209
so, we used correlation analyses corrected according Clifford’s method to obtain the
effective degrees of freedom for Pearson’s coefficients while controlling for spatial
autocorrelation (Clifford et al. 1989). Statistical analyses were performed in R using
the package SpatialPack (Vallejos and Osorio 2014).
9.3.2 Results
Live oak models calibrated using different sources and combinations of environmental predictors (Table 9.1) within the ensemble framework generally provided
similar suitability distributions (Fig. 9.5). Interestingly, increasing the number of
predictors or increasing model complexity (S-SR2 in Table 9.1) affected model performance as measured by the Cohen’s Kappa coefficient and AUC (area under the
receiver operating characteristic curve) indices (Table 9.2), and thus affected the
geographic predictions: Complex models tended to have higher statistical performance but to underestimate the distributions of live oak species when compared
with simpler models (Fig. 9.5). Individual live oak species models made from S-RS
products and WorldClim differed somewhat in their performances (see Table 9.2
and Fig. 9.5). Models from WorldClim tended to have slightly better statistical performance in inferring species distributions based on the AUC and Kappa criteria.
However, these metrics do not capture differences in the precision and spatial resolution of the approaches. In several species, the WorldClim models predicted low
precision locations compared to the S-RS data. In particular, the IUCN (International
Union for Conservation of Nature) red-listed narrow endemic Brandegee Oak
(Quercus brandegeei) in southern Baja California is very imprecisely predicted
compared with the S-RS data. Using high-resolution interpolated climatic predictors did not improve the performance of individual models (WC25 in Table 9.2) and
returned similar suitability predictions to those estimated under lower spatial resolution climatic predictors (Fig. 9.5). Although WorldClim models seems to have
better statistical performance as shown in Table 9.2, we can at most discriminate the
accuracy of interpolating continuous surface-derived models, only when we are
inferring habitat suitability models (ENM) and not the projected species geographical distribution (SDM). Using S-RS data as predictors not only helps to identify the
species habitat suitability but also incorporates local ecological conditions necessary to predict local species distributions and co-occurrence (Radeloff et al. 2019).
This is because S-RS data have the potential to get at biological mechanisms, for
example, through the detection of species phenological variation over space and
time (Figs. 9.3c and 9.4).
When macroecological patterns of species richness and uncertainty maps were
constructed, we observed similar patterns of species richness between maps made
from the simpler combination of S-RS and WorldClim models (Fig. 9.6a, c, and d;
see Table 9.1 for a description of the environmental combinations of S-RS and
WorldClim). Notably, species richness estimation from the complex S-RS tends to
restrict live oak assemblages to southeastern North America (Fig. 9.6b), which is the
9 Using Remote Sensing for Modeling and Monitoring Species Distributions
so, we used correlation analyses corrected according Clifford’s method to obtain the
effective degrees of freedom for Pearson’s coefficients while controlling for spatial
autocorrelation (Clifford et al. 1989). Statistical analyses were performed in R using
the package SpatialPack (Vallejos and Osorio 2014).
9.3.2 Results
Live oak models calibrated using different sources and combinations of environmental predictors (Table 9.1) within the ensemble framework generally provided
similar suitability distributions (Fig. 9.5). Interestingly, increasing the number of
predictors or increasing model complexity (S-SR2 in Table 9.1) affected model performance as measured by the Cohen’s Kappa coefficient and AUC (area under the
receiver operating characteristic curve) indices (Table 9.2), and thus affected the
geographic predictions: Complex models tended to have higher statistical performance but to underestimate the distributions of live oak species when compared
with simpler models (Fig. 9.5). Individual live oak species models made from S-RS
products and WorldClim differed somewhat in their performances (see Table 9.2
and Fig. 9.5). Models from WorldClim tended to have slightly better statistical performance in inferring species distributions based on the AUC and Kappa criteria.
However, these metrics do not capture differences in the precision and spatial resolution of the approaches. In several species, the WorldClim models predicted low
precision locations compared to the S-RS data. In particular, the IUCN (International
Union for Conservation of Nature) red-listed narrow endemic Brandegee Oak
(Quercus brandegeei) in southern Baja California is very imprecisely predicted
compared with the S-RS data. Using high-resolution interpolated climatic predictors did not improve the performance of individual models (WC25 in Table 9.2) and
returned similar suitability predictions to those estimated under lower spatial resolution climatic predictors (Fig. 9.5). Although WorldClim models seems to have
better statistical performance as shown in Table 9.2, we can at most discriminate the
accuracy of interpolating continuous surface-derived models, only when we are
inferring habitat suitability models (ENM) and not the projected species geographical distribution (SDM). Using S-RS data as predictors not only helps to identify the
species habitat suitability but also incorporates local ecological conditions necessary to predict local species distributions and co-occurrence (Radeloff et al. 2019).
This is because S-RS data have the potential to get at biological mechanisms, for
example, through the detection of species phenological variation over space and
time (Figs. 9.3c and 9.4).
When macroecological patterns of species richness and uncertainty maps were
constructed, we observed similar patterns of species richness between maps made
from the simpler combination of S-RS and WorldClim models (Fig. 9.6a, c, and d;
see Table 9.1 for a description of the environmental combinations of S-RS and
WorldClim). Notably, species richness estimation from the complex S-RS tends to
restrict live oak assemblages to southeastern North America (Fig. 9.6b), which is the
9 Using Remote Sensing for Modeling and Monitoring Species Distributions