Model comparison
After conducting repeated random k -fold cross-validation, I
observed that the ten-year resolution approach outperformed the standard
30-year average approach in terms of the omission rate for withheld
data, while the other resolutions performed the same or worse based on
some validation metrics. First, the AUC validation values of the
standard approach were significantly higher than those of the one-year
and five-year time-matched approaches (Fig. 6a). When comparing with the
ten-year time-matched approach, the AUC values of the standard approach
were still higher, but the difference was not significant. The
Continuous Boyce Index (Fig. 6b) and the 10thpercentile of training omission rate (Fig. 6c) showed similar
(non-significant) values between the standard and time-matched
approaches. However, the omission rate based on the 18 fully withheld
occurrences (OR.W) indicated a higher omission rate for the 30-year
standard approach than for each of the three temporal resolutions (Fig.
6d).
Visual inspection showed that the standard 30-year average and the
ten-year time-matched approaches provide more accurate predictions than
the other two approaches (Fig. 7 and Fig. S1). The one-year matching
approach overpredicts suitability in the northern and southern parts of
the study area, while the predictions of the five-year matching approach
are narrower than the one-year approach but still contain some
unrealistic predictions of suitable conditions to the west of the Sierra
Madre Oriental between 20°N and 22°N. In contrast, the standard and
ten-year matching approaches have similar predictions. However, there is
a larger prediction at the north of the potential distribution for the
ten-year matching approaches. A more detailed comparison of ecological
plausibility between these two approaches is found in Supplemental
material (Fig. S2). Moreover, the number of times the withheld data were
predicted as suitable (represented by black circles in Fig. 7) shows
that the time-matched approaches have higher prediction rates than the
standard approach (which supports the quantitative test).