Discussion
This study found differences in the performance and ecological plausibility of ENMs that incorporated temporal occurrence-environment matching compared to models that used a standard 30-year average. Using Cryptotis mexicanusas a use-case, the results indicated that the ten-year temporal resolution performed similar to the standard 30-year average approach based on cross-validation AUC, CBI and OR (with no significant difference). Moreover, the results even showed an improvement in performance when considering the fully withheld omission rate, with an average rate that is lower by 0.1 than the standard approach. However, given the significantly lower AUC values of other shorter temporal resolutions (one and five years), this result may not be generalized to all temporal resolutions. The ecological plausibility interpretation supported the findings of model performance, showing that the ten-year resolution had as much of a realistic geographical prediction of suitability as the standard 30-year approach for almost all the known shrew range, while the shorter temporal resolutions resulted in unrealistic estimates of the potential shrew distribution.
Time-matched approaches have proven successful in obtaining reliable model predictions for other species of small non-volant mammals (with low vagility), but none of these studies compared the performance and ecological plausibility of different temporal resolutions, as presented in this study. For instance, time-specific models of American Pika (Smith et al. 2019) and Australian Bettongs (Bateman et al. 2012) achieved favourable results using resolutions of ten and five years, respectively. The observed differences in model performance across the three temporal resolutions here suggest that using an uninformed time-matched approach will not translate into better models for small non-volant mammals or other species with low mobility. Ideally, the temporal resolution selection should be based on the species’ biology, specifically the time lag in which a population could be impacted by the change in the variables used in the model training (Levin 1992, Nadeau et al. 2017). For instance, shorter-term resolutions, such as yearly, monthly, or even daily resolutions (Reside et al. 2010, Abrahms et al. 2019), might be more useful for highly mobile species (Ingenloff and Peterson 2021) than for species of low mobility, such as C. mexicanus . Unfortunately, most species worldwide lack the necessary demographic information to establish the time lag of population response, although often information could be borrowed from taxonomically related species with similar body sizes and natural history information. As an alternative, time-matched models should explore several temporal resolutions and tune them empirically.
For the ten-year temporal resolution, the modified version of the target-group background obtained realistic predictions of the shrew’s distribution, except for the northern region. This large area predicted in the north is not expected, as it corresponds to xeric shrublands and mesic lowland areas in which populations of C. mexicanus never have been found. This could be due to the minimal representation of the background data in the northern area (as seen in Fig. 3b), which may fail to capture the full range of environmental conditions due to under-sampling by biologists. In such cases, the target-group background tends to overcompensate for the least sampled areas by increasing the number of false positives (Syfert et al. 2013, Barber et al. 2022). This tendency can be observed even in the standard 30-year average when used to correct for spatial bias, but it increases when also correcting the temporal bias in occurrence-environment approaches. Furthermore, while the model performance may be affected by occurrences removed because they fell outside the temporal range between 1971 and 2000, the time-matched approaches allow the use of occurrences traditionally discarded in model training, such as occurrences from the exact location but captured at different times, which could experience different environmental conditions. This situation can benefit species with a restricted distribution or where spatial thinning (Boria et al. 2014) may significantly reduce the occurrence data available for modelling.