Impacts of sea level rise and climate change on coastal plant species in the central California coast

Species occurrence data

Using the CalFlora Plant Database available from The CalFlora Database (http://www.calflora.org), we selected 88 species in the Tri-County Area (an area of 23,948 km²) of California, USA (Fig. 2) that were likely candidates for exposure to SLR, given their occurrence at low elevations (0–30 m) and represented in our study a variety of different taxonomic families, habitat types, life histories, and status (CalFlora, 2014). There were other species that also existed below our elevation criteria but were not chosen for a variety of reasons such as lack of data, limited distribution etc. The selected 88 species represent 31 different taxonomic families; 6 habitat types including coastal fresh and brackish marshes, coastal dunes, scrub, coastal bluffs, and meadows and grasslands; multiple life histories including annuals, herbs, succulents, woody, and deciduous shrubs; a variety of elevation ranges; and a mix of state and federally listed species, as well as unlisted but rare species (Table S1).

Species occurrence data were extracted from the ‘RareFind’ dataset of the California Natural Diversity Database (CNDDB) (http://www.dfg.ca.gov/biogeodata/cnddb/). The CNDDB maintains information about the natural history and locations of rare, threatened, endangered, and special status species and natural communities of California and has been used for a variety of species distribution models (Hernandez et al., 2006; Williams et al., 2009; Regan et al., 2012). In CNDDB, location data for a species takes the form of polygonal occurrences, which are a rough proxy for populations. An occurrence is defined as the area of a cluster of individuals within 1/4 mile of one another and separated by at least that distance from other occurrences. We excluded all occurrences recorded before 1970 and any that were greater than 4 km in diameter in order to minimize outdated and uncertain values. Due to incomplete and unknown data on a number of individuals present within each occurrence, we assumed that populations were distributed evenly across occurrences. Thus, we included occurrences regardless of the number of individuals or clusters of populations known to be extant within them. The 88 species accounted for a total of 1091 occurrences used in our analyses.

SLR projections

The SLR scenarios in this study were generated as part of the California Climate Impact Assessments which were produced from a downscaled global climate model (GCM) analyzed by the Scripps Institution of Oceanography (Cayan et al., 2009). The “high scenario” was a 1.4 m rise by 2100, while the “low” scenario was a 1.0 m rise by 2100 (Cayan et al., 2009). The coastal hazards of erosion and flooding associated with the impacts of the GCM outputs were projected for a variety of planning horizons using a total water level (tides + wave run-up) methodology (Revell et al., 2011). Coastal erosion model projections mapped all of San Luis Obispo County and most of Santa Barbara County, while the coastal flood extents were projected and mapped for the entire state of California. These projections of future coastal hazards were made available by the Pacific Institute, which conducted an initial statewide vulnerability assessment identifying critical infrastructure, habitats, and social demographics at risk from SLR (Heberger et al., 2011).

For coastal flooding, the mapped hazard extent was extrapolated from existing FEMA 100-year coastal Base Flood Elevations (BFEs), escalated by the projected amount of sea level rise. A 100-year flood is defined as a flood extent that has a 1% chance of being equaled or exceeded in a given year (FEMA, 2005). These BFEs, which calculated a maximum elevation of wave run-up at the shoreline, were mapped inland using a simple bathtub approach (FEMA, 2005). This approach likely overestimates the inland extent of coastal flooding, but in areas of combined fluvial and coastal flooding, may suitably represent the joint probability of a combined fluvial and coastal storm event (Revell et al., 2011). The coastal erosion hazards contained 3 components in the projected outputs: the effects of shoreline transgression from SLR, historic trends in shoreline change which provided an indirect accounting of sediment budget considerations, and the impact on erosion of a 100-year storm wave event (Revell et al., 2011). Inundation was mapped as the current extent of Mean High Water elevated by the SLR scenario over time by using a bathtub approach and ignoring hydraulic connectivity (Heberger et al., 2011). The resulting projections took the form of four general types of SLR related-threats (inundation, flooding, and cliff and dune erosion).

SLR threat analysis

In order to analyze the threat of SLR to each species, the occurrences for the 88 species were combined with the above four SLR threat layers for the year 2100, including inundation, flooding, and cliff and dune erosion in the Tri-County Area. We compared the geographic area of the occurrence data with the geographic area of the SLR threat layers to determine the area of overlap. We used the area of overlap to calculate the percent of each occurrence exposed to SLR for each species. We examined the area of exposure by aggregating the geographic areas of the four SLR-related threats to determine where any threat might occur.

SLR risk analysis

In order to determine the best predictors of exposure to SLR for our 88 species, we gathered a variety of physical, spatial, and biological plant characteristics related to each species, including life history, federal and California listing status, as well as each occurrence’s area, elevation, and distance from the coast (see Table S1). These variables included both continuous (e.g., elevation, distance) and categorical (e.g., life history, listing status) data. The continuous variables all had occurrence-level specificity, whereas the categorical variables only had species-level specificity. We ran multiple logistic regressions using R 2.15.1 (R Development Core Team, 2012), to determine which variables (including interactions) resulted in the best predictive models for exposure to SLR. We selected the best model based on two measures: the lowest Akaike Information Criterion (AIC) value (Akaike, 1973; Bozdogan, 1987) and statistically significant coefficients.

Species distribution modeling

We modeled current and future habitat suitability using MaxEnt version 3.3.3k (Phillips, Anderson & Schapire, 2006), a machine-learning technique often used to model the spatial distribution of a species using environmental variables and species’ occurrence data (Gogol-Prokurat, 2011). Our species provides presence-only data. Although many SDMs require both presence and absence data to predict distributions, MaxEnt has been recognized to be particularly effective with presence only data (Phillips, Anderson & Schapire, 2006; Regan et al., 2012). Moreover, MaxEnt can partially compensate for incomplete and small data sets on species occurrence and perform with nearly maximal accuracy level under these conditions (Hernandez et al., 2006). This is ideal for rare species that typically have small populations.

Based on the results of the SLR Risk Analysis, we identified the 10 species that were most likely to be substantially impacted by SLR in the Tri-County Area. These were Centromadia parryi ssp. australis, Chloropyron maritimum ssp. maritimum, Cirsium rhothophilum, Dithyrea maritima, Erigeron blochmaniae, Lasthenia glabrata ssp. coulteri, Monardella crispa, Monardella frutescens, Scrophularia atrata, and Suaeda californica. We examined the effect of climate change on each species by modeling current and future habitat suitability in MaxEnt, based on current location data calculated from centroid of species occurrence polygons in California and six environmental inputs consisting of four bioclimatic (i.e., Mean Diurnal Range; Annual Precipitation; Precipitation in the Wettest Quarter; Growing degree days above 5 C) and two edaphic variables (i.e., Soil pH; and Available Water Holding Capacity). These environmental inputs have been used previously to model plant species distributions (Fitzpatrick et al., 2008; Riordan & Rundel, 2009; O’Donnell et al., 2012; Sheppard, 2013) because these variables were general factors influencing the distribution of a wide range of plant taxa (Woodward, 1987). The inclusion of soil characteristics has also been known to improve SDM performance when assessing climate change impacts (Austin & Van Niel, 2011) and has been used in various SDM studies (Syphard & Franklin, 2009; Regan et al., 2012; Belgacem & Louhaichi, 2013; Conlisk et al., 2013).

Historical climate was obtained from the Parameter-Elevation Regressions on Independent Slopes Model (PRISM) at Oregon State University, a method for extrapolating the measured historical data (Daly et al., 2002). Due to the large variability in long-range climatic predictions for 2100, we selected two GCMs: the Parallel Climate Model (PCM) (Washington et al., 2000) and the Geophysical Fluid Dynamics Lab (GFDL) (Delworth et al., 2006; Knutson et al., 2006) model, both used by the State of California for assessing climate change impacts because they produce accurate simulations of California’s recent historical climate but show different levels of sensitivity to greenhouse gas forcing (Cayan et al., 2008b). As all GCMs, GFDL and PCM project warmer conditions for southern California by the end of the 21st century, but PCM projects a more modest annual temperature increase (2.5 °C for PCM vs. 4.4 °C for GFDL) and winter precipitation change (+8% for PCM vs. −26% for GFDL) while the GFDL projects a generally drier future based on the IPCC’s A2 emissions scenario (i.e., business-as-usual) (Regan et al., 2012). We used downscaled monthly climate data from the two GCMs and PRISM (historical climate) at a grid size of 90-m resolution (Flint & Flint, 2012), and then calculated bioclimatic parameters based on the methods described in Sork et al. (2010) for Growing Degree Days and used the WORLDCLIM database (www.worldclim.org) for other bioclimatic parameters. The time horizon for this data is centered on 2085, as opposed to 2100, though it represents an end-of-century 30-year average with 2085 being the median (Flint & Flint, 2012).

We calibrated the MaxEnt model using the default value settings suggested by Phillips, Anderson & Schapire (2006). We set the random test percentage to 33%, which retains a percentage of the occurrences at random in order to evaluate the model and the rest of the occurrences were used to build the final models. We ran 10 replicate runs and averaged the results. We evaluated our models under the current climate by using the area underneath the receiver operating curve statistic (AUC) (Phillips, Anderson & Schapire, 2006). The AUC produces a single number between 0 and 1, where a higher AUC indicates a better model fit (Fielding & Bell, 1997; Giannini et al., 2012).

MaxEnt outputs are continuous probability layers for species occurrence under: (i) the historical climate with the PRISM climate model; and (ii) the two future projected climates with PCM and GFDL climate models. We converted the continuous probability maps from MaxEnt into binary presence/absence layers using a threshold value that minimizes the sum of sensitivity and specificity of the model (Jiménez-Valverde & Lobo, 2006). We removed current urban areas, which we deemed as unsuitable areas, from each of the three binary layers. However, these urban areas may not have included specific features such as seawalls that might have a direct impact on the future suitability of habitats. We then calculated the area of presence data to compare the relative gain or loss in habitat between the current and future scenarios allowing us to quantitatively compare the habitat change from impacts of climate change with SLR.

Evaluating relative impacts of SLR and climate change

Using the GIS layers of modeled current and future suitable habitat, along with the layer of projected SLR impact, we calculated several intermediate variables to quantify the relative effects of, and interactions between, SLR and climate change in determining changes in potential habitat for each species. Figure 3 illustrates a conceptual model of how the change in suitable habitat was affected by SLR and climate change. The direct effect of climate changes in air temperature and precipitation (C) is the change in suitable habitat area disregarding SLR, estimated as the difference in areas between the projected suitable habitat layer under one of the future climate projections (F) and that for historical climate (P) (Fig. 3A): (1)${\displaystyle C=F-P.}$ Specifically the area projected to be lost from the current suitable habitat is identified as C⁻ and the area gained from future suitable habitat is C⁺.

The direct impact of SLR (S) is the reduction in current suitable habitat area caused by SLR: (2)${\displaystyle S=P_S-P,}$ where P_s is the portion of the projected habitat layer based on the historical climate that does not overlap projected SLR (Fig. 3B). The total change in suitable habitat area (H) combines the effects of SLR and climate change: (3)${\displaystyle H=F_S-P,}$ where F_s is the portion of the projected habitat layer based on future climate that does not overlap projected SLR (Fig. 3C). To further specify the direction of loss or gain in the change in habitat, H⁺ represents a gain in habitat while H⁻ represents the loss in habitat.

We write the combined impact of climate change and SLR as contributions from their direct effects and any spatial interactions (I) between the two: (4)${\displaystyle H=C+S+I,}$ where (5)${\displaystyle I=H-\left(C+S\right).}$ This interaction can be positive (Fig. 1A), negative (Fig. 1B) or zero (Fig. 1C).

We also calculated the proportional impact of SLR on habitat area under the future climate, A: (6)${\displaystyle A=1-\left(\frac{F_S}{F}\right).}$

SLR effects on current occurrences

We found that under the SLR projections for the year 2100, 17% of the 1091 occurrences of all species in our analysis would be affected by SLR, with a total of 10.6% threatened by routine inundation, 15.6% by a 100-year coastal flood, 5.9% by dune erosion, and 4.6% by cliff erosion. On the species level, we found that 65% of the 88 studied species are projected to have at least one occurrence impacted by SLR, with 12% of species having all of their occurrences within the SLR hazard zones (Fig. 4). However, nearly two thirds (63%) of the species are projected to have less than 20% of their occurrences at risk. The risk profile of the remaining species is fairly uniformly distributed between 20% and 100% (Fig. 4). Among all SLR threats, the threat profile from flooding alone closely mirrors the aggregate SLR threat profile. By contrast, inundation, dune erosion, and cliff erosion, are projected to affect almost 50% of species, with less than 5% of the species having all occurrences in the hazard zone (Fig. 5).

SLR risk as a function of elevation and distance

The best-fitted logistic regression model to explain the SLR exposure of species occurrences incorporated occurrence area, elevation, and distance from the coast (Table 1). None of the species-level variables (life history and listing status) were significant predictors of exposure to SLR (Table S1). Adding interaction terms did not improve the model. SLR threat to a species occurrence increases with occurrence area but decreases with elevation and distance from the coast (Table 1 and Fig. 6). Occurrences that are within 0.25 km of the coast and below 0.1 km in elevation are predicted to have a 100% chance of exposure to SLR.

The probability of exposure to inundation and flooding is qualitatively similar to that for the aggregate threat, with risk from flooding extending further inland than inundation (Table 2 and Fig. 7). In contrast, exposure to dune and cliff erosion depends only on distance from coast and occurrence area, but not elevation (Table 2 and Fig. 7).

Effects of climate change and SLR on habitat (species distribution modeling)

All runs for our 10 species produced consistently high AUC values greater than 0.95, indicating that MaxEnt modeled and predicted the current distribution of species effectively. Four species (Cirsium rhothophilium, Erigeron blochmaniae, Monardella crispa, and Monardella frutescens) were projected to have no habitat left in the study region under both the PCM and GFDL future climate models.

Under the GFDL climate model, four species (C. maritimum ssp. maritimum, C. parryi ssp. parryi, D. maritimum, and L. glabrata ssp. coulteri) are projected to significantly expand habitats with minimal loss to current modeled habitat (Fig. 8). With SLR, only C. maritimum ssp. maritimum loses as much as 40% of the current habitat. S. atrata is projected to have only a very small amount of future suitable habitat, and this habitat does not overlap with the current habitat projected for this species. S. californica is projected to maintain about 25% of its current habitat under the GFDL model, with a very modest habitat expansion into new areas and no significant losses to SLR.

The PCM climate model primarily projects a contraction in future habitat (Fig. 9). Only two species are projected to gain significant habitat under the PCM climate model; L. glabrata ssp. coulteri will gain extensive suitable habitat (+339% habitat relative to current habitat) and C. parryi ssp. parryi will gain some new suitable habitat (+65% habitat relative to current habitat) All species, except L. glabrata ssp. coulteri, maintain less than 45% of their current habitat under the PCM future climate model, with notable losses from SLR for C. maritiumum ssp. maritimum (Fig. 9).

The total loss of current habitat due to SLR is projected to be similar across species (Table 3). In contrast, the projected changes in habitat resulting from climate change are much more variable across species and climate models. In terms of the area of habitat lost, the impact of SLR can be as much as half the magnitude of the projected impact of climate change (C. maritimum under PCM), but is generally a much smaller component of future habitat change (as little as 0.1%). Comparing the percent area lost due to SLR for the current and future climate models reveals that the proportional impact of SLR is generally less in the future than at present (the exceptions are D. maritima and S. californica under the PCM climate model). Additionally, the interaction between SLR and climate change is on the same order of magnitude as the effect of SLR alone (Table 3).