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. Ecologists have long used the concept of ecoregions to provide a framework for visualizing, understanding, and managing complex environmental factors, plant and animal habitats, and ecosystem processes (McMahon et al. 2001, Omernik 1995, Omernik 1987, Bailey 1983). While ecoregions were traditionally based on human expertise, quantitative methods, combined with multivariate observational and remote-sensing data, have more recently been applied to produce custom-developed regionalizations for specific analytical purposes (Hargrove and Hoffman 1999, Hargrove et al. 2003, Hargrove and Hoffman 2004, Hoffman 2004), including analyses involving temporal changes in environmental factors (Saxon et al. 2005, White et al. 2005, Hoffman et al. 2005, Hoffman 2010). Similarly, geologists often classify landscape areas into geomorphological units based on their geophysical and hydrological features (Ulrich et al. 2009, Schneider et al. 2009, Jorgenson 2000, Jorgenson and Ely 2001, Jorgenson and Brown 2005, Gude et al. 2002). For NGEE Arctic, we propose to unify these two stratification concepts to produce biogeomorphic units at relevant spatial scales for landscape characterization, identification of ecological and geomorphological features, assessment of the representativeness of measurements, and provision of a framework for scaling measurements and model parameters to larger domains or the entire Arctic.
J.Rowland (LANL) will lead this team.
The development of landscape units relies on the fusing of plot-scale derived data with landscape-scale measurements obtained from towers, aircraft, satellites, and regional mapping. In published applications for the Arctic, such unit characterization incorporates a suite of landscape properties such as topography, hydrology, vegetation type and productivity, soil moisture and ice content, soil characteristics, and land surface age (Hinkel and Nelson 2003, Ulrich et al. 2009, Schneider et al. 2009). Characterization of the landscape using biogeomorphic units requires (1) a definition of unit characteristics, which will depend on the intended use of the unit classification and the availability of relevant data, and (2) determination of the spatial distribution of these characteristics through remote sensing or up-scaling of point measurements. In well-studied regions with extensive data on characteristics such as vegetation type, bioclimatic factors, above- and below-ground carbon content, geophysical feature age, and soil structure and composition, maps of biogeomorphic units may be constructed by developing relationships between the properties of interest and the spectral properties of imagery from satellites such as Landsat-7 and by classifying the landscape using a combination of supervised and unsupervised data-mining techniques (Hinkel and Nelson 2003, Ulrich et al. 2009, Schneider et al. 2009). In the case of areas with limited field data, development of biogeomorphic units will be an iterative process with the level of complexity varying depending on the intended application of the classification.
As an example, a preliminary landscape-scale regionalization was performed using model-derived bioclimatic and observed topographic factors for the state of Alaska at a nominal resolution of 2 km2. A total of 37 characteristics, with model results averaged for the period 2000–2009, were included in the analysis. An unsupervised k-means algorithm, called Multivariate Spatio-Temporal Clustering (MSTC; Hargrove and Hoffman 2004), was applied to these data at various levels of division, yielding multiple maps of ecoregions for the state of Alaska. The map in Figure 6, shown in random colors, depicts the 20 most-different regions defined by MSTC. The North Slope, Brooks Range, central boreal forest, and
|
|
|
Multivariate Spatio-Temporal Clustering (MSTC) of20 regions in Alaska. |
other broad features in Alaska are clearly identifiable as distinct colors on the map. Blue circles on the map identify the geographic locations that best represent the mean combination of environmental conditions within each ecoregion. As such, these sites represent optimal sampling locations for each ecoregion. At a large scale, this technique is useful for delineating distinct broad regions and optimal measurement sites. However, this technique and similar methods―both supervised and unsupervised―can be applied at finer spatial scales, with inclusion of other geophysical characteristics and remote-sensing data, to inform measurement site selection within these broader ecoregions.
Perform landscape characterization for site selection and data gap assessment. In support of site selection and to establish a landscape-based framework for the assimilation of Year 1 data collected in the field and laboratory, an initial classification of landscape properties will be conducted to develop biogeomorphic units using published and unpublished datasets of climatology, topography, and other characteristics derived from existing remote-sensing data. These data sources may include
- basin ages (Hinkel and Nelson 2003);
- soil maps and vegetation indices [e.g., the Normalized Difference Vegetation Index (NDVI)] derived from high-resolution satellite imagery obtained from the National Geospatial-Intelligence Agency (NGA);
- polygonal ground topographic/geometric characteristics such as size, curvature, and geometric patterns (Gangodagamage et al. 2012);
- soil moisture from multispectral and/or radar data;
- NGEE Arctic-derived and prior study data on active layer thicknesses, such as the Circumpolar Active Layer Monitoring (CALM) Program;
- existing soil carbon inventories (e.g., Barrow Area Information Database (BAID) datasets; and
- topographic data such as slope, drainage networks, microtopography, and storage elements derived from the University of Texas-El Paso (UTEP) LiDAR dataset (Tweedie, unpublished data).
Incorporate field and laboratory datasets into spatially distributed datasets to project landscape properties. As new datasets, such as thaw layer depths, ground ice content, thermal characteristics, and carbon and methane fluxes, are generated from field and laboratory work, they will be scaled up to larger domains by developing relationships between these properties and surrogate data. An example of this approach was documented by Hinkel and Nelson (2003), who developed the relationships between drained lake basin age (in the Barrow region) to remotely sensed observations of vegetation type, surface water ponding, degree of polygonalization, basin wetness, and texture. In NGEE Arctic, progress has already been made to correlate geophysical and subsurface data on thaw layer and snow depth thicknesses to topographic and vegetation characteristics (Hubbard et al. 2012; Wainwright et al. 2012, Gangodagamage et al. 2012).
Apply supervised and unsupervised clustering algorithms to classify the landscape based on observable, proxy, and mapped landscape properties. A variety of agglomerative and divisive techniques have been applied to the general problem of classification, but data-mining methods for feature extraction and change detection that are designed to accommodate very large and complex datasets are being applied with marked success in the Earth sciences (Hoffman et al. 2011, Kumar et al. 2011). The k-means and similar approaches described by Hartigan (1975) have proven to be particularly useful for landscape characterization (e.g., Hargrove and Hoffman 2004) and detection of disturbance (e.g., Hoffman 2004, Hoffman et al. 2010). In addition to such established methods, novel unsupervised clustering methodologies using adaptive sparse representations (Moody et al. 2012) may offer promise as new techniques for characterizing the landscape and identifying unique assemblages of biogeomorphic properties at regional scales.
Develop model-specific landscape characterizations for model initialization and parameterization. Not all surface and subsurface properties used to characterize the landscape into distinct biogeomorphic units determined may be directly relevant to models. Additionally, data needs and parameterizations will vary from high-resolution process resolving models to regional and global climate models. Therefore, a critical task will be the development of a suite of model-relevant landscape properties associated with mapped biogeomorphic units. Ideally, the biogeomorphic units developed will be associated with unique sets of model-relevant properties; however, revision and remapping of units will likely be required to capture the full range of inputs and parameterizations needed across modeling scales and platforms. A critical task in this process will be the identification of observable and proxy characteristics (such as from remote-sensing or surface geophysical measurements) for the mapping of model parameterizations. In many cases this will likely be an iterative process requiring field validation. As an example, ground ice content will likely be a critical subsurface attribute, but it is not directly observable. Direct field measurements will be required to test the relationships between the preliminary mapped biogeomorphic units and ice content. The need to identify these types of relationships also highlights the importance of landscape characterization in the site selection process.
Empty