Efficiency in assessment and monitoring methods : scaling down gradient-directed transects

Rapid survey methods are needed that accurately detect both species richness and relative abundance for surveying new sites and/or for long-term monitoring. We investigate whether the large-scale gradsect survey method of Gillison can be adapted for a smaller scale more suitable for monitoring or research. Three different designs that are compared are typical random small plot sampling and two transect designs that are modifications of gradient-directed transects (gradsects). We did intensive work at one site, and surveyed another eight sites. In contrast to most methods comparison work, we develop a baseline reference dataset for ‘true species richness and relative abundance’ by using more than 400 randomly-located small plots. Monte Carlo randomizations determined the minimum sample number for each type of sampling method for thresholds of species richness and abundance. The modified gradientdirected transects were accurate for both richness and abundance, even for uncommon species, and with much greater efficiency than random sampling.


INTRODUCTION
As climate change accelerates and anthropogenic impacts broaden (Richardson et al. 2009), assessment and monitoring methods that efficiently capture both species richness and quantitative measures of abundance are urgently needed.A variety of sampling methods are currently used to assess species composition or monitor sites for restoration, invasive species or other conservation issues.Considerable research has evaluated sampling effectiveness under different conditions and habitat types (e.g., Floyd and Anderson 1987, Bra ˚kenhielm and Qinghong 1995, Carlsson et al. 2005, Godı ´nez-Alvarez et al. 2009), but few have incorporated comparisons of time expenditure.We are interested in methods that are both rapid and accurately determine species richness and relative abundance.
One method, gradient-directed transects ('gradsects'; Gillison and Brewer 1985), seems adaptable to a variety of environments and conditions and has been modified for both plant and animal studies (e.g., Wessels et al. 1998, Sandmann andLertzman 2003).This method was originally developed for large-scale resource surveys (Gillison 1983, Gillison andBrewer 1985) and assumes that the steepest environmental gradients structure the greatest biological diversity in any region and orients transects along these gradients.This approach has been modified for different questions and tested for a number of different vegetation types or key animal groups (e.g., Austin andHeyligers 1989, Wessels et al. 1998).In most of these cases, the scale of the tested system was large, on the order of 100-10,000 km 2 (Gillison and Brewer 1985, Ludwig and Tongway 1995, Sandmann and Lertzman 2003).Environmental gradients, however, arise at different spatial scales and the spatial extent of many assessment or restoration sites are at substantially smaller scales.Here we test whether the technique can be modified to encompass a smaller scale and still capture species richness and relative abundance, both accurately and efficiently.
The effectiveness of gradsects in a variety of systems encouraged us to modify the method and test the hypothesis that gradient-directed transects are equally effective at smaller scales.A common problem in site assessments or in developing monitoring programs is determining an effective sampling scheme that reliably will document species richness and abundance.The design phases of monitoring programs often do not distinguish between sampling method and design, and yet comparison among alternate designs may yield information on the benefits and limitations of different approaches (Stohlgren 2007).Because monitoring labor costs can be expensive, monitoring programs are often minimal without regard to the power of the sampling design (Burbridge 1991, Legg andLaszlo 2006).
Given these issues, we tested the accuracy and time efficiency of two modified gradsect sampling designs against randomized plot sampling for measuring plant species richness and cover.For a test system, we selected tidal wetlands because of their clear environmental gradients along a relatively small spatial scale and because tidal systems are at substantial risk due to climate change and accelerated sea-level rise (Vermeer and Rahmstorf 2009).Wetland species are restricted to specific inundation and salinity regimes.Tidal wetlands are characterized by channels through which water moves during tidal cycles, modifying the physical environment of the adjacent wetland areas and affecting species composition and productivity in a gradi-ent with distance from channel edges toward the interior of the wetland (e.g., Zedler et al. 1999, Sanderson et al. 2000).Because tidal channels create a spatial gradient with respect to species distributions, we sought to compare methods directed specifically on the spatial gradient created by channels against a random plot method.We asked the following questions: (1) how accurate are different sampling designs; (2) which method is most time efficient at reaching thresholds of species richness and cover; and (3) do designs that take advantage of spatial gradients increase efficiency while maintaining accuracy?

Study location
We conducted vegetation surveys at nine tidal wetlands throughout the San Francisco (SF) Bay-Delta estuary, California, USA (hereafter referred to as 'Bay-Delta'; Fig. 1, Table 1); these sites vary in species richness (12-119 species) and vegetation complexity (0.4-3.5 m height).We conducted intensive work using three sampling designs on Coon Island, a brackish marsh; most of the results reflect data collected here.We selected Coon Island for intensive work because it is intermediate in species richness among wetlands in the region, and is among the larger in spatial extent.To compare their efficiency and accuracy in other wetlands, we used the three sampling designs at five additional wetlands, and only two at three other wetlands for which we lacked aerial photos, including a mix of salt, brackish and freshwater wetlands, as well as both natural and restored wetlands (Fig. 1, Table 1).

Random point vegetation surveys
Orthorectified color infrared (CIR) imagery was collected in 2003 over all sites except for China Camp, Palo Alto Baylands, and Petaluma River marsh, and the images were used to generate random points in the wetlands in a Geographic Information System (GIS).Sampling points were located in the field using Garmin GPSMap 76 units (horizontal accuracy: 3-7 m).At Coon Island, we visited 206 random points in 2003 and 219 in 2004.At each random point, we surveyed a 3-m diameter circular plot, recording presence and cover of all species.Cover was  species cover; we used the 'true' richness and cover data to determine accuracy of both the belt and parallel transects as well as subsets of random plots.

Transect vegetation surveys
We tested two transect-based sampling designs that were oriented from tidal channel edges: a continuous belt transect and a set of parallel transects (Fig. 2).To determine the initial transect locations, channels were manually delineated using CIR imagery, and random points were generated that were constrained to be within 5 m of channel edges.Transects were initiated at the closest channel edge to the random point.We used both transect sampling designs at all nine wetlands in 2004, simultaneous with the second set of random point plots described above.
Ten random channel edge locations were visited at each marsh.For the continuous belt transect (hereafter referred to as 'belt transect'), a 40-m tape was extended perpendicularly from the channel edge, and plant presence and cover were recorded in 40 sequential 1 3 2-m plots to the one side of the tape.A transect length of 40 m was chosen because this reflects the average distance to the mid-point between channels at most sites.At the same starting location as the belt transects, five 10 3 1-m transects oriented parallel to the channel (hereafter referred to as 'parallel transect') were established at 0, 10, 20, 30, and 40 m from the channel, oriented along, but on the other side of, the same transect tape as the belt transect (Fig. 2).Along each parallel transect, three 1 3 2-m plots were randomly selected from five potential plots, and species presence and cover were recorded.The distances between parallel transects were selected to sample species transitions based on our experience in these tidal wetlands.Distances would have to be modified for other vegetation types based on the steepness of gradients.

Calculating observer time
At all of the sites but one, the amount of time required and the number of people involved in data collection were also recorded.We calculated time two ways: (1) as the total number of minutes required per plot, regardless of the number of people involved, calculated as total time divided by number of plots for that day; and (2) as the total number of minutes per plot, times the number of people involved.Data for each field day were kept separate, and the time required for each random plot or for each transect was determined every day.The average of those field days was used to calculate time efficiency for each method.Teams generally only used a single person to estimate cover and multi-person teams were spread evenly among methods as much as possible.

Randomization
We analyzed data from a wetland intermediate in size and species richness, Coon Island, for thresholds of species richness and percent cover using a Monte Carlo resampling approach.Plots were resampled without replacement for the random dataset, and individual transects for the two transect methods.Either a plot-by-species matrix or a transect-by-species matrix was created with either presence/absence or cover class data in each cell; these matrices were used in the randomizations.For species richness randomizations using the presence/absence matrices, the order of individual plots or transects was randomized and sampled in sequence.The number of plots or transects required to reach different species richness thresholds was assessed.We determined species richness thresholds for 5, 10, 15 and 20 species at Coon Island, which had a total of 20 wetland species.This process was repeated sequentially for 100 randomization trials.For the cover randomizations, the average cover based on the total number of plots (425 plots from 2003-2004 dataset) was used as the threshold; the number of plots required to reach the average cover value, and then subsequently stay within 610% of the average, was used as the final value.We selected 10% as the threshold range because it is commonly recommended (McCune and Grace 2002).Randomizations for cover on the transect data used all ten samples, and cover means for 100 randomizations for 1, 5 and 9 transects were compared to that of all ten transect samples as well as to the 'true' values based on the 425 random plots.To assess performance differences among the methods, one-way ANOVA followed by Tukey-Kramer HSD tests were performed on the randomization data for area sampled and time required for all richness thresholds using JMP 8.0.

Species richness
Fewer transects were required to reach species richness thresholds than small random plots (Fig. 3A).Both transect methods averaged the 20species threshold at slightly over 8 transects each (belt, 8.3 6 0.16 SE; parallel, 8.46 6 0.14), while an average of almost 72 random plots (71.66 6 2.75) were required to reach this threshold.For all three methods, substantial increases in sampling number were required to reach each richness threshold.
However, each method assessed different areas (a single, circular 3-m diameter plot '7 m 2 , a continuous belt transect is 80 m 2 , and 15 plots v www.esajournals.orgfrom the parallel transect sum to 30 m 2 ).Consequently, different total areas were sampled for each richness threshold (one way ANOVA F 2, 297 ¼ 366.90 at five, 161.64 at ten, 213.74 at fifteen, and 234.78 at the twenty species richness threshold, all P , 0.0001 for each richness level).At lower richness thresholds, the belt transects sampled significantly greater areas.At the highest richness threshold, the parallel transect method was the most efficient in terms of area sampled, and the belt transect and random plot method were roughly equal although statistically different (Fig. 3B) (all means statistically different in a post-hoc Tukey-Kramer HSD test).
After converting data collection to time expended, a different pattern developed among the methods.As with total area sampled, the random plot method and parallel transects were the most time-efficient at the lower richness thresholds but the random plot method became the least efficient method for the higher thresholds (Fig. 3C) (one way ANOVA, F 2, 297 ¼ 56.22 at five, 3.68 at ten, 45.13 at fifteen, and 221.21 at the twenty species richness threshold, P , 0.0001 for five, fifteen, and twenty; P , 0.0268 at ten).To reach the 20-species threshold, or all the wetland species on the 162.4 ha Coon Island, random plots required 711.01 6 272.61 person-minutes.The belt method required only 44% of that time, and the parallel transect method only 38% (Fig. 3C).With respect to time efficiency, at the 20species threshold, the two transect methods were not statistically different from each other in a post-hoc Tukey-Kramer HSD test.We also calculated time efficiency incorporating the number of people involved, and while the number of person-minutes increased considerably, the results were proportionately similar to the values calculated without regard to the number of people (data not shown); this is likely due to the fact that only one person per multi-person crew estimated cover while the others recorded and helped with setting up transects.In addition, multi-person crews were equally distributed among the three different methods.
At the eight additional wetlands, the effectiveness of each method varied among sites.At the five sites in which random plots were taken, the transect methods again were more efficient in time required (data not shown).Comparing the two transect methods, the belt transect captured more species in five of the sites, was equal to the parallel transect in two sites, and the parallel transect captured more species in two sites (Table 2).The parallel transect method was more time efficient at seven of eight sites and similar to the belt method at the other site (Table 2).In tidal freshwater sites with diverse, tall (3 mþ), dense, vegetation, transect methods were much more accurate in determining species richness than the random plot method.

Cover
The total plant species cover at Coon Island was similar between the two years assessed for the random plot data and those data were combined to represent the true cover values of species in the wetland (425 plots total) (Table 3).The ten belt or parallel transects provided a close match for the species cover compared to the 'true' cover values based on the complete random plot dataset.While there was a tendency to slightly overestimate or underestimate a few species using both of the transect methods, an important outcome was how relatively accurate these methods were with respect to less common or relatively rare species.This was an unexpected outcome given the small number of transects used and the sometimes patchy distribution of some of the species.
Based on randomizations of the Coon Island random plot percent cover data, more plots were required for species with lower cover (Fig. 4).The values in this figure represent the minimum number of plots needed to reach the percent cover threshold in the 'complete' dataset and maintain that value 610%.Less common or rare species always require more random sampling plots for accurate means (Grieg-Smith 1983), and a negative exponential relationship existed between mean species cover and the number of random plots required for this site (y ¼ 320e À0.026x , R 2 ¼ 0.94).In contrast, randomizations of the transect data demonstrated a rapid stabilization of cover values regardless of whether species were abundant or uncommon (Fig. 5).

DISCUSSION
Scaled-down modifications of the gradsect method can be an effective rapid assessment and monitoring methodology.While we focused on tidal wetlands, the methodology can be modified to suit any vegetation or habitat if scaled to the local environmental gradient.The general principle of sampling along environmental gradients is scale-free (p.123, Gillison and Brewer 1985), but the application of the gradsect methodology has generally been restricted to or suggested for rather large spatial extents (Bullock 2006).Our two gradient-directed transect methods were clearly more efficient than the random plot sampling method, both in terms of the number of samples required to be equally accurate and in sampling time.In much less time Fig. 4. Relationship between the mean cover of dominant species and the minimum number of small random plots required for the Coon Island (focal site) dataset.The x-axis values are based on randomizations.Minimum number of plots was determined by how many plots were required before the species average cover remained within 10% of the average species cover value for the whole dataset.Y-axis values represent the mean of all 217 plots.The data were calculated to fit y ¼ 320.27eÀ0.026x with R 2 ¼ 0.94.v www.esajournals.organd in one case with significantly less area sampled, the modified gradsects were able to accurately assess both species richness and relative abundance (Fig. 3, Table 3).
A critical unexpected finding was that gradsects also were able to reasonably assess relative abundance of less common species as well as dominants in contrast to the random plot design (Figs. 4,5).While random plots can assess species richness with some efficiency at a relatively small number of plots (averaging 72 in our randomizations), a much greater number of plots was required to similarly estimate relative cover (Fig. 4).In random plot designs, the number of plots needed to reliably estimate cover increases exponentially from the most dominant species for each subsequent taxon as cover values declined (Grieg-Smith 1983), limiting its effectiveness for rare or less common species.
Scale-modified gradsects should be successful in any habitat because virtually all habitats are characterized by environmental gradients.The effectiveness of the original gradient-directed transect method is based on vegetation reflecting underlying environmental gradients (Gillison and Brewer 1985).Species distributions in our study wetlands are arrayed along strong tidal gradients of inundation, soil aeration and salinity, thus meeting this environmental gradient assumption.Wetlands of any type generally exhibit elevation gradients and sampling along those gradients should be similarly effective (e.g., Galatowitsch and van der Valk 1996, Edwards and Proffitt 2003, Leck 2003).A principal critique of gradsect methodology is the concern that the primary environmental gradient will not be selected (e.g., Bullock 2006).Most habitats with significant topography will usually display recognizable gradients.For example, an analogous situation to wetlands occurs in rocky intertidal zones, in which an environmental gradient is created by tides, exposure, and temperature gradients; transects along elevation gradients consequently should be, and turn out to be, more efficient than random plots in accurately Fig. 5. Percent cover results of randomization runs for wetland species using belt transects.Results are shown for belt transects (parallel not shown) for 4 representative species differing in cover and frequency, Sarcocornia pacifica, Grindelia stricta var.angustifolia, Baccharis pilularis and Chenopodium californicum.Three types of random sampling (with replacement) results are shown for each species, the cumulative average for one belt (dotted line); the cumulative average for five belts (dashed line), and the cumulative average for nine belts (solid line).In the results for five belts, for example, in each run, 5 of the 10 belts were randomly selected and % cover averaged for the five belts.The x-axis represents cumulatively averaging the results of separate randomization runs.v www.esajournals.orgestimating species presence and cover (Dethier et al. 1983, Andrew andMapstone 1987).Further, in an intertidal study that utilized Monte Carlo simulations, transects along an elevational gradient were far more efficient than any other design (Miller and Ambrose 2000).
To capture both species richness and abundance accurately, the methods used in this study are not without critique.One critical design consideration is that the length of transects should be adjusted to the length of environmental gradients in the vegetation or habitat sampled.In other words, transects have to be long enough to capture all phases of the mosaic or gradient patterns (Grieg-Smith 1983).In our study, for most of the tidal marshes sampled, the transect length reached the mid-point between channels.However, in the upper Bay-Delta, oligohaline and freshwater wetlands have much larger and fewer channels than salt marshes due to differences in tidal flow and amplitude.Following data collection, we found that transect lengths needed to be increased in those sites to capture the proportionate patterns in species abundance because we under-sampled higher elevation areas.Similarly, the number of transects we used was also arbitrary as we considered this an exploratory study; in lower species richness sites, regardless of the spatial area, this small transect number was effective.At the more species rich sites, not all species were found and more transects would have been more effective; the optimal number could be estimated by traditional richness to number of sample plots.Finally, we spaced the parallel transects at equal distances because those distances accommodated the changes in environmental gradients in these wetlands.In habitats with a change in life form or slope, rates of change in environmental gradients would similarly change; distances between parallel transects should reflect the rates of change in gradients.
Of the two transect methods, the parallel transect method was almost always the most efficient with respect to time required, principally because it sampled less than half the area of the belt transect.The parallel transect method, however, more frequently missed species and thus sometimes yielded a lower species richness.Our parallel distances were arbitrary even though they were appropriate for some of the wetlands; in others they gave misleading results when the transect spacing interacted with the spacing of species shifts or mosaics (Dale 1999).In contrast, the belt transect method, like random plots, requires no prior knowledge of species turnover patterns but does require sampling greater areas.Consequently, the belt method is more accurate in sampling species richness.While the difference between the two transect methods is small, and each has advantages, the parallel transect method requires prior surveys to customize the spacing among the parallel transects.Otherwise, it may result in less reliability in estimating species presence and overall species richness (Dale 1999).Both methods require adjusting transect length to assess proportionate species abundances.With respect to time expended, at sites with vegetation that was tall, dense and difficult to move through, both transect methods were much more efficient.
For most conservation applications, e.g., the rapid assessment of the conservation value of new sites, the recovery of sites after restoration (e.g., Callaway et al. 2007), species responses to the impacts of climate change or invasive species, estimates of both species richness and abundance are critical measures.Some have argued for restricting plant inventories to presence-absence lists (e.g., Hintermann et al. 2002).However, abundance is a critical measure in these circumstances because species lists are subject to error (e.g., Vittoz and Guisan 2007), and pseudoturnover is common in studies of observer variation (Fischer andSto ¨cklin 1997, Vittoz andGuisan 2007).Also, species colonize at different rates following restoration (Tuxen et al. 2008), disperse or invade at different rates (e.g., Diggory and Parker 2011), and vary in sensitivity to different aspects of climate change (Parker et al. 2011) and abundance measures tend to indicate conservation issues more accurately in those conditions.
In summary, gradsect sampling represents an important survey and assessment method at scales typical for most conservation or restoration work.They would also be effective for detecting change along ecological gradients of different spatial and temporal scales.Long-term monitoring of vegetation or focal animal populations is a critical aspect of understanding directional species change due to species invasion, restoration or climate change.In this study, abundance of species in a small number of transects was able to match a large random dataset for both dominant and uncommon species (Table 3, Fig. 5) and would be a more sensitive monitoring measurement than presence-absence alone.Transects have the advantage of speed and as long as their length appropriately adjusted, they should accurately capture both species diversity and relative abundance (Grieg-Smith 1983).Because costs are always an issue with long-term monitoring (Strayer et al. 1986, Legg andLaszlo 2006), these transect methods also have the advantages of accuracy with considerably less labor investment (Fig. 3).Some have argued that a targeted approach to monitoring is more likely to detect ecosystem changes over time, whether predicted or unanticipated (Stohlgren 2007).The application of gradsects at smaller scales represents a methodology that could well be applied to this targeted approach.

Fig. 1 .
Fig. 1.Site map of the San Francisco Bay Estuary, California, USA, marshes where vegetation surveys were conducted.

Fig. 2 .
Fig. 2. Diagram of the sampling design for the two transect techniques.For the parallel transect method, small transects were established parallel to the channel at 0, 10, 20, 30, and 40 m from the channel edge.On each parallel transect, three 1 3 2-m randomly-placed plots were assessed; shaded plots represent the plots that were not sampled.For the belt transect, a 2 m wide 40 m transect was set up and plots assessed at each meter from the channel (1 3 2-m plots).

Fig. 3 .
Fig. 3. Differences (mean and SE) between three sampling methods (belt, parallel, random plot) for (A) the number of samples, (B) the total area sampled, and (C) the amount of time taken to reach different species richness thresholds.Based upon a one-way ANOVA and Tukey-Kramer HSD tests for the 20-species thresholds, sampling methods (B) were significantly different; and time required for random plots (C) were highly significantly different from both parallel and belt transects methods, which did not differ statistically.

Table 1 .
Sample sites, locations within the San Francisco Bay-Delta estuary, size and number of sampling methods tried and where the site is a natural or restored wetland.The sites with only two methods tried did not try random plots.

Table 2 .
Comparison of belt and parallel transect methods for all nine wetland sites.The fourth column (labeled Design) indicates which method discovered more species; the fifth column (labeled Time efficiency) show the time ratio for belt vs parallel methods by dividing the time required for the belt transect by the time required for the parallel transect.The final column indicates how effective the arbitrary ten transects were for discovering the total species richness of the site.

Table 3 .
Cover data comparisons for the focal site, Coon Island, comparing random plots (Random Plots) combining data from two different years (206 and 217 plots, ''true'' values), versus 10 continuous belt transect data (Belt) and 10 parallel plot data along transects (Parallel).Data shown are average percent cover (standard error).