Conceived and designed the experiments: DA CK SL. Analyzed the data: DA CK SL BC SM. Contributed reagents/materials/analysis tools: SL BC KH RM. Wrote the paper: DA CK SL BC KH SM.
The authors have declared that no competing interests exist.
The flora of California, a global biodiversity hotspot, includes 2387 endemic plant taxa. With anticipated climate change, we project that up to 66% will experience >80% reductions in range size within a century. These results are comparable with other studies of fewer species or just samples of a region's endemics. Projected reductions depend on the magnitude of future emissions and on the ability of species to disperse from their current locations. California's varied terrain could cause species to move in very different directions, breaking up present-day floras. However, our projections also identify regions where species undergoing severe range reductions may persist. Protecting these potential future refugia and facilitating species dispersal will be essential to maintain biodiversity in the face of climate change.
The California Floristic Province has over 5500 native plant taxa; 40% of them are endemic, that is, their entire native distributions are within the Province
Empirical examples of species' range shifts resulting from climate change have been recorded for numerous taxa
A recent study of two California oak species projected significant range reductions for both species
Outside of North America, regional studies have addressed both range shifts and potential levels of extinction in the face of climate change. Studies of the Proteaceae in the Cape Floristic Province – another Mediterranean hotspot — estimate that this group may lose up to 20% of the species considered
Currently, there are no published assessments of potential impacts of climate change on regional endemic floras for any part of North America. California is particularly well suited to such a study, as it has high endemic plant diversity and the quality of plant distribution and climate data across the region are excellent. California also provides an interesting case study because of its topographic complexity, extensive urban and agricultural land use, and Mediterranean climate characterized by distinctive rainfall and temperature patterns.
We assess 8 different potential scenarios for the future of the California flora in the face of climate change. These are the combinations of three pairs of possibilities. First, we compared two projections of future emission levels from human activities. One is higher, with global CO2 emissions reaching almost 30 GtC per year, or 4 times present-day levels, by 2100 (SRES A1FI) while the other emission scenario is lower, with CO2 emissions rising slightly by mid-century before dropping to below present-day levels by the end of century (SRES B1)
Projecting the impacts of climate change to an entire endemic flora is complicated by scarce and variable distribution data. Studies conflict on how many geo-referenced specimens are necessary to obtain robust species projections
A recent study recommends using Maxent and at least 30 non-validation specimens for robust species projections
To assess whether excluding poorly known species biases diversity patterns, we build a multilevel generalized linear model (MLGLM)
To summarize the impacts of climate change on the California flora and to compare the projections with other studies, we ask four questions. First, where will endemic species diversity be most influenced by climate change? Second, if species are permitted to move, where will they go? Third, how do we project range sizes to change? Fourth, where do we expect future refugia — locations where species at risk from climate change will persist under future climates? To date no studies have mapped the locations of such refugia.
The California Floristic Province (
(A) The province divided into six floristic regions (solid lines): Northwestern California (NW), Central Western California (CW), Southwestern California (SW), the Cascade Ranges (CaR), the Great Central Valley (GV), and the Sierra Nevada (SN). The province includes most of California (dashed line) and portions of Oregon and Mexico. We include a surrounding buffer of equal area (colored areas outside solid line). Colors represent elevation in meters. (B) Projected present diversity. (C–J) Projected diversity 80 years from now modeled with increasing amounts of future climate change: (C–F) Plants cannot disperse. (G–J) Plants can disperse to all suitable areas. (C, F, G, H) Simulations based on the lower sensitivity PCM model. (E, F, I, J) Simulations based on the higher-sensitivity HadCM3 model. (C, E, G, I) Lower emissions scenario (B1). (D, F, H, J) Higher emissions scenario (A1FI).
We created diversity maps by summing modeled species distributions as is commonly done in Gap analysis
Based on these 591 species, we project present-day endemic diversity to peak at 340 species per km2, with the highest concentrations from southern Northwest California through most of Central Western California and in the foothills of the Sierra Nevada (
Our models yield projections of future diversity under a range of climate change scenarios (
Under the highest level of climate change examined here (mid-high climate sensitivity and higher emissions, as represented by HadCM3 A1FI projections), with the assumption of no dispersal, we project peak diversity to drop as low as 247 species per km2 (
Across all scenarios, the general trend is that diversity shifts towards the coast and northwards. Coastal areas, especially Northwestern California and Central Western California, are presently rich in species. Even under significant climate change, they will continue to be so. In contrast, the foothills of the northern Sierra Nevada are extremely vulnerable to species loss. Under scenarios that allow dispersal, the areas that straddle the California-Oregon border also become rich in species — as expected from northward dispersal.
The number specimens and range size derived from the TJM1 range maps were positively correlated (ρ = 0.49). Summed range maps for all 2387 endemic species indicate that species richness peaks at 621 species (
(B) Projected present diversity from the Multi-level Generalized Linear Model for all species with >2 specimens (2068). (C–J) Projected diversity 80 years from now modeled with increasingly increasing amounts of future climate change: (C–F) Plants cannot disperse. (G–J) Plants can disperse to all suitable areas. (C, F, G, H) Simulations based on the lower sensitivity PCM model. (E, F, I, J) Simulations based on the higher-sensitivity HadCM3 model. (C, E, G, I) Lower emissions scenario (B1). (D, F, H, J) Higher emissions scenario (A1FI).
Changing patterns of diversity projected from the multilevel model are very similar to the patterns of diversity projected from Maxent. In general, diversity shifts towards the coast and northwards, and the degree depends on the dispersal assumptions, emission scenarios, and the sensitivity of climate simulations. The following results on species movement and range size change are from Maxent projections of the 591 best known species.
Changes in diversity reflect the overall consequences of local extirpation and species dispersal. These patterns do not address the potential fate of individual species. For that reason, we also examined individual species fate in terms of projected geographic shifts in species' mean elevation, range centroid, and percent change in range size. In high emission scenarios (A1FI) with dispersal, we project species range centroids to shift by an average of up to 151 kilometers (see
As one might expect, species tend to move to higher elevations and often northward (see
(A) Two representative species that have adjacent present ranges (lighter colors) and are projected to move in opposing directions (arrows and darker colors). (B) Projected centroid movements for all species. Individual polar plots group species by the floristic region in which their centroid originates. Within each plot, species are grouped by the elevation in which their centroids originate. The magnitude of the directions represents the percentage of the regional flora moving in each direction.
The results shown here are for the largest projected changes in temperature (HadCM3, A1FI), allowing dispersal. We obtain similar patterns under lower projections of climate change and without dispersal (when species ranges can only shrink). In short, even relatively moderate projections suggest that climate change has the potential to break up local floras, resulting in new species mixes, with consequent novel patterns of competition and other biotic interactions.
As in previous studies in Europe and southern Africa, we project both reductions and increases in range sizes, depending on the degree of climate change and the abilities of the species to disperse
(A - D) Percent geometric mean change in range size (Future/Present with colors stretched from a <-10% decrease to a >10% increase). (E - H) Diversity of species gains (future diversity with migration minus future diversity without migration) for the quarter species suffering the largest range contractions. (A, B, E, F) Simulations based on the lower-sensitivity PCM model. (C, D, G, H) Simulations based on the higher-sensitivity HadCM3 model. (A, C, E, G) Lower emissions (B1). (B, D, F, H) Higher emissions (A1FI).
Green areas are dominated by species with expanding ranges. Red areas harbor shrinking species; they are climate change refugia for the species that a future generation of biodiversity managers may classify as “threatened”. In the future, the lower sensitivity simulations (PCM:
The red refugia in
The projections of diversity change are comparable with other studies from Africa and Europe
The magnitude of our range centroid shifts is similar to those reported for Eastern North American trees
The positive correlation between range map derived range size and number of museum specimens raises legitimate concern that excluding poorly known species may bias the results. From the comparison of the Maxent results from 591 species and the MLGLM results from 2068 species, we did not find the exclusion of these poorly known species to influence the general patterns of projected present and future biodiversity. These results suggest that the patterns of projected biodiversity presented here are robust despite the exclusion of poorly known species.
The bioclimatic models implemented in this study make a number of simplifying assumptions that may bias the projections
A key simplifying assumption is the “equilibrium postulate”
As described in the
These results present a sobering picture of the potential impacts of climate change on California's diverse and distinctive flora. The severity of projected impacts is closely linked to the magnitude of climate change. That, in turn, depends crucially on human emissions of greenhouse gases over the next few decades. The projected impacts are also very sensitive to the potential rate of plant movement, and rapid dispersal could mitigate much of the impact on individual species and overall diversity. However, rapid movement by natural dispersal is unlikely on a century time-scale, except for weedy species with short generation time and highly dispersable propagules. Human assisted dispersal must be considered as a critical component of conservation and biodiversity management in the next century.
The results of this study present a dilemma for conservation planning in the face of climate change. Future diversity will likely peak along the coast and to the north of its present concentrations (
We compiled geo-referenced specimens from the Consortium of California Herbaria
Additionally, we built range maps for each species from The Jepson Manual, 1st edition (TJM1)
We created four largely independent climate variables to represent present climate, derived from average monthly mean temperature and monthly total precipitation from the 1 km resolution DAYMET 1980–1998 mean climate database (
The four climate variables were the first two axes of two principal components analyses (PCA), one based on the 12 monthly mean temperatures and one on the 12 monthly precipitations, respectively (
For each of the 591 best known species, we used Maxent (version 2.3)
We used the test specimens to evaluate the performance of the Maxent projections using two widely used statistics that are recommended when evaluation absences are unavailable. The first was the area under the receiver operating characteristic curve
The second statistic was prediction success, the percentage of positive evaluation occurrences correctly classified as positive
Despite being statistically defensible, the chosen thresholds produced diversity maps that exceeded the diversity calculated from the TJM1 range maps. Since range maps are known to overestimate range size by over interpolating patchy species distributions
Unlike Maxent, generalized linear models require presence and absence data. To generate absence data for each of the 2068 species with at least 2 specimens, we generated a random (from 1 to 54) number of informed pseudo-absence data by randomly sampling points from outside the species' range map. See
The multi-level model has two levels: a flora level and an individual species level. At the flora level, the model estimates 9 parameter values for a data matrix consisting of an intercept, linear versions of the four climate variables, and quadratic versions of the four climate variables. Predicting
For each individual species, the model estimates random parameters for linear versions of the intercept and the four climate variables. The model estimates all parameters simultaneously, and the structure of the model allows poorly known species to draw strength from the rest of the flora. Effectively, this causes poorly known species to behave more like the average of the flora. The individual influence of error prone, poorly known species is thus appropriately weighted in diversity maps for the entire flora.
The future climate simulations are from the U.K. Meteorological Office Hadley Climate Centre Model version 3 (HadCM3)
The HadCM3 and PCM simulations project increases in mean annual temperatures averaged across the state of California of 2.3–2.2°C under B1 and 3.8–5.8°C under A1FI by 2070–2099. The models also project increases in the magnitude of seasonal temperature differences in most areas. Rainfall predictions are more variable among models. Changes range from decreases of 157 mm to increases of 38 mm of total annual precipitation. Within the United States, the global climate outputs were statistically downscaled to 1/8th-degree resolution
Histograms of the density of species centroid shifts in kilometers for each climate change scenario. (A–D) Scenarios in which species are permitted to move. (E–H) Scenarios in which species are not permitted to move. (A, B, E, F) Climate simulated by the PCM model. (C, D, G, H) Climate simulated by the HadCM3 model. (A,C,E,G) Scenarios with B1 emission levels. (B,D,F,H) Scenarios with A1FI emission levels.
(7.17 MB TIF)
Density histograms of mean elevation of species ranges in the present (blue) and future (red) for each climate change scenario. (A–D) Scenarios in which species are permitted to move. (E–H) Scenarios in which species are not permitted to move. (A, B, E, F) Climate simulated by the PCM model. (C, D, G, H) Climate simulated by the HadCM3 model. (A,C,E,G) Scenarios with B1 emission levels. (B,D,F,H) Scenarios with A1FI emission levels.
(3.89 MB TIF)
Directional histograms of species centroid movement for selected scenarios. Histograms are overlaid for different elevational zones, based on the species present elevation. The length of the vector in each direction is the percent of the corresponding flora that moves in that direction based on 591 species modeled with Maxent. (A) Climate simulated by the HadCM3 model with A1FI emission levels (severe scenario) where species are not permitted to move. (B) Climate simulated by the PCM model with B1 emission levels (less severe scenario) where species are not permitted to move. (C) Climate simulated by the PCM model with B1 emission levels (less severe scenario) where species are permitted to move.
(13.49 MB TIF)
Distributions of range size changes across all scenarios grouped by 6 range size change categories. (A–D) Scenarios in which species are permitted to move. (E–H) Scenarios in which species are not permitted to move. (A, B, E, F) Climate simulated by the PCM model. (C, D, G, H) Climate simulated by the HadCM3 model. (A,C,E,G) Scenarios with B1 emission levels. (B,D,F,H) Scenarios with A1FI emission levels.
(10.89 MB TIF)
(A–D) Current climate layers derived from PCA analyses and (E–H) corresponding climate variables. (A) Temperature magnitude (Axis 1 of a PCA of monthly mean temperature representing 69% of variation). (B) Temperature seasonality (Axis 2 of a PCA of monthly mean temperature representing 20% of variation). (C) Precipitation magnitude (Axis 1 of a PCA of monthly total precipitation representing 48% of variation). (D) Precipitation seasonality (Axis 2 of a PCA of monthly total precipitation representing 21% of variation). (E) Mean annual temperature ({degree sign}C). Correlation with Temperature Axis 1 is 1.000. (F) Standard Deviation of mean monthly temperatures ({degree sign}C). Correlation with Temperature Axis 2 is 0.998. (G) Total Annual Precipitation (cm). Correlation with Precipitation Axis 1 is 0.980. (H) Coefficient of variation of total monthly precipitation (cm). Correlation with Precipitation Axis 2 is 0.673.
(13.28 MB TIF)
Projected change in temperature magnitude (Axis 1, arbitrary units) (A–D) and temperature seasonality (Axis 2, arbitrary units) (E–H) under future climate change scenarios. (A, E) PCM B1. (B, F) PCM A1FI. (C, G) HadCM3 B1. (D, H) HadCM3 A1FI.
(19.98 MB DOC)
Projected change in precipitation magnitude (Axis 1, arbitrary units) (A–D) and precipitation seasonality (Axis 2, arbitrary units) (E–H) under future climate change scenarios. (A, E) PCM B1. (B, F) PCM A1FI. (C, G) HadCM3 B1. (D, H) HadCM3 A1FI.
(19.89 MB TIF)
Distribution data for each endemic species. The first two columns list the TJM1 range sizes in sq. kilometers and the number of specimens. The next two columns indicate whether the species was modeled with Maxent (>41 specimens) and included in the MLGLM (>1 specimens). The last column indicates the number of randomly selected informed pseudo-absences for use in the MLGLM.
(0.27 MB XLS)
Maxent model performance for the 591 best known species. The first two columns list the number of specimens used to test and train the models. The next column lists the area under the receiver operating characteristic curve (AUC) evaluation statistic which ranges from 0.5 to 1. The next column lists the threshold used to create binary ranges from the cumulative index ranging from 0 to 100. The last column lists the prediction success evaluation statistic which is the percent of test specimens correctly predicted by the binary ranges.
(0.10 MB XLS)
We thank all the authors of The Jepson Manual who compiled information on species ranges in California, P. Thornton for DAYMET assistance, and S. Pimm for mentorship. T. Dawson, J. Harte, M. Hanneman, M. Moritz, B. Mishler, B. Baldwin, and S. Stephens provided critical comments on an earlier draft.