0.5, and were therefore
missed. When we classified only those nests for which predictions were
made with high certainty (>75%), the overall accuracy
increased to 91.8%.
The prediction of breeding success for the data from Germany was less
accurate: out of the 55 individual seasons that raised at least one
fledgling, our breeding success model predicted 83.6% as successful,
but erroneously predicted 18.6% of failed nests as successful, leading
to an overall accuracy of 82.7% (Table 1). For 69% of nesting attempts
the outcome could not be predicted with high certainty (<75%
assignment probability) and would have required manual annotation;
omitting these cases with low predictive certainty increased the
accuracy to 92.3%. When the data from Germany were used to train nest
success models, the predictive accuracy increased to 93.1% for all
cases and to 97.1% for cases that were classified with
>75% certainty (Fig. S5).