Study site and species
The Haleakalā silversword, Argyroxyphium sandwicense subsp. macrocephalum (A. Gray) Meyrat, is a federally-listed threatened taxon in the family Asteraceae that occurs only on East Maui, Hawaiʻi. The Mauna Kea silversword (Argyroxyphium sandwicense subsp. sandwicense (A. Gray) Meyrat), is sister to the Haleakalā silversword, and is a federally-listed endangered subspecies growing at high elevations on Mauna Kea volcano, Hawaiʻi Island. Two additional silversword taxa grow on West Maui and Mauna Loa, Hawaiʻi Island. Unspecified references to silversword plants in this study refer to the Haleakalā subspecies.
The Haleakalā silversword is a long-lived (estimated 20–90 year, [50]), monocarpic, acaulescent rosette plant that today grows on the largely barren cinder cones, cinder flats, and rocky cliffs in a broad geographic area spanning the central to western portions of Haleakalā crater up to the summit, in the alpine zone from 2150 to 3050 m elevation (roughly 2300 ha). Long-term averages of annual rainfall across the silversword range are estimated to vary from 1018 to 1352 mm [51], with approximately 70 % of this falling in the wet season (November-April, [13]. Silversword distributions within the total range on upper Haleakalā volcano are clumped, with distinct aggregations often separated by large areas devoid of individuals. These aggregations were first comprehensively mapped and censused by H. Kobayashi in 1971 [32]. Following Kobayashi’s recommendations, a total population census was subsequently repeated approximately every decade (see below).
Early accounts suggest that the silversword population underwent a dramatic decline around the turn of the 19th century, owing to ungulate browsing and human vandalism [27]. Anecdotal descriptions by several visitors to Haleakalā crater in the 1800s imply a large abundance of plants, including the description of “thousands of silverswords … making the hillside look like winter or moonlight” in 1873 [52]. By the 1920s, however, serious concern arose among National Park staff and local residents, as it was apparent that the population had been decimated by feral goats and cattle and by zealous over-collection by people [27]. As before, no numbers were attached to these observations, with the exception of an estimate of “barely 100 plants” for the entire population made by a botanist in 1927 [53]. This was almost certainly a drastic underestimate, as a careful count of plants on one cinder cone, Ka Moa o Pele, in 1935 recorded 1470 plants [27]. This 1935 count, and a repeated count of the same cone in 1962, are the only reliable figures for silversword abundances prior to 1971. The 1935 count was previously used to estimate a total population of approximately 4000 in that year [28], by projecting the proportion of flowering plants on Ka Moa o Pele (88 of 1470) onto the entire population (in which 217 flowering plants were counted). In response to the aforementioned concern regarding silversword persistence, rigorous protections against plant collection were instituted by the National Park Service in the 1930s, and feral ungulates were controlled through hunting and were eventually completely excluded in the 1980s with a fence encircling the park [27].
Census procedures
Total population censuses were conducted in 1971, 1982, 1991, 2001 and 2013. In all censuses prior to 2013, one to three observers attempted to count all plants during a two to four week period. For this purpose, the silversword range was divided into 82 sectors that corresponded to a combination of known silversword aggregations and topographical features such as cinder cones or portions of cinder cones, and observers visited as many of these sectors as possible during each census. Total number of plants in each sector was estimated by counting all visible live plants with either the naked eye or with binoculars, depending on the observer’s distance. Observation distances are known to have varied tremendously, from several meters or tens of meters in flat regions that can be traversed on foot, to several hundreds of meters in regions situated on steep cinder cone faces or on distant cliffs. However, these observation distances and/or vantage points were not recorded. Methods of earlier censuses are also reported in Kobayashi [32] and Loope & Crivellone [27].
In the 2013 census, the procedure was modified to allow comparison with prior censuses while improving methods for future censuses. Four observers completed the census between October 2013 and March 2014, using the same methods described above. However, the count sectors were modified to more accurately represent current boundaries of silversword aggregations, rather than relying more heavily on topographic features, and the resulting 111 sectors were mapped in GIS (Geographic Information Systems) to allow more accurate relocation and delineation in future censuses (Additional file 1: Figure S1). In addition, observation distances were estimated for each region, either by estimating average distances to plants in regions that were walked through, or by using ArcGIS™ 10.2.2 (ESRI, Redlands, CA, USA) to measure the distances from observer vantage points to the midpoints of sectors in cases where binoculars were used. Finally, more effort was expended in the 2013 census, compared to prior censuses, to visit all previously reported locations supporting silverswords, making it the most thorough census to date.
Census comparisons over time
Because each census counted plants in a slightly different subset of sectors, and because these sectors were modified somewhat in 2013, we delineated 19 regions that were fully counted in all five censuses, grouping multiple sectors within each region, to enable comparison between censuses (Additional file 1: Figure S1). Sectors were generally grouped by cinder cone or contiguous lava flows. These 19 regions comprised from 86.7 to 99.6 % of the actual total counts in each census, with the lowest percentage (86.7 %) corresponding to 2013, the most spatially complete census. Therefore, the summed totals of the 19 regions in common for each census were subsequently divided by 0.867 to calculate an adjusted total count for each census, which accounted for areas missed and made the censuses directly comparable. This procedure assumes that the small minority of areas not counted in the 1971 to 2001 censuses fluctuated in the same manner as the remainder of the population.
Counts of plants on Ka Moa o Pele cinder cone were 1470 in 1935 and 2248 in 1962 [32]. This cone accounted for between 8.0 and 12.7 % of the five adjusted census totals (1971–2013). We therefore used this range to calculate a rough estimate of the lower and upper bounds for the total population in 1935 and 1962, by dividing the Ka Moa o Pele counts in those years by 0.127 and 0.080. We feel that this is likely to be a more accurate method for estimating the total population size than the one used previously, in which the proportion of plants flowering on the cone was projected to the entire population [28]. This is because proportions of plants flowering are known to vary considerably from region to region in a given year, much more so than the magnitude of variation observed in the proportion of plants on Ka Moa o Pele cone relative to the estimated population total over the five censuses.
Related work has suggested that silverword population declines over the past two decades have been most severe in lower elevation portions of the range [13]. To assess whether the decadal censuses have recorded a similar pattern, the 19 common geographic regions described above, which spanned an elevation of approximately 2200 m to 2700 m, were split into two even elevation zones (2200–2450 m, 2450–2700 m), and temporal trends between 1971 and 2013 were compared in each zone. We took this approach of lumping plant numbers across large geographic areas, rather than relating percent population change within each region to its elevation, because strong temporal variation and/or counting error in smaller geographic units led to very high and probably unrealistic variation in the resultant percent population changes, making spatial patterns difficult to detect.
Assessment of count error rate and true total population estimate
For 18 of the 111 count sectors in the 2013 census, we performed a second ‘true’ count, in which every live plant was individually approached and coordinates were recorded using a Garmin eTrex Legend H GPS unit (Garmin Ltd., Olathe, KS, USA). The ratios of the initial counts to the second true counts were used as estimates of count error rates for these sectors. Error rates were then regressed against the log-transformed observation distance and the log-transformed initial count (number of plants). These relationships were subsequently used to estimate count error rates (and 95 % prediction intervals) and to predict the true plant numbers for all 111 sectors in the 2013 census, based on the observation distances and number of plants estimated for each of these sectors. To avoid negative as well as unreasonably high predictions, we bounded the lower ends of the 95 % prediction intervals for the count error rates to 0.1, which seems reasonable given that the lowest measured count error rate was 0.32 (see Results). The resultant predicted true counts and 95 % intervals were summed to estimate the true total population size in 2013.
Influence of climate on decadal trends
We examined the explanatory power of patterns in rainfall, air temperature and the frequency of occurrence of the TWI on the magnitude and direction of decadal-scale population changes. For rainfall, we used a database of spatially-explicit (250 m grid cell resolution) hind-casted monthly rainfall estimates for the island of Maui from 1920 through 2012 [54]. We extracted all grid cells coinciding with the silversword distribution, and used these to calculate the average dry season (May-October), wet season (November-April) and annual (by water year, November-October) rainfall across the silversword range during each inter-census period. Rainfall data for 2012–2013 were taken from the average of six weather stations installed across the silversword range in 2010. Monthly mean air temperature data through 2010 were taken from the Haleakalā Ranger Station located at the park headquarters, 2125 m elevation (National Climatic Data Center). Air temperature data from 2010 to 2013 were taken from HaleNet station 151, also located at the Haleakalā NP headquarters [55]. We used the combined data to calculate average dry season, wet season, and annual air temperatures for each inter-census period. Frequency of occurrence of the TWI was calculated from atmospheric sounding data collected daily at 2 pm HST in Hilo, Hawaiʻi Island, from 1973 to 2013; these data are maintained by the University of Wyoming and can be accessed at http://weather.uwyo.edu/upperair/sounding/html. TWI presence was determined according to the methods in Longman et al. [22], and frequency of daily occurrence (%) was calculated for all dry seasons, wet seasons and water years possessing at least 75 % of daily readings. This resulted in TWI data gaps during the 1971–1982 inter-census period: for the dry season and wet season, respectively, 64 and 54 % of seasons had available data, while only 45 % of water years had available data during this period. We therefore chose to not use the water year data, and only calculated average dry season and wet season TWI frequency of occurrence for each inter-census period. There were no gaps in the TWI data after the 1970s.
We used the above inter-census climate averages as explanatory variables in simple linear regressions with percent decadal change in population size as the response variable. For the latter, we used the adjusted census totals, and standardized the percent population change between each census to ten year periods. Because there are only four inter-census periods, we added a fifth period, 1962–1971, to increase sample size, by using the average of the lower and upper bounds of the 1962 total population estimate (however, only rainfall and temperature data were available for this earlier period). Because non-linearity was suggested in some of the fitted relationships, we also attempted quadratic regressions against the rainfall and temperature data, but not against the TWI data which only had four data points. Due to the small sample sizes and intercorrelations, we did not fit multiple regressions using combinations of explanatory climate variables.
To test the influence of the TWI on rainfall patterns in Haleakalā silversword habitat, we regressed seasonal rainfall totals against TWI frequency of occurrence for all years with available data between 1973 and 2013 (n = 38 and 37 for dry and wet seasons, respectively); we also regressed the decadal averages of dry and wet season rainfall on decadal averages of TWI incidence (n = 4 for each). We further examined graphically how the reported shift in TWI frequency around 1990 may have differentially affected rainfall across elevation within silversword habitat. To do this, we plotted the mean estimated annual rainfall for the periods 1920–1990 and 1991–2012, for each 250 m grid cell within silversword habitat (n = 377), as a function of elevation. We also plotted the difference between these two time periods (pre- and post-1990) relative to elevation.