Model selection
Along with the 53 occurrences, 1401 background points were selected and
used for model training. The best models selected by the delta.AIC for
each of the four approaches shared a similar combination of the feature
classes (Linear, Quadratic, and Hinge), without hinge features in the
one-year temporal resolution (parameters and validation metrics of each
model are shown in Fig. 5). Selected regularization multipliers range
from 0.3 to 2.2. The standard approach yielded the highest validation
AUC, although it had the lowest CBI compared to the time-matched
approaches. The validation omission rate (OR) of the 30-year average was
slightly higher for the standard approach than for the other
time-matched approaches. In terms of the geographic predictions of each
approach, shorter temporal resolutions resulted in more extensive
predictions (Fig. 5).