Representativeness-based sampling network design for the State of Alaska

The Arctic contains vast amounts of frozen water in the form of sea ice, snow, glaciers, and permafrost. Extended areas of permafrost in the Arctic contain soil organic carbon that is equivalent to twice the size of the atmospheric carbon pool, and this large stabilized carbon store could be released by widespread thawing of permafrost, resulting in a positive feedback to climate warming (Schuur et al. 2008). The Intergovernmental Panel on Climate Change (IPCC) Fourth Assessment Report (AR4) has documented strong evidence for warming of the Earth’s climate over the last century and has attributed the increase in global temperatures primarily to the rising anthropogenic greenhouse gas burden (IPCC 2007). Climate warming is projected to continue with broad implications for sensitive ecosystems and globally important climate feedbacks (Anisimov et al. 2007). Warming is projected to be especially pronounced at high latitudes and accompanied by significant regional impacts. Evidence of Arctic-wide responses are already being observed (Hinzman et al. 2005). Despite these potential implications, the Arctic has a limited record of low density observations. The Arctic Climate Impact Assessment (ACIA) (2005) emphasized the need for studies of the complex and interacting processes of the atmosphere, sea ice, ocean, and terrestrial systems to improve the interpretation of past climate and projections of future climate. The Committee on Designing an Arctic Observing Network (2006) identified critical needs and gaps for observations in the Arctic. It recommended an Arctic Observing Network to satisfy current and future scientific needs and offered recommendations on key physical, biogeochemical, and human dimensions variables to monitor.

Conducting systematic and continuous field observations and long term monitoring are challenging, particularly in the Arctic. Resource and logistical constraints limit the frequency and extent of observations, necessitating the development of a systematic sampling strategy that objectively represents environmental variability at the desired spatial scale. Statistical design of the network, particularly the location of sampling sites, is critical for maximizing the representativeness of the sampled data, given a fixed number of sampling locations. A methodology that provides a quantitative framework for stratifying sampling domains, informing site selection, and determining the representativeness of measurements is required to ensure that observations are well distributed across geographic and environmental data space. This information is needed for up-scaling and extrapolating point measurements to a larger landscape with similar environmental characteristics. This study addresses these needs by developing a quantitative methodology, based on the concept of ecoregions, for objectively delineating sampling domains, identifying optimal sampling locations for these domains, and quantifying representativeness of sites and measurements. This methodology is applied at the landscape scale to inform the design of a sampling network for the U.S. Department of Energy’s Next Generation Ecosystem Experiment (NGEE) Arctic project in the State of Alaska. The National Science Foundation’s (NSF’s) National Ecological Observatory Network (NEON) adopted an objective, data-based methodology to define 20 optimal sampling domains across the conterminous United States (Schimel et al. 2007; Keller et al. 2008). An extension of that same methodology was applied both across space and through time to support identification of measurement sites and provide a framework for scaling measurements and model parameters for the NGEE Arctic project.

Evidence of environmental change in the Arctic and resulting impacts on aquatic productivity and biodiversity, terrestrial ecosystems, and local economies were highlighted by Anisimov et al. (2007). Increased shrub abundance has been observed in Alaska (Sturm et al. 2001, 2005; Tape et al. 2006). During the last 50 years, the tree line along the Arctic to sub-Arctic boundary has moved 10 km northward and 2 % of Alaskan tundra on the Seward Peninsula has been replaced by forests. Ecoregions derived for the present and future (Fig. 1) show a similar northward shift, indicating a dramatic change in environmental conditions due to a warming climate by the end of this century, as projected by models using the A1B emissions scenario (Nakićenović et al. 2000). By tracking changes in the spatial area and migration of ecoregions statistically derived from a hypervolume of environmental gradients (Hutchinson 1957), this objective approach for mapping landscapes undergoing environmental change can be applied to predict shifts in species ranges and constrain estimates of changes in the carbon balance of sensitive environments.

Figure 4a shows the percent area distribution of each ecoregion, at the k = 10 level of division, for the present and future time periods. Correspondence between these MSTC Ecoregions and Nowacki et al. (2001) Level 2 ecological groups is shown in Table 4. A significant decrease in the area of Ecoregion #3, representing most of the North Slope of Alaska as shown in Fig. 3a, is observed. This contemporary Arctic Tundra environment is predicted to be reduced to about 0.78 % of its present area by the end of the century. About 76 % of the area will be replaced by conditions typical of the warmer Bering Tundra environment (Ecoregion #2). Meanwhile, the Bering Tundra (Ecoregion #2) environment moves northward by the end of the century and more than doubles in areal extent. About 70 % of its current area, especially over the Seward Peninsula, will change to conditions similar to contemporary Bering Taiga (Ecoregion #4). In the future, the Bering Taiga (Ecoregion #4) environment decreases in extent by 32 % and migrates northward. Under increased temperatures and reduced permafrost conditions, the present-day Aleutian Mountains (Ecoregion #7) environmental conditions are predicted to replace 65 % of Bering Taiga (Ecoregion #4), and Alaska Range Transition (Ecoregion #10) environmental conditions are expected to replace 28 % of Bering Taiga (Ecoregion #4). Aleutian Mountain (Ecoregion #7) and Alaska Range Transition (Ecoregion #10) environments, which exist in the southern coastal regions of Alaska, are expected to grow in extent northward and occupy a larger portion of Alaska. Alaska Range Transition (Ecoregion #10) environmental conditions are also expected to replace about 75 % of the Intermontane Boreal (Ecoregion #5) environment in the future, which will be reduced to 18 % of its current area by the end of the century. While similar trends of large scale northward migrations and changes in the areal extents of the environments discussed above are observed at 20 and higher levels of divisions, these ecoregion refinements highlight the changes that are occurring in smaller, more uniquely defined environments.

Open image in new window

Figure 4b shows the percent area distribution of k = 20 ecoregions for the present and future time periods. In addition to areal extent, changes and geographic redistribution of ecoregions between the present and future, at this level of division one present-day ecoregion ceases to exist in the future (i.e., becomes extinct) while another ecoregion exists only in the future (i.e., is born) and has no analog in the present. Ecoregion #13 (Fig. 5a), which represents the most northern portion of Arctic Tundra on the North Slope, becomes extinct in the future due to projected climate change. Ecoregions #2 and #17, which presently occupy the Seward Peninsula and nearby coasts (Fig. 5b), replace Ecoregion #13 in the future (Fig. 5c). Approximately 46 % of the area of Ecoregion #13 is replaced by Ecoregion #2 and 53 % is replaced by Ecoregion #17. Under this climate change scenario, the ecoregions replacing the extinct region in the future have characteristically higher precipitation, higher temperatures, earlier thaw dates, later freeze dates, a longer growing season, increased active layer depth, and higher ground surface temperatures. At the end of the century, much of the Seward Peninsula and nearby coasts are occupied by an entirely new combination of environmental conditions, defined by Ecoregion #1, which has no analog in the present (Fig. 5d). This new ecoregion, which appears only in the future time period, represents an environment with higher precipitation and temperature, an increased growing season length, increased active layer depth, and higher soil temperatures.

Open image in new window

As the level of division is increased in the MSTC algorithm, more specialized ecoregions are delineated. As a result, the number of present-day ecoregions that become extinct and the number of non-analog future ecoregions will both increase. Identification of regions representing new combinations of environment conditions that did not previously occur together is important for forecasting species range distributions, conservation planning, and climate change impacts on biodiversity (Fitzpatrick and Hargrove 2009).

Selection of sampling locations for long term monitoring of ecosystem properties and processes should be guided by an objective, quantitative, systematic, and defensible methodology. Instead, sampling locations in large-scale networks have often been established in opportunistic, political, or logistically-driven ways, resulting in unquantified representation of heterogeneity, biased sampling, uncharacterized uncertainty, and undirected network growth. Finite resources and logistical constraints limit the spatiotemporal frequency and extent of environmental observations, necessitating the development of a systematic sampling strategy to objectively represent environmental variability at the desired spatial scale. An appropriately designed observation strategy should be employed to quantitatively delineate sampling domains, sites, and frequencies. The NSF’s NEON adopted the objective, data-based methodology described above to define 20 optimal sampling domains across the conterminous United States (Schimel et al. 2007; Keller et al. 2008). Accurate characterization of the landscape and translation of data collected in the field and laboratory into useful datasets, process algorithms, and model parameters requires classification of the landscape into discrete units based on ecological, hydrological, and geological properties. In much the same way that ecologists develop ecoregions, geologists often classify landscape areas into geomorphological units based on their geophysical and hydrological features. For complex and evolving landscapes featuring interacting vegetation and geomorphological dynamics responding to changes in climate, such as in the Arctic, these stratification concepts may be unified to produce biogeomorphic units at relevant spatial scales for landscape characterization, identification of ecological and geomorphological processes, assessing the representativeness of measurements, and providing a framework for scaling measurements and model parameters to larger domains.

An important aspect of site selection and the up- and down-scaling approach to integration of models, observations, and process studies is the estimation of representativeness. The MSTC methodology described above for landscape characterization offers useful metrics for indicating the representativeness of sites, measurements, and model parameters, assuming the environmental characteristics included in the analysis covary with the measured variables. Hargrove et al. (2003) described this technique for understanding the representativeness of a sampling network based on a suite of environmental gradients considered to be useful proxies for the characteristics being measured. Maps identifying poorly represented regions can be produced, suggesting where new measurements should be taken to maximize sampling network coverage. As discussed in the "Alaska ecoregions" section, since the cluster centroid represents the mean value of all the state variables in an ecoregion, the realized centroid for an ecoregion is the location that best represents the combination of environmental conditions of the entire ecoregion. Therefore, statistically defined realized centroids, indicated by blue dots in each ecoregion in Figs. 1 and 2, are the optimal sampling locations for each ecoregion. Logistical constraints—including accessibility, availability of electric power and telecommunications infrastructure, and geologic stability—may prevent establishment of sampling sites at such optimal locations, particularly in an Arctic environment. Nevertheless, the MSTC Ecoregion framework provides a means for quantifying the representativeness of measurements taken at sub-optimal locations, either within an ecoregion or across any larger domain for which the desired state variables are available.

While most in situ field measurements are made at relatively small, individual geographic points, ecosystem processes operate at many scales. To utilize limited point measurements at larger spatial and temporal scales for input to or evaluation of process modeling or for estimating landscape-scale characteristics, the representativeness of those measurements must be quantified in the context of a heterogeneous and evolving landscape. A useful representativeness metric is one that can inform the selection of sampling locations, up-scaling of point measurements, down-scaling of remote sensing data, and extrapolation of measurements to unsampled domains. This requires that the underlying variables used to define ecoregions covary with the point measurements (i.e., the surrogate variables have and maintain predictive power). The representativeness metric described by Hargrove et al. (2003) provides a unit-less, relative measure of the dissimilarity between the ecoregion of interest, which may contain a sampling site, and any other ecoregion. It is calculated as the Euclidean distance between two ecoregion centroids within the standardized n-dimensional state space. Ecoregions with similar combinations of environmental conditions will have centroids located near to each other in state space. Therefore, the Euclidean distance between those centroids will be small, representing a low dissimilarity or high representativeness measure. Meanwhile, ecoregions with very different combinations of environmental conditions will have centroids located far from each other in state space, resulting in a large Euclidean distance between them. Such ecoregions will have a high dissimilarity or low representativeness measure. To best capture the natural heterogeneity at the scale of interest, this ecoregion-based representativeness should be calculated using MSTC Ecoregions with a large number of divisions (i.e, a large value of k).

While Hargrove et al. (2003) calculated representativeness in the context of ecoregions, this same approach can be applied to every map cell projected individually onto the n-dimensional state space used to perform the cluster analysis that produced MSTC Ecoregions. This point-based representativeness metric captures the full range of heterogeneity in the combinations of environmental conditions, providing a continuously varying measure of dissimilarity for every map cell with respect to a map cell of interest, which may contain a sampling location. When a single ecoregion centroid or map cell of interest is considered, a map of site representativeness can be produced. However, multiple ecoregions or map cells of interest may be considered simultaneously, for instance, to provide a quantitative measure of the representativeness of an array or network of sampling sites. The result is a map of network representativeness for which the dissimilarity measure for every ecoregion centroid or map cell is the Euclidean distance between that point and the nearest ecoregion centroid or map cell of interest (i.e., the minimum value from a stack of site representativeness maps, one for each ecoregion centroid or map cell containing a measurement site). This representativeness metric, whether ecoregion- or point-based, can be calculated not only between different geographic points in space, but also between different (or the same) geographic points through time. For example, the Euclidean distance between the present combination of environmental conditions and those of the future for any single map cell represents a measure of the magnitude of environmental change over time. Therefore, with this metric it is possible to calculate not only the present-day representativeness of measurements from a site, but also the future representativeness of those present-day measurements, based on future projections of the state variables used in the analysis.

Site representativeness

Due to significant logistical constraints when working in the Arctic, a set of eight potential sites were identified as candidates for measurements, long term monitoring and potential manipulative experiments for the U.S. Department of Energy’s NGEE Arctic project in the State of Alaska: Barrow, Council, Atqasuk, Ivotuk, Kougarok, Prudhoe Bay, Toolik Lake, and Fairbanks. Because of available support infrastructure, Barrow was selected as an initial location for collecting field measurements. To adequately capture the heterogeneity of environmental gradients, an ecoregion-based representativeness analysis employed ecoregion maps at the k = 1,000 level of division. Figure 6a shows the present-day representativeness of the monitoring site at Barrow for the present period. In this map, white to light gray land areas are well-represented by the Barrow location, while dark gray to black land areas are poorly represented by Barrow. The Arctic Tundra of the North Slope is well represented by the Barrow site, but the representativeness drops rapidly at the Brooks Range, which experiences different climate conditions driven by high topography. If a field researcher were attempting to select one additional sampling location to provide optimal coverage of the environments within the state of Alaska, that next site should be chosen within the darkest land areas shown in the map. Once a new candidate site has been selected, a new map of representativeness can be generated with simultaneous consideration of both sites. Using this relative representativeness metric, optimal sampling locations can be chosen to maximize the coverage of environmental conditions for any domain at any scale for which sufficient state variable data are available.

Open image in new window

Since climate model projections for the future were included in the MSTC procedure, the future representativeness of the present-day Barrow-containing ecoregion can also be mapped (Fig. 6b). Since the climate is projected to change significantly, the future representativeness of the present-day ecoregion is relatively lower, which is indicated by darker colors in Fig. 6b as compared with Fig. 6a. Such changes in representativeness are especially large in the Northern Arctic Coastal Plains since this Arctic Tundra is projected to warm significantly and has been identified as a sensitive environment ("Mapping sensitive environments" section). Similarly, Fig. 7a, b contain maps of the present and future representativeness of present-day Barrow, respectively, calculated using the point-based representativeness method. As expected, the large-scale pattern of maps in Fig. 7 is the same as that of the maps in Fig. 6, but the maps in Fig. 7 show more detail and are less generalized than those in Fig. 6. Point-based site representativeness maps for each of the eight candidate sites for the present time period are shown in Fig. 8.

Open image in new window

Open image in new window

Since the representativeness metric—or measure of dissimilarity—can be computed between any two map locations, a table quantitatively characterizing dissimilarity of the eight individual candidate sampling locations may be useful for site selection purposes. Table 5 shows point-to-point dissimilarity values for the eight candidate sampling locations for the present time period. Of those locations, Barrow and Fairbanks are the most dissimilar, having a dissimilarity value of 12.16. Atqasuk and Prudhoe Bay are the most similar of the sites. Both Atqasuk and Prudhoe Bay are near-coastal sites at the northern extent of the North Slope; therefore, the environmental conditions are expected to be similar. In addition, according to Table 5, the Prudhoe Bay site is most similar to Barrow, while the Council site is the most dissimilar to Barrow, ignoring Fairbanks. This example analysis suggests that if Barrow were the first sampling site selected, Council may be a strong candidate for a second site in the northern half of the State of Alaska because of its dissimilarity to Barrow. Similarly, Table 6 shows point-to-point dissimilarity values for the eight candidate sampling locations for the future time period. While the dissimilarity values for the future are similar to those of the present, it is apparent that some sites become more similar while others become less similar. For example, Barrow and Council become less dissimilar in the future (i.e., their dissimilarity value of 9.13 in the present changes to 8.87 in the future), indicating that the environmental conditions in Barrow and Council are more different in the present than they are projected to be in the future.

Table 5

Site state space distances for the present (2000–2009)

Sites	Council	Atqasuk	Ivotuk	Toolik Lake	Kougarok	Prudhoe Bay	Fairbanks
Barrow	9.13	4.53	5.90	5.87	7.98	3.57	12.16
Council		8.69	6.37	7.00	2.28	8.15	5.05
Atqasuk			5.18	5.23	7.79	1.74	10.66
Ivotuk				1.81	5.83	4.48	7.90
Toolik Lake					6.47	4.65	8.70
Kougarok						7.25	5.57
Prudhoe Bay							10.38

Table 6

Site state space distances for the future (2090–2099)

Sites	Council	Atqasuk	Ivotuk	Toolik Lake	Kougarok	Prudhoe Bay	Fairbanks
Barrow	8.87	4.89	6.88	6.94	8.04	4.18	11.95
Council		8.82	6.93	7.74	2.43	8.24	5.66
Atqasuk			5.86	5.84	8.15	2.30	10.16
Ivotuk				2.01	7.27	4.75	7.51
Toolik Lake					7.81	5.00	8.33
Kougarok						7.89	6.42
Prudhoe Bay							9.81

Table 7 shows a full matrix of point-to-point dissimilarity values for the eight candidate sites between the present and the future. This table quantifies the dissimilarity of present-day sites to those same sites in the future. For this list of widely dispersed locations, the environmental conditions for any single site in the present will be most like the environmental conditions for that same site in the future. Therefore, the smallest dissimilarity values are along the diagonal in Table 7. The largest value on the diagonal is for the Barrow site, indicating that environmental conditions at Barrow are projected to change more than at any other candidate site. In addition, this table shows that environmental conditions at Barrow in the future are more similar to those at Council in the present (8.38) than are the conditions at Barrow in the present to Council in the future (9.67). This result is consistent with the MSTC Ecoregion migration shown in Fig. 5. This point-to-point analysis through time is a novel method for quantifying relationships between sampling locations and how those relationships evolve over time due to environmental change.

Table 7

Site state space distances between the present (2000–2009) and the future (2090–2099)

	Sites	Future (2090–2099)
		Barrow	Council	Atqasuk	Ivotuk	Toolik Lake	Kougarok	Prudhoe Bay	Fairbanks
Present (2000–2009)	Barrow	3.31	9.67	4.63	6.05	5.75	9.02	3.69	11.67
	Council	8.38	1.65	8.10	5.91	6.87	3.10	7.45	5.38
	Atqasuk	6.01	9.33	2.42	5.46	5.26	8.97	2.63	10.13
	Ivotuk	7.06	7.17	5.83	1.53	2.05	7.25	4.87	7.40
	Toolik Lake	7.19	7.67	6.07	2.48	1.25	7.70	5.23	8.16
	Kougarok	7.29	3.05	6.92	5.57	6.31	2.51	6.54	5.75
	Prudhoe Bay	5.29	8.80	3.07	4.75	4.69	8.48	1.94	9.81
	Fairbanks	12.02	5.49	10.36	7.83	8.74	6.24	10.10	1.96

Network representativeness

A monitoring network often consists of a geographically distributed constellation of measurement sites or may be locations where samples are collected for further analysis in the laboratory. Quantifying the representativeness of the network as a whole is important for optimal network design to avoid unnecessary duplication and to maximize the coverage of the monitoring network. By combining multiple maps of site representativeness for every sampling location, and calculating the minimum value for every map cell, maps of network representativeness are produced. Figures 9a, b contain maps of ecoregion-based network representativeness for all eight candidate sampling sites for the present and future time periods, respectively. Similarly, Fig. 10a, b contain maps of point-based network representativeness for the same eight candidate sampling sites for the present and future time periods, respectively. White to light gray land areas are well-represented by the network of sites, while dark gray to black land areas are poorly represented by the network of sites. If the objective were to maximize the coverage of all environments in the State of Alaska, the next sampling location should be chosen within the darkest land areas shown in the map. Most of Alaska is well represented by this network of eight sampling locations.

Open image in new window

Open image in new window

Systematic sampling strategies are essential for understanding ecosystem responses to climate change and informing model development. In the harsh Arctic environment—where climate change appears to be most rapidly affecting sensitive ecosystems and vulnerable, carbon-rich permafrost—filling critical gaps in observations is expensive and technically challenging. To fully explore the regional and global implications of climate change in the Arctic, global Earth System Models must capture the important processes and feedbacks. Such models must be developed based on a rich body of observational data as representative as possible of multiple spatial and temporal scales. Meanwhile, finite resources and logistical constraints place restrictions on the number of sampling sites, spatial extent, frequency, and types of measurements that can be collected. This study proposes a quantitative, data-based methodology for stratifying sampling domains, informing site selection, and determining the representativeness of measurement sites and sampling networks.

Multivariate spatiotemporal clustering, based on k-means cluster analysis, was applied to down-scaled GCM results and observational data for the State of Alaska at a nominal resolution of 4 km² to define a set of ecoregions at multiple levels of division across two decadal time periods. Maps of ecoregions for the present (2000–2009) and future (2090–2099) were produced, showing how combinations of 37 environmental conditions are distributed across Alaska and how these combinations shift as a result projected climate change in the 21st century. Using this statistical approach, optimal sampling locations, called realized centroids, were identified for each ecoregion at every level of division. In addition, the resulting geographic shifts and changes in areal distribution of ecoregions suggested that some environments may disappear, many will be redistributed, and new ones will appear in the coming century. This analysis provides insights into the identification of the most sensitive and potentially vulnerable Arctic ecosystems. The Euclidean distance within the 37-dimensional state space used for MSTC provides a metric for representativeness. Gray-scale maps of representativeness, showing the similarity of every map cell to a list of eight candidate samples locations near town sites in Alaska, were produced for each site. Tables quantitatively characterizing the similarity of candidate sampling locations to each other across space and through time were generated. These tables are useful for understanding the strength of the environmental gradients between sites and how those gradients may change based on model projections of the future. Taken together, these analysis products provide model-inspired insights into optimal sampling strategies across space and through time, and these same techniques can be applied at different spatial and temporal scales to meet the needs of individual measurement or monitoring campaigns.

The representativeness of a sampling network is best maximized before the network is deployed. Even if additional “optimized” sites are added to an existing network, it will require many more additions to approach the theoretical maximum representativeness for a given number of initial sites. It is difficult, with only the sequential addition of new optimized sites, to achieve the same representativeness once some sampling sites have been established. Representativeness resulting from such network “repairs” rarely ever equal the representativeness of a network initially designed de novo with that same number of sampling sites. Even if the network is to be constructed in stages, it is best to design site placement using the final, ultimate complement of sites and to operate sub-optimally until the full network can be completed. Otherwise, many more sites will have to be added to the existing network to achieve the same representativeness than could otherwise have been designed in initially.

Cluster analysis and n-dimensional data space regressions offer quantitative methods for up-scaling and extrapolating measurements to land areas within and beyond the sampling domain and provide a down-scaling approach to the integration of models, observations, and process studies. The success of these methods depends upon selecting appropriate surrogate environmental characteristics that covary with the observations and parameters that will be up- or down-scaled. The accuracy of the up-scaled data will be higher for areas represented well by the monitoring network and lower for areas that are poorly represented. At a large scale, these techniques are useful for delineating distinct, broad regions and optimal measurement sites. However, this methodology can also be applied at finer spatiotemporal scales, with inclusion of other geophysical characteristics and remote sensing data, to inform measurement frequency and site selection within these broader ecoregions.