Results

Model development with data from Switzerland

We trained our models with Red Kite tracking data from Switzerland that encompassed 697 individual seasons from a total of 258 tracked individuals, many of which were tracked over several years (range 1 - 7). The presence (n = 360) or absence (n = 337) of a home range and a nesting attempt (yes = 296, no = 401) were known for all individual seasons, and the outcome of a nesting attempt was known for 257 individual seasons, of which 177 successfully raised at least one fledgling, while 80 nesting attempts failed.
Our model to predict the presence of a home range classified 97.6% of individual seasons correctly in the cross-validated test data (Table 1). The most influential variables in predicting the existence of a home range were the total amount of daytime spent within a potential nest radius (residence_time_day), the median distance between daytime and the most frequently used nighttime locations (median_day_dist_to_max_night), and the number of revisits to a potential nest site during the day (revisits_day; Fig. S2).
Our model predicting the presence of a nesting attempt classified 96.7% of individual seasons correctly in the cross-validated test data (Table 1), while 9.3% of individual seasons could not be classified with certainty (Fig. 1). The most influential variables in predicting a nesting attempt were again the total amount of daytime and nighttime spent within a potential nest radius (residence_time_day, residence_time_night), and the median distance between daytime and the most frequently used nighttime locations (median_day_dist_to_max_night; Fig. S3). The process to prepare the data, train the model, and predict the existence of nesting attempts for 697 individual seasons took 17 minutes.
Our model to predict the outcome of a nesting attempt classified 83.1% of individual seasons correctly in the cross-validated test data (Table 1). If only those seasons with high (> 75%) probability of assignment were used, the accuracy increased to 97.9% of correctly classified outcomes, while 26.5% of all seasons could not be classified with 75% certainty and were therefore flagged up for manual annotation (Fig. 2). The most influential variables in predicting breeding success were the number of revisits to the nest site during the second chick phase (revisitsChick2), and the amount of time spent within the nest radius during the second chick phase (timeChick2; Fig. S4). The process to train the model, and predict the outcome of nesting attempts for 257 individual seasons took 2 minutes.
Table 1. Proportion of correctly predicted reproductive parameters from GPS tracking data of Red Kites in Germany and Switzerland using the newly developed NestTool. Data subset indicates whether the data were used for model fitting (‘training’) or not (‘validation’). nindicates the number of individual breeding seasons that were classified, but note that the number for breeding success was lower because not all individuals initiated a nesting attempt.