Hierarchical multi-scale classification of nearshore aquatic habitats of the Great Lakes: Western Lake Erie

Journal of Great Lakes Research 36 (2010) 757–771

Contents lists available at ScienceDirect

Journal of Great Lakes Research j o u r n a l h o m e p a g e : w w w. e l s e v i e r. c o m / l o c a t e / j g l r

Hierarchical multi-scale classification of nearshore aquatic habitats of the Great Lakes: Western Lake Erie James E. McKenna Jr. a,⁎, Chris Castiglione b a b

Tunison Laboratory of Aquatic Science, US Geological Survey, Great Lakes Science Center, 3075 Gracie Road, Cortland, NY 13045, USA Lower Great Lakes Resources Office, US Fish and Wildlife Service, 405 North French Rd., Amherst, NY 14228, USA

a r t i c l e

i n f o

Article history: Received 29 March 2010 Accepted 30 August 2010 Communicated by Rudstam Index words: Aquatic habitat classification Nearshore Fish assemblages Neural network models Great Lakes

a b s t r a c t Classification is a valuable conservation tool for examining natural resource status and problems and is being developed for coastal aquatic habitats. We present an objective, multi-scale hydrospatial framework for nearshore areas of the Great Lakes. The hydrospatial framework consists of spatial units at eight hierarchical scales from the North American Continent to the individual 270-m spatial cell. Characterization of spatial units based on fish abundance and diversity provides a fish-guided classification of aquatic areas at each spatial scale and demonstrates how classifications may be generated from that framework. Those classification units then provide information about habitat, as well as biotic conditions, which can be compared, contrasted, and hierarchically related spatially. Examples within several representative coastal or open water zones of the Western Lake Erie pilot area highlight potential application of this classification system to management problems. This classification system can assist natural resource managers with planning and establishing priorities for aquatic habitat protection, developing rehabilitation strategies, or identifying special management actions. Published by Elsevier B.V. on behalf of International Association for Great Lakes Research.

Introduction Coastal aquatic organisms of the Great Lakes reside within ecosystems that are physically and biologically complex, consisting of many continua of environmental and biological conditions and interactions (Christie, 1974; Mills et al., 2003; Munawar, 2003). Benthic substratum and attached structure are tied to fixed locations in space, but water of various condition and organisms move through that space. Peterson (2003) has described these two basic aspects of aquatic environments as stationary (i.e., structural) and dynamic (i.e., physicochemical). Physical and chemical influences occur across a range of spatial and temporal scales and can significantly affect different processes at different scales (Ricklefs and Miller, 2000; Wetzel, 2001; Schertzer, 2003). For example, the distance over which wind blows (i.e., fetch) can determine the wave energy at any particular location, the depth to which waters are well mixed, and generate currents and upwelling zones (Li et al., 1975; Pickett, 1977). Turbidity and benthic substratum conditions can be strongly influenced by the size and location of rivers flowing into the coastal zone (Allen, 1995; Wetzel, 2001). Organisms and water (along with its suspended load) move freely throughout nearshore areas and interact with deeper offshore waters as well as mixing with waters and organisms entering from coastal tributary systems. Ecological ⁎ Corresponding author. E-mail addresses: [email protected] (J.E. McKenna), [email protected] (C. Castiglione).

problems (e.g., habitat modification and water quality degradation) in the Great Lakes vividly illustrate the connected nature of coastal systems and the problems that can arise (Christie, 1974; Mills et al., 2003). This complexity of aquatic environments has been recognized by the National Fish Habitat Action Plan (NFHSDC, 2008) and is an integral part of the foundation for the national strategy for assessment and management of aquatic ecosystems. Nearshore habitats are critical to aquatic biodiversity and Great Lakes fish populations; more than 120 native or established fish species use coastal habitats as spawning and nursery grounds (Goodyear et al., 1982). Some of these species range widely and migrate through many different habitats annually (Scott and Crossman, 1973). The complex and dynamic coastal environment makes it difficult to determine which set of habitat conditions significantly affect aquatic organism survival and movement. In addition, increasing evidence that many local ecological problems stem from conditions at larger spatial scales highlights the need for tools to effectively address multi-scale problems (Matthews, 1998; Wetzel, 2001; Gido et al., 2006). Classification is a tool that helps us simplify complex systems into sets of discrete units with known characteristics and relationships to each other, summarizing the essence of what is pertinent to a particular problem. Hierarchical classification systems also allow for determination of relationships among classified elements and may be applied to habitats, biota, or both. Several classification systems have been developed for coastal areas to address various needs (Cowardin et al., 1979; Busch and Sly, 1992; Hudson et al., 1992; Olson et al., 2001; Connor et al., 2004; Madden et al., 2005; Spalding et al., 2007)

0380-1330/$ – see front matter. Published by Elsevier B.V. on behalf of International Association for Great Lakes Research. doi:10.1016/j.jglr.2010.09.005

758

J.E. McKenna Jr., C. Castiglione / Journal of Great Lakes Research 36 (2010) 757–771

and numerous researchers have discussed important aspects of habitat within the context of landscapes (e.g., Davis and Henderson, 1978; Kotliar and Wiens, 1990; Seelbach et al., 2006; Riseng et al., 2008). An effective aquatic habitat classification system should function to (1) generate multiple classifications, because classification is needed for a wide variety of problems; (2) both generate hypotheses about the natural world and serve as a practical tool; (3) provide a basis for cataloging habitat status and making spatial comparisons; (4) provide a framework for developing management strategy; and (5) communicate knowledge and experience about aquatic habitat. In order to perform these functions, a coastal classification system should (1) be inclusive of all space and habitat types; (2) include both nearshore and tributary influences; (3) include hydrologic regime; (4) account for the mosaic of habitat patches on the landscape; (5) be a nested hierarchy of scale to address issues at a variety of scales; (6) be ecological in nature, including both biotic and abiotic features; (7) be mappable and use new technologies for organization and analysis (e.g., geographic information systems); and (8) be readily buildable for an entire region. Many different systems accommodate one or a few of these features, but there remains a need to effectively characterize, compare, and contrast coastal habitat units with a practical system that accommodates the range of scales and scale-dependent processes that significantly affect fish and other aquatic resources. In this paper, we describe a classification system that incorporates all of these characteristics, including practical instructions and data requirements for coastal aquatic habitats of the Great Lakes. The Great Lakes Regional Aquatic Gap Analysis Coastal Project Classification System (C_Gap) improves on existing classifications by providing a flexible, multi-scale framework for construction of organism- or habitat-specific classifications (or ecological hybrids) that are appropriate for different needs and situations in coastal ecosystems. Our objectives were to describe the C_Gap hydrospatial framework and demonstrate the development of a habitat classification using a fish-based example. We discuss application of the fish-based classification to important aquatic resource issues. Effective prediction of organism occurrence and abundance within a lakescape is integral to an effective biologically based spatial classification and analysis tool. Therefore, we begin with an explanation of our spatial data and predictive model development. We deal first with the stationary aspect of aquatic habitat and show how the resulting framework may be used to deal with the multitude of dynamic factors that determine the character of a given unit of habitat at any particular time. Our focus is on western Lake Erie, where extensive fish data are available from survey work of the Ohio Environmental Protection Agency (OEPA) and Ohio Department of Natural Resources (ODNR). Methods The Western Lake Erie Demonstration Area Western Lake Erie is a 4100-km2, shallow (typically b10 m deep) system with a coastline that includes two major embayments, two large peninsulas, and numerous islands and shoals. Prevailing winds blow from the southwest and the fetch can be as large as 70 km. Western Lake Erie receives most water from the Detroit River (80%) and a smaller amount from the Maumee River (5%) and a number other tributaries (Bolsenga and Herdendorf, 1993). These coastal tributaries deliver varying amounts of dissolved and particulate materials and are themselves habitat for fish and other aquatic organisms. Benthic substratum ranges from rock to fine mud (Environment Canada, 1997) and submerged aquatic vegetation (SAV) grows in some areas. Extensive coastal wetlands exist along the southern and western shores (Herdendorf, 1992). Western Lake Erie is roughly bounded on the east by two major peninsulas, Point Pelee on the north shore and Scott Point on the south shore. We chose

to use a line extending from just east of Point Pelee south to the Huron River, Ohio (approximately 82° 33' W longitude), as the eastern boundary of our study area because it allowed us to include Sandusky Bay and adjacent habitat to the east (Fig. 1). Due to the great complexity of deep water systems (e.g., upwelling and circulation, and weaker coupling of pelagic and benthic systems), we chose not to extend our scope beyond the nearshore zone, except in the small pockets that intruded into the western Lake Erie study area. Habitat data Energy in large lentic systems like the Great Lakes is generally provided by wind and waves and differential exposure creates areas of high or low energy. The wave regime of a particular coastline determines to what depth of water bottom sediments are physically reworked. We defined this depth as the offshore boundary of nearshore zones, calculated as:

√ðgdhÞdT = 2 ; or Zn = 10 m; Zn =

ð1Þ

whichever was deeper, where Zn is the depth of the offshore boundary (m), g is the acceleration due to gravity (9.8 m/s2), h is wave height (m), and T is wave period (Knauss, 1978). The alternative criterion of 10 m was chosen because in typically calm areas of the lakes, the dynamic definition of the nearshore zone becomes impractically small (b1 m in many cases). The SAV is an important aspect of nearshore zone habitats (Wetzel, 2001) and the band of waters ≤10 m usually encompasses the areas where light in Lake Erie penetrates close enough to bottom sediments to allow for SAV growth (when other conditions permit). In the case of western Lake Erie, the limited fetch distances and shallow waters define nearly all of it as nearshore habitat (≤10 m). Locations within the study area were represented in raster form (i.e., simply as an array of equally sized cells) in a Geographic Information System (GIS) using ArcGIS 9.2 (ESRI, Inc., Redlands, California). The 270-m cell (0.07 km2) was chosen as the best compromise among data spatial resolutions and computational capacity and is the basic unit of space containing all habitat and biotic information in this study. Twenty-two georeferenced habitat variables were available to describe summer habitat conditions at each local cell (Table 1). Two basic data sets describe the nearshore zone in this work, (1) the lake shoreline, which provides spatial reference to coastal features, and (2) the lake bathymetry, from which bottom slope, aspect, and other physical geometry characteristics are derived. In addition to lentic geometry, the values of areal habitat characteristics, such as bedrock geology, bottom sediments, and water temperature, were determined for each spatial cell. These variables have different inherent spatial scales. For example, identity of the major circulation gyre and bedrock geology type are the same at all points throughout the western Lake Erie pilot area (N80 km range). Exposure varies with orientation and effective fetch (1–80 km range). Geomorphology is an indication of the structure and shape of the coastline and how it has been modified (0.1–10 km range). Local variables (e.g., water temperature and bottom sediment) indicate conditions in and on the bottom and in the overlying water column (0.1 km range). These data were acquired from publicly available, georeferenced databases, with the exception of bottom sediment (Environment Canada, 1997) (Table 1). No complete spatial coverage of SAV was available for our pilot areas. Therefore, we applied a modification of an algorithm developed by Minns et al. (1995) that uses available habitat features of bottom sediment type (substrate as sand or finer), effective fetch (b2 km), and bottom slope (b15%) to estimate where SAV is expected to cover more than 50% of the bottom (Fig. 1).

J.E. McKenna Jr., C. Castiglione / Journal of Great Lakes Research 36 (2010) 757–771

759

Fig. 1. Western Lake Erie Aquatic Lake Unit (study area) with 5-m and 10-m isobaths separating progressively darker gray areas, predicted distribution of SAV coverage ≥50% (darkest gray), and OH EPA (triangles) and OH DNR (circles) fish collection locations.

Fish data

Model development process

Observed fish abundances were acquired from two extensive and well-organized databases maintained by the Ohio Department of Natural Resources (ODNR) and Ohio Environmental Protection Agency (OEPA). The OEPA fish data were collected via electrofishing (1982– 2002), typically from waters ≤5 m deep (Fig. 1) (OEPA, 2003). The ODNR fish data were collected via bottom trawls (1987–2002), typically from waters N5 m deep (ODNR, 2003). These data were georeferenced and accompanied by measures of collection effort and local habitat conditions. Catch data for each fish species were standardized (CPUE), depending on the gear used and effort applied, and mean standardized CPUE within sampling areas was classified on a log abundance scale: 0, 1, 2–10, 11–100, 101–1000, 1,001–10,000, and N10,000 (number / 1000 m2) to further smooth variability and gear differences.

Identification of species–habitat relationships was achieved by direct gradient analysis of habitat and matching fish observations using the canonical correspondence analysis (CCA) (CANOCO program, ter Braak, 1995; ter Braak and Smilauer, 2002). The forward selection process of CCA ensured that only significant habitat variables were included in the analysis and results indicated the relative influence of each habitat variable on the fish community. The 10–15 habitat variables that ranked highest in influence were used in neural network development. The standardized observed fish and matching habitat data were used together to train artificial neural networks (NN) and develop predictive models for each species that was represented by at least 10 occurrences within a data set. Separate models were developed for

Table 1 Georeferenced habitat variables used in this analysis. Variables used in neural network models based on a particular data set are indicted under the Model heading as E for OH EPA data or D for OH DNR data. Abbreviated variable codes are provided in brackets after each name. Habitat variable Benthic 3-D structure (% of rock, boulder, and SAV) [3-D] Benthic Substratum [Substrat] Coastline geomorphology [Geomorph] Coastline protection [Protect] Coastline sediment [Shorsed] Coefficient of variation of water temperature [TempCV] Density of rivers entering ALU (rivers/km) [Rivdens] Direction to nearest delta-type wetland (degrees) [Deltadi] Direction to nearest protected-type wetland (degrees) [Protecdi] Direction to nearest open-type wetland (degrees) [Opendi] Distance to mouth of nearest river of Strahler 4 or larger (m) [Rivdist] Distance to nearest delta-type wetland (m) [Deltdst] Distance to nearest open-type wetland (m) [Opendis] Distance to nearest protected-type wetland (m) [Protdis] Effective fetch (m) [Fetch] Ice cover duration (days) [Icedur] Mean water temperature (°C) [Xtemp] Secchi Depth (m) [Secchi] Sinuosity of coastline (straight-line distance:coastline distance) (m/m) [Sinuous] Slope of submerged bottom (change in depth m/m) [Slope] Submerged aquatic vegetation (% cover) [SAV] Water depth (m) [Depth]

Model

E E E D E E D, E E

D D D, E D

Source Computed from Environment Canada (1997) and NOAA (1994) Environment Canada (1997) NOAA (1999) NOAA (1999) NOAA (1999) Computed from NOAA (1994) Computed from USGS (2004) Computed from GLC (2004) Computed from GLC (2004) Computed from GLC (2004) Computed from NOAA (1994) and USGS (2004) Computed from GLC (2004) Computed from GLC (2004) Computed from GLC (2004) Computed from NOAA (1994) NOAA (1994) NOAA (1994) OH DNR 2002 and OH EPA 2002 Computed from NOAA (1999) Computed from NOAA (1994) Computed from Environment Canada (1997) and NOAA (1994) NOAA (1994)

760

J.E. McKenna Jr., C. Castiglione / Journal of Great Lakes Research 36 (2010) 757–771

each data set (ODNR and OEPA). Neural networks are excellent at distinguishing among entities described by non-linear, multivariate suites of characteristics and typically explain most data variation with low error rates (Olden and Jackson, 2001; McKenna, 2005; McKenna et al., 2006). We used the best results from training two different but simple NN designs: simple back propagation with a single hidden layer of neurons or a Wards Net model with three hidden layers of neurons (Ward Systems, 2000). The number of neurons in the hidden layer was determined by the following equation,

water depth and the ODNR-based model was applied to deeper areas. If a species' model was developed from only one data set, it was applied throughout the study area. A minimum diversity index (Shannon–Wiener Information Index, H′min; Shannon and Weaver, 1949) was computed based on the set of mean modeled species abundances expected within each habitat patch. This is a minimum measurement, because the rarest species were not modeled.

NH = OðNI + NO Þ + √DT ;

Hydrospatial framework We distinguish between the basic hydrospatial framework and specific habitat classifications generated from it (Davis and Henderson, 1978). The framework is composed of generic, spatially nested building blocks (Fig. 2). This design provides the flexibility needed to allow generation of any number of habitat classifications, while encompassing the entire study area within a common set of spatially related components. Any biotic or habitat attributes (or both) of interest may be assigned to the finest scale of the framework (the spatial cell) in order to generate a particular classification. The fishbased classification described below uses these generic building blocks to classify habitat according to predicted fish distributions. Fine-scale values are aggregated up the hierarchy scale and grouped within each successive scale level according to similarity, providing broader-scale classes and analyses. Abbreviations for framework elements (e.g., COZ and AHA) will be represented in bold-face throughout this document to easily distinguish them from classification elements. The hierarchical framework includes spatial scales from the major continental drainage basin to the 270-m spatial cell. The largest scale discussed here is the Great Lakes Region and nested finer scales include, each Great Lake (Subregion), circulation compartment (Aquatic Lake Unit), coastal zone (Coastal or Openwater Zone), geomorphological coastline section (aquatic habitat area), and local site (spatial cell) (Fig. 2). Large-scale units (Zone, Subzone, and Region)

ð2Þ

where NH is the number of neurons in the hidden layer, NI is the number of input neurons, NO is the number of output neurons (always one in this study), and DT is the number of observations in the training data set (McKenna, 2005). Leaning (0.1) and inertia (0.1) rates were implemented to ensure global rather than local convergence during training. The criterion for acceptance of an NN model was explanation of ≥70% of observed classified fish abundance variability, as measured 2 by the coefficient of multiple determination (i.e., Radj ≥ 0.7). 2 Radj

=


∑ y− yˆ 2 ∑ðy−yÞ2

;

ð3Þ

_ where y is the observed value, ŷ is the predicted value, and y is the mean of the observed values (Ward Systems, 2000). Twenty percent of the data were held out as a validation data set, which provides the models with greater ability to extrapolate beyond the training data and prevents overfitting (Ripley, 1996; Olden and Jackson, 2002; Olden et al., 2002). The models were then applied to predict the abundance of each species in each local cell throughout the study area and the GIS was used to map habitat conditions and both observed and predicted fish abundances. When a species' model was successfully developed from each data set, the OEPA-based model was applied to locations of ≤5 m

Habitat classification

Fig. 2. Diagram of hydrospatial framework (left) and example fish-based habitat classification (right). Sections within dashed boxes are discussed in detail in the text. The scale level is identified in bold; a specific example unit is named below each.

J.E. McKenna Jr., C. Castiglione / Journal of Great Lakes Research 36 (2010) 757–771

match those defined by Maxwell et al. (1995) and have been described in other aquatic habitat classification efforts (e.g., Sowa et al., 2005). The Great Lakes Region is the Great Lakes watershed (Maxwell's St. Lawrence Region) and is a component of the ArcticAtlantic Subzone and the Nearctic Zone, which are continental and sub-continental in scale. The subregion corresponds to each Great Lake and its local drainage basin (similar to Sly and Busch's (1992) System), in this case Lake Erie. Aquatic lake units (ALU) are generally defined by broad circulation basins within each lake and may be units of broad ecological division; these share some characteristics with Sly and Busch's (1992) Basin Division. In our example, Lake Erie consists of three ALUs, the Western, Central, and Eastern Basins (Fig. 1), and our study area is the western ALU. Coastal or open water zones (COZ) are broad areas within ALUs that have distinct habitat conditions associated with major coastline features, islands, or open water (Fig. 3). The COZs adjacent to mainland areas are defined as (1) large embayments (length: width ≥ 1), (2) great river mouths (those of Strahler order ≥ 8), and (3) areas along open shores between mouths of rivers of order 5 or larger or other major coastline features. The offshore boundaries are determined by projecting these coastline boundaries out from the shoreline to the 5-m isobath, adding adjacent areas where SAV is expected to exist in deeper water. The fourth type of COZ includes the areas around island archipelago groups, including shallow seamounts (b5 m) and adjacent areas expected to support SAV growth. The last type covers the remaining open water area of depths less than or equal to that of the nearshore boundary (10 m in this case). Aquatic habitat areas (AHA) are subdivisions of each COZ that characterize habitat and processes that occur at different spatial scales within the ALU. Local sites are each of the 270-m cells within each AHA. The AHA boundaries were determined by values of a multimetric index that provides a measure of energy level; intensity of upland effects; habitat structure, productivity, and quality; and habitat volume and degree of benthic-pelagic coupling. This index consists of effective fetch, distance to the nearest large river (Strahler order 4 or greater), presence or absence of SAV, and water depth values at each local site. Each variable value was classified as high (3), medium (2), or low (1) (or SAV covering b50% [0] or ≥50% [1] of the

761

bottom) to provide a code for each metric of the index. Those values were then concatenated in the order given above to provide a unique identification code for each AHA type. For example, an AHA coded 3311 is at the end of a large fetch, is distant from a major river, supports SAV, and is shallow. These were computed at each local site and all adjacent cells of the same type (i.e. index code) constituted an AHA (Fig. 3). Adjacent similar units on opposite sides of a boundary of the next larger scale unit (COZ in this case) were not fused into a single AHA patch. This maintains the integrity of larger scale unit boundaries throughout the framework hierarchy. Fish-based classification Aquatic habitat classification for any specific purpose may be generated by populating the local spatial cells of the generic framework with the appropriate georeferenced data and combining similar spatial units at each scale. We illustrate this with a classification based on predicted fish assemblages. Model-predicted abundances for each fish species were computed and assigned as attributes of each local site (270 m cell) within the Western Lake Erie ALU, along with the habitat attributes used in the NN models. Thus, both predicted fish abundances and habitat conditions to which the fish respond were available to describe environmental conditions at any location within the study area. Bootstrapping cluster analysis provides a statistically objective means of predicted fish assemblage identification. These cluster groups represent the distinct (potential) fish assemblages expected to be supported by the habitat of each AHA, with the assemblage of each constituent local site as a replicate sample within each AHA (McKenna, 2003). The cluster analysis used the Bray-Curtis similarity index and the UPGMA linkage method. One hundred bootstrap samples were used to test the significance of each cluster group fusion (α = 0.05). Highly similar AHAs were then considered to support the same fish assemblage and similar adjacent AHAs (within the same COZ) were merged to form fish habitat units that we label Fisheries Conservation and Management units (FCM) (Fig. 3). Larger scales of the classification were then formed by averaging characteristic assemblages within their boundaries from the next finer scale. Assemblages of each FCM within a given COZ were averaged (on a species-specific basis) to form Aquatic Ecological Systems (AES, akin to those of Sowa et al., 2005). For example, all fish assemblages within AHAs of the Sandusky Bay COZ were averaged to represent the fish assemblage of the Sandusky Bay AES (Fig. 7). This aggregation process is continued up the hierarchy of scale. Other attributes (e.g., habitat, etc.) of these ecological units are also averaged to determine the characteristic abiotic (or other) conditions that are associated with each distinct ecological type at each spatial scale. The result is a classification of every unit of habitat within the study area with known characteristic fish assemblage and associated habitat conditions, and with clear relationships in space and across scales. Results Observed fish assemblages and species–habitat relationships

Fig. 3. Western Lake Erie Aquatic Lake Unit (ALU) classification elements. Coastal or open water zones (COZ) (heavy solid lines), aquatic habitat areas (AHA) (gray dashed lines), and Fisheries and Conservation Management units (FCM) (colored polygons) within the Western Lake Erie Ecological Lake Unit (ELU).

The OEPA data set provided 225 fish collections from 43 sampling locations within the Western Lake Erie study area (Fig. 1) and included 69 fish species. The most widespread, frequently occurring species were freshwater drum Aplodinotus grunniens and white perch Morone americana; white perch was also the most abundant along with gizzard shad Dorosoma cepedianum. Mean standardized abundance reached 339/1000 m2 for gizzard shad. The rarest species (those occurring only once) included, coho salmon Oncorhynchus kisutch, fantail darter Etheostoma flabellare, northern pike Esox lucius, rainbow trout Onchorynchus mykiss, river chub Nocomis micropogon, striped shiner Luxilus chrysocephalus, suckermouth minnow Phenacobius

762

J.E. McKenna Jr., C. Castiglione / Journal of Great Lakes Research 36 (2010) 757–771

mirabilis, and three-spine stickleback Gasterosteus aculeatus. Although 34 species occurred ≥10 times and 18 of those species occurred at N1/2 the locations, only freshwater drum occurred at all sites. The ODNR data set provided 3083 fish collections from 62 sampling locations within the study area (Fig. 1) and included 66 fish species. The most widespread, frequently occurring species were yellow perch Perca flavescens and freshwater drum, while gizzard shad and white perch were most abundant. Mean standardized abundance reached nearly 700/1000 m2 for gizzard shad. The rarest species included, bowfin Amia calva, fathead minnow Pimephales promelas, largemouth bass Micropterus salmoides, Muskellunge Esox masquinongy, brook silverside Labidesthes sicculus, three-spine stickleback, and yellow bullhead Ameiurus natalis. Thirty species occurred ≥10 times and 23 of those species occurred at N1/2 the locations, but only freshwater drum and white perch occurred in all sample areas. The Morone spp. pair (white perch + white bass Morone chrysops) were reported as a group frequently enough and were abundant enough that the genus was also modeled. For the OEPA data, the species–environment correlations of the first three CCA axes ranged from 0.67 to 0.87 and explained 70% of that relationship (Table 2). CCA results also indicated that eight habitat variables most strongly influenced the OEPA fish assemblages (Table 1). For the ODNR data, the species–environment correlations of the first three CCA axes ranged from 0.77 to 0.83 and explained 79% of that relationship (Table 2). CCA results indicated that six habitat variables most strongly influenced the ODNR fish assemblages (Table 1). Neural network model development and predictions A total of 65 neural networks were constructed to predict speciesspecific fish abundances, 34 based on OEPA data and 31 (including the Morone generic group) based on ODNR data. All models, except for log-perch Percina caprodes (R2 = 0.52) explained at least 70% of variability (Table 3); most explained N80% and about half accounted for N90%. Model predictions estimate abundance for any species at any location or area from the local cell to the entire study area (ALU). Spatial aggregation and assessment units are discussed below. As expected, predicted fish distributions show different abundances in different areas. We use examples of a forage species (alewife Alosa pseudoharengus), two game fish (smallmouth bass Micropterus dolomieu and yellow perch), and H′min to illustrate model prediction results. Alewife displayed a wide range of predicted abundances within the study area with the greatest generally in waters N5 m deep, particularly north of the Crane Creek area, east of Maumee Bay (Fig. 4a). Patches of moderate abundance also occurred in major embayments and river mouths as well as scattered areas along the southern and western shores. Predicted smallmouth bass abundances were greatest in the western portion of the study area, particularly in waters b5 m deep, and around the islands closest to the south shore. Patches of moderate abundance were found in major embayments and river mouths and along the southern and western shores (Fig. 4b).

Table 2 CCA results for each observed fish database. SEC = species–environment correlation and CVS = cumulative percentage of the species–environment variance explained. Data Set

OH DNR

OH EPA

Axis

Eigenvalues SEC CVS Eigenvalues SEC CVS

1

2

3

4

0.079 0.833 47.8 0.170 0.875 40.4

0.028 0.698 64.9 0.073 0.739 57.8

0.023 0.772 78.8 0.053 0.668 70.3

0.014 0.717 87.2 0.039 0.615 79.6

Total inertia 0.677

2.737

Yellow perch also displayed a wide range of predicted abundances with the greatest in extensive patches of open water N5 m deep and around the southern islands. Patches of moderate abundance occurred in some shallow areas along the entire coastline (Fig. 4c). In the classic debate about whether ecological communities are fixed associations of organisms (Clements, 1936) or independent arrangements of organisms responding to environmental conditions (Gleason, 1939), elements of both schools are correct. However, our models were developed as independent, species-specific models driven by abiotic variables. The habitat variables used for each model are shown in Table 1 and the relative influence of each varied by species. Predictions for each species were made for each local spatial cell. Thus, the resulting local species assemblages are strictly Gleasonian, interspecific interactions had no influence. The cluster analysis compared both species composition and relative dominance to identify distinct types among the local assemblages. Because most anthropogenic factors were excluded from model development, the resulting assemblages represent the potential for a given unit of habitat and might be considered “target fish communities” (sensu Bain and Meixler, 2008). However, the lack of rare species and interspecific influence limits the accuracy of these assemblages. Despite these limitations, they represent a major cross-section of the natural species assemblages and potential interspecific interaction can be examined with additional directed study. The predicted assemblages also provide a consistent (though truncated) measure of diversity. Predicted assemblages with the highest minimum diversity (H′min N 2.3) were expected to occur in waters b5 m deep from the west side of Scott Point to the Detroit River (Fig. 4d). Those of moderate diversity (H′min = 1.5–2.0) were expected in Sandusky and Maumee Bays and throughout extensive areas of deeper water, particularly along the north shore and in deep water to the east. The full range of assemblage diversity occurred in the islands area in many small and medium size patches, with the areas supporting the highest diversity located closest to small islands and shoal areas. The basic hydrospatial framework The study area is contained within a single compartment of the 4th level of the nested hydrospatial framework; specifically, the westernmost of the three aquatic lake units in the Lake Erie Subregion (part of the Great Lakes Region). This ALU is composed of 11 coastal or open water zones (plus a deepwater zone, N10 m water depth) (Fig. 3). There are five general types of COZs in the study area: embayment, open coast, great river delta, islands, and open water. Each COZ is composed of various numbers of aquatic habitat areas of varying sizes (i.e., number of local sites contained) for a total of 603 of 32 types. For example, the Sandusky Bay COZ consisted of two AHAs (each a different type), while the open water COZ consisted of 298 (25 types) (Table 4). A total of 56,816 local sites (i.e., 270-m spatial cells) comprise the study area. All of these spatial containers have known location, definable size, and together account for all aquatic space within the study area at each of the framework levels (i.e., scales). Fish-based classification of aquatic nearshore habitat Each local site was attributed with the predicted abundance of each modeled species; together, this array of abundances forms the potential fish assemblage supported by the habitat in that local site. As described above, various numbers of adjacent local sites comprise each AHA. Cluster analysis of the predicted fish assemblages within the AHAs identified 82 distinct assemblages; these are the Fisheries Conservation and Management unit types (Figs. 2 and 3). Adjacent patches of these types, within each COZ, combined to create a mosaic of 308 FCMs (i.e., patches of those 82 types) within the study area

J.E. McKenna Jr., C. Castiglione / Journal of Great Lakes Research 36 (2010) 757–771

763

Table 3 Neural network model performance. Max class indicates the mid-point of the largest abundance class predicted to be present, followed by the number of observed occurrences of each species in the data (f). The n = the total number of observation. OM% = omission error (predicted to be absent when observed present), COM% = commission error (predicted to be present when observed absent), R2 = adjusted coefficient of determination, and Method indicates the neural network architecture used (SIMPLE = simple back propogation model and WARD = Ward's net model). Common name

Alewife Gizzard shad Lake whitefish Rainbow smelt Goldfish Spotfin shiner Common carp Silver chub Emerald shiner Spottail shiner Mimic shiner Sand shiner Bluntnose minnow Quillback White sucker Smallmouth buffalo Bigmouth Buffalo Shorthead redhorse Brown bullhead Channel catfish Trout-perch White perch White bass Morone spp. Rock bass Green sunfish Pumpkinseed Orangespotted sunfish Bluegill Smallmouth bass Largemouth bass White crappie Black crappie Greenside darter Johnny darter Yellow perch Walleye Logperch Freshwater drum Round goby

Scientific name

Alosa pseudoharengus Dorosoma cepedianum Coregonus clupeaformis Osmerus mordax Carassius auratus Cyprinella spiloptera Cyprinus carpio Macrhybopsis storeriana Notropis atherinoides Notropis hudsonius Notropis volucellus Notrpis strameus Pimephales notatus Carpiodes cyprinus Catostomus commersonii Ictiobus bubalus Ictiobus cyprinellus Moxostoma macrolepidotum Ameiurus nebulosus Ictalurus punctatus Percopsis omiscomaycus Morone americana Morone chrysops Morone spp. Ambloplites rupestris Lepomis cyanellus Lepomis gibbosus Lepomis humilis Lepomis macrochirus Micropterus dolomieu Micropterus salmoides Pomoxis annularis Pomoxis nigromaculatus Etheostoma blennioides Etheostoma nigrum Perca flavescens Sander vitreus Percina caprodes Aplodinotus grunniens Apollonia melanostoma

OH EPA (n = 43)

OH DNR (n = 62)

Max class (f)

OM%/COM%

R2

Method

Max class (f)

OM%/COM%

R2

Method

50 (16) 5,000 (40)

14/30.2 11.6/0

0.90 0.92

SIMPLE WARD

50 (27) 500 (18) 500 (40) 1 (5) 500 (40) 500 (38) 500 (25) 5 (11) 50 (18) 5 (14)

0/20.9 7/37.2 20.9/7 9.3/11.6 11.6/2.3 7/2.3 7/18.6 7/30.2 4.7/34.9 7/32.6

0.82 0.94 0.96 0.90 0.80 0.72 0.88 0.93 0.94 0.73

SIMPLE SIMPLE WARD WARD WARD WARD WARD WARD WARD WARD

5,000 (57) 50,000 (60) 1 (14) 5,000 (53) 500 (18)

22.6/6.5 21/1.6 0/37.1 19.4/11.3 8.1/33.9

0.84 0.73 0.93 0.93 0.98

SIMPLE WARD SIMPLE WARD WARD

(54) (58) (60) (58) (34) (38)

8.1/11.3 17.7/4.8 1.6/3.2 30.6/1.6 4.8/29 12.9/21

0.97 0.86 0.91 0.77 0.99 0.96

WARD SIMPLE WARD WARD SIMPLE SIMPLE

50 (36) 5 (39)

1.6/29 8.1/25.8

0.96 0.81

SIMPLE SIMPLE

50 (20) 5 (12) 50 (21) 50 (19) 50 (31) 5 (10) 5,000 (42) 50 (40)

0/30.2 2.3/23.3 2.3/20.9 14/37.2 9.3/16.3 0/34.9 23.3/2.3 9.3/2.3

0.98 0.84 0.97 0.99 0.74 0.93 0.84 0.89

WARD SIMPLE WARD SIMPLE WARD WARD WARD WARD

500 (32) 50 (16) 50 (31) 5 (10) 500 (32) 50 (32) 50 (30) 5 (10)

0/18.6 4.7/23.3 2.3/20.9 0/53.5 9.3/11.6 11.6/9.3 0/30.2 0/30.2

0.95 0.99 0.64 0.99 0.98 0.84 0.86 0.89

WARD SIMPLE GRNN SIMPLE WARD WARD SIMPLE WARD

17.7/24.2 3.2/12.9 3.2/3.2 0/0 17.7/1.6 16.1/17.7 3.2/35.5

0.95 0.77 0.91 0.96 0.75 0.94 0.90

WARD SIMPLE WARD WARD WARD SIMPLE SIMPLE

5 (10)

0/35.5

0.84

WARD

1 (19) 5 (31)

0/40.3 1.6/25.8

0.75 0.91

SIMPLE WARD

50 (36) 5 (13)

25.8/16.1 1.6/51.6

0.73 0.90

WARD WARD

0/30.2

0.99

SIMPLE

2.3/23.3 0/11.6 9.3/14 0/0 7/34.9

0.79 0.89 0.52 0.99 0.75

WARD SIMPLE WARD WARD SIMPLE

1.6/43.5 3.2/1.6 14.5/1.6 16.1/0 0/0 4.8/32.3

0.80 0.88 0.81 0.51 0.95 0.83

WARD SIMPLE WARD SIMPLE WARD WARD

50 (11) 50 (28) 5 (30) 500 (36) 500 (43) 50 (13)

(Table 4). The FCM types are characterized by the average of local assemblages within each cluster group. Richness of these assemblages ranged from 21 to 40 (H′min: 0.53–2.66) and dominance varied widely, although white perch and gizzard shad often accounted for more than a third of assemblage abundance. The islands zone displayed the most complex pattern of fish assemblages (Fig. 3). Temperate basses, gizzard shad, freshwater drum, and yellow perch were common in most of these assemblages, but relative dominance varied among these and other species (Table 4). In general, broad areas of the Open Water COZ were predicted to support similar fish assemblages, but there was greater variety of types within shallower COZs of the southern and western coasts (Fig 3). Aggregation across spatial scales Predicted biotic (and abiotic) conditions at any spatial scale of the classification are determined by aggregation of conditions within the framework containers of the next finer scale (Fig. 7). The FCM assemblages are aggregations (means) of the local assemblages within each AHA. Characteristics of Aquatic Ecological Systems (AES) are determined by averaging fish assemblages of the FCMs within each framework zone (COZ). The next larger scale unit of the

500 500 5,000 5,000 500 50

500 (33) 50 (49) 500 (59) 50,000 (62) 5,000 (60) 5,000 (24) 5 (30)

1 (10) 5,000 (61) 500 (61) 500 (47) 5,000 (62) 500 (37)

fish-based classification is the ecological lake unit (ELU) and its characteristics were determined by aggregating the fish assemblages of each AES within the Western Lake Erie ALU. For example, the two Sandusky Bay FCMs form the Sandusky Bay AES and the 16 Crane Creek FCMs form the Crane Creek AES. These two Aquatic Ecological Systems combine with the other nine AESs to form the Western Lake Erie ELU. Fisheries Conservation and Management units (FCM) may be best suited for identifying areas in need of specific “on-the-ground” management or assessment efforts. Aquatic Ecological Systems have been identified as appropriate units for addressing many broad-scale conservation and assessment issues in lotic systems (Higgins et al., 2005; Sowa et al., 2007) and may serve a similar purpose in coastal systems. Each AES may be a single or group of COZs predicted to support similar fish assemblages (Table 4). Cluster analysis indicated that each Western Lake Erie AES contained a distinct fish assemblage. Ecological lake units and higher scale elements are valuable for largescale assessment and planning. To illustrate the aggregation process and highlight some classification features, we describe here a few example assemblage patterns from three of the general AES types, representing an embayment (Sandusky Bay), an open coast (Crane Creek), and the islands area. The primary compositional differences in

764

J.E. McKenna Jr., C. Castiglione / Journal of Great Lakes Research 36 (2010) 757–771

Fig. 4. Predicted distributions of (a) alewife, (b) smallmouth bass, (c) yellow perch, and (d) minimum diversity.

Table 4 Summary of hydrospatial framework and habitat classification based on modeled fishes. Mean diversity is the average of the minimum diversity estimated for each assemblage of modeled species and is weighted by the area occupied (spatial cells) by each. Dominant taxa are those that together account for ≥ 75% of predicted fish abundance. Dominant species: DRUM = freshwater drum, EMRL = emerald shiner, GIZZ = gizzard shad, MIMC = mimic shiner, MRON = Morone spp., PRCH = yellow perch, SFSH = spotfin shiner, SMLT = rainbow smelt, WBAS = white bass, and WHPR = white perch. Rare species: BUFF = bigmouth buffalo, CRAP = black crappie, GSUN = green sunfish, JOHN = Johnny darter, PUNK = pumpkinseed, SAND = sand shiner, and WFSH = lake whitefish. The average aquatic habitat area code (each digit indicating the class of fetch, distance to large river, SAV presence, and depth) is provided below each COZ name. COZ

Spatial cells

AHA units (types)

FCM units (types)

Mean diversity

Dominant taxa

Rarest modeled species

Crane Creek (1101) Deepwater (3203) Detroit River (1102) Erie (1101) Estral Beach (1101) Huron (1102) Islands (1202) Maumee Bay (1102) North Shore (2102) Open Water (2202) Sandusky Bay (1102) Scott Point (1202)

3,051 11,036 1,051 1,818 992 372 5,374 1,258 1,192 27,511 2,161 986

26 (6) 55 (10) 11 (2) 32 (3) 30 (5) 29 (3) 141 (13) 29 (3) 64 (10) 298 (12) 2 (2) 37 (6)

16 (8) 26 (16) 14 (5) 10 (5) 14 (8) 6 (5) 102 (27) 6 (6) 23 (11) 71 (31) 2 (2) 18 (12)

1.98 1.21 1.73 2.07 2.00 1.79 1.30 1.94 1.59 1.17 1.86 1.90

MRON, GIZZ, WHPR, EMRL, DRUM, WHPR, WBAS, SMLT, GIZZ, PRCH WHPR, MRON, GIZZ, PRCH, WBAS GIZZ, MRON, WHPR, EMRL, DRUM, SFSH MRON, WHPR, GIZZ, EMRL, PRCH, DRUM MRON, GIZZ, WHPR, EMRL WHPR, GIZZ, WBAS GIZZ, MRON, WHPR, EMRL WHPR, MRON, PRCH, GIZZ, WBAS WHPR, GIZZ, PRCH MRON, WHPR, GIZZ MRON, WHPR, GIZZ, EMRL, MIMC

WFSH, JOHN, BUFF JOHN, WFSH, GSUN JOHN, WFSH, PUNK JOHN, WFSH, BUFF WFSH, JOHN, BUFF WFSH, JOHN, BUFF WFSH, JOHN, BUFF JOHN, WFSH, SAND PUNK, WFSH, JOHN JOHN, BUFF, WFSH WFSH, JOHN, CRAP WFSH, BUFF, JOHN

J.E. McKenna Jr., C. Castiglione / Journal of Great Lakes Research 36 (2010) 757–771

765

Fig. 5. Stacked bar chart comparing predicted species composition within each COZ of the Western Lake Erie ALU. Predicted H′min is indicated by the solid line; error bars indicate the full range of predicted values.

average fish assemblages among these AESs were due to substantial differences in proportions of the dominant taxa, white perch, gizzard shad, Morone spp., yellow perch, white bass, and freshwater drum (Fig. 5). The proportion of uncommon species was also different in most cases. Minimum fish diversity (H′min) differed among these units as well, but the 95% confidence intervals were generally wide. The rarest species (among those modeled) expected to be present in these AESs were usually, bigmouth buffalo Ictiobus cyprinellus, black crappie Pomoxis nigromaculatus, green sunfish Lepomis cyanellus, johnny darter Etheostoma nigrum, pumpkinseed Lepomis gibbosus, sand shiner Notropis ludibundus, and lake whitefish Coregonus clupeaformis (Table 4). Example Aquatic Ecological Systems Embayment: Sandusky Bay Within the basic framework, the Sandusky Bay COZ was composed of two AHAs, containing a total of 2161 local sites (i.e., spatial cells) (Fig. 3). Analysis of predicted fish assemblages indicated that each AHA supported a different fish assemblage, and thus each was also a single distinct Fisheries Conservation and Management unit (FCM). The westernmost fish assemblage at the head of the bay (FCM type 35, Table 5) had more white perch and less gizzard shad than the other near the mouth of the bay (type 36, Fig. 6a). Smallmouth bass abundance was high in both of these FCMs compared with most others within the study area (89th and 95th percentiles, Table 5). The H′min was N1.9 for the predicted assemblages of both FCMs. The basic hydrospatial framework unit characteristics showed that, relative to the entire Western Lake Erie ALU, Sandusky Bay habitats are generally shallow, close to a river, have little SAV, and short fetches (Table 4). In addition to these basic conditions, habitat attributes used in the NN models showed that most habitat values were lower than average for the ELU, but River Density, Distance to a Protected Wetland, and Temperature CV were higher than average (Fig. 8a). When this COZ as a whole is characterized by the average fish assemblage, it becomes the Sandusky Bay AES. Open coast: Crane Creek The Crane Creek COZ is composed of 26 AHAs of six types. Assigning predicted fish assemblages to these areas and analyzing the resulting distributions indicated that these habitat units supported eight distinct fish assemblages (i.e., FCM types) distributed within 16 patches (FCMs) (Fig. 3). The most extensive FCMs were arranged in

bands from inshore to offshore. The most inshore of these occupied some of the coastal estuaries1 (FCM types 2 and 4, Table 5) and had higher total fish abundance (dominated by Morone spp.) than the other widespread assemblages (Fig. 6b). The other two widespread assemblages (FCM types 1 and 3) also had large components as Morone spp., but there was a trend of increasing gizzard shad abundance moving offshore and the middle band (FCM types 3 and 7; Table 5) had a large minnow component (mainly emerald shiner Notropis atherinoides, spotfin shiner Cyprinella spiloptera, and spottail shiner Notropis hudsonius). Both the middle and offshore (FCM type 1) bands supported relatively high abundances of smallmouth bass (91th percentile and 96th percentile). A small pocket of FCM type 6 intruded from the Open Water AES into the northeast section of the Crane Creek AES and supported the highest fish abundance, dominated by white perch. The H′min diversity of the Crane Creek FCMs ranged from 1.56 to 2.38 and types 1 and 3 were most diverse. Like Sandusky Bay, Crane Creek habitats are generally shallow, near coastal tributaries, and have short fetches, but distances to larger rivers and wetlands are generally less than average and they support some of the most extensive SAV growth (Table 4, Fig. 8b). Islands The islands COZ is composed of a high diversity of habitat types and fish assemblages. There were 141 AHAs of 13 types in the islands zone. Analysis of the predicted fish assemblages assigned to those areas indicated that these habitat units supported 27 distinct fish assemblages (FCM types) distributed within 102 patches (Table 4). The four most extensive FCMs (types 9, 10, 11, and 12) were located in the large expanses of water between the islands and occupied 72% of the islands AES area (Fig. 7). These habitats supported fish assemblages of moderate overall abundance and high species richness (40) (Fig. 6c); each was dominated by white perch. Two FCM types had assemblages with predicted total fish abundance N50,000 (types 24 and 34) but were rare, occupying only 1–4 FCMs (patches). The FCM types 11, 14, 20, 25, 29, and 33 supported assemblages with relatively high abundances of yellow perch or smallmouth bass (≥90th percentile). The abundant smallmouth bass areas were associated with the southern islands, while the abundant yellow perch areas were associated with more open waters. The FCM types 9, 10, 1 Habitats where tributary and Great Lake waters mix most intensively and lake levels affect tributary conditions.

766

J.E. McKenna Jr., C. Castiglione / Journal of Great Lakes Research 36 (2010) 757–771

Table 5 Characteristics of FCMs within selected COZs, including FCM type, area occupied by each type, number of habitat patches supporting each type, mean predicted abundance of fishes in patches of each type, predicted minimum fish diversity (H′min), and relative abundance of alewife, yellow perch, and smallmouth bass. COZ

Fish assemblage type

Area (km2)

No. FCMs

Mean fish abundance (1000 m-2)

Alewife abundance (percentile)

Yellow perch abundance (percentile)

Smallmouth bass abundance (percentile)

H′min

Crane Creek Crane Creek Crane Creek Crane Creek Crane Creek Crane Creek Crane Creek Crane Creek Islands Islands Islands Islands Islands Islands Islands Islands Islands Islands Islands Islands Islands Islands Islands Islands Islands Islands Islands Islands Islands Islands Islands Islands Islands Islands Islands Sandusky Bay Sandusky Bay

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 5 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36

96.2 46.4 38.3 12.8 11.2 9.6 6.6 1.1 130.0 68.9 45.1 39.9 28.6 16.1 14.4 11.1 9.0 7.9 4.7 3.2 2.0 1.8 1.3 1.3 1.2 0.9 0.8 0.7 0.4 0.4 0.4 0.3 0.2 0.1 0.1 104.3 52.9

2 5 2 2 1 1 2 1 13 3 10 11 12 4 3 1 4 2 2 3 6 2 2 3 3 2 3 2 2 1 4 1 1 1 1 1 1

3,800 5,600 2,600 6,500 8,500 25,200 2,700 4,700 39,100 11,500 20,500 28,500 4,300 1,900 26,900 3,600 2,500 11,100 1,500 15,400 8,500 4,500 14,000 15,300 52,600 6,700 21,900 3,600 18,800 13,600 1,800 2,600 7,000 6,600 5,400 6,600 5,400

32% 35% 38% 33% 62% 95% 0% 36% 90% 91% 88% 85% 52% 23% 98% 64% 28% 96% 59% 86% 62% 64% 94% 84% 0% 60% 100% 21% 93% 22% 40% 64% 0% 41% 31% 41% 31%

31% 25% 17% 35% 44% 70% 10% 19% 72% 49% 93% 81% 36% 11% 74% 14% 33% 16% 30% 90% 44% 12% 0% 53% 67% 38% 0% 32% 54% 46% 40% 28% 23% 37% 21% 37% 21%

96% 85% 91% 17% 73% 41% 15% 86% 58% 35% 52% 46% 62% 99% 53% 0% 72% 26% 37% 51% 73% 0% 19% 44% 16% 100% 27% 0% 25% 93% 0% 0% 0% 95% 89% 95% 89%

2.38 1.56 2.37 1.62 2.15 1.66 2.23 1.919 1.39 1.99 1.77 1.40 2.15 2.48 1.66 1.75 2.60 1.98 2.66 2.03 2.15 1.44 1.87 2.20 0.93 1.15 1.97 2.20 2.17 1.49 2.03 2.01 1.20 1.94 1.99 1.94 1.99

18, and 22 are expected to support a relatively high abundance of alewife in open water areas of this AES. Nearly all of the highest diversity (H′min N 2.2) FCMs (types 14, 17, 19, 23, and 27) were closely associated with islands and shoal areas. The islands AES habitats are generally deeper than average, far from coastal tributaries, have moderate fetches, and support little SAV growth. In fact, most other habitat conditions had higher than average values (Fig. 8c). Model predictions indicate that certain habitats should support high abundances of some species, such as smallmouth bass. Although our models do not identify cause and effect, fish–environment correlations can be examined. For example, a signature of habitat conditions (Fig. 9) indicates that smallmouth bass should be most abundant where fetch, distance to nearest large river, distance to open or delta wetlands, amount of shoreline modification, water depth, and Secchi depth are lower than average; river density, distance to a protected wetland, SAV cover, and water temperature and temperature variability were higher than average; and sediment was softer than average. This is consistent with what is known about appropriate smallmouth bass habitat (Scott and Crossman, 1973). Similar associations between fish predictions and habitat (or other ancillary information) can be examined for any species, assemblage, or habitat unit. Discussion The above examples demonstrate the ability of the classification system (i.e., the framework and its implementation) to highlight the

diversity of habitats that exist (or may exist) in a coastal aquatic system and accommodate all space comprising that system, across the range of spatial scales. However, our example focuses on summer daytime fishes (collected by trawl or electrofishing) and habitat units and conditions that they perceive (at least in a statistical sense). Distribution is, of course, not the only aspect of the ecology of Great Lakes fish and aquatic organisms, others include, diel shifts in assemblages and predatory activity, seasonal assemblage changes, and long-term alterations due to exotic species introductions and climate changes. Examinations based on fish collections with different types of gear or in different seasons would produce different results. The hydrospatial framework provides objectively defined habitat patch boundaries, but the implementation redefines those boundaries to identify important habitat units and conditions appropriate for any of these classification issues (or others). This dynamic definition of spatial units accommodates the variety of habitat patches in a landscape (Kotliar and Wiens, 1990). General classification tool Natural systems are complex and include many biotic and abiotic continua. Classification simplifies this complexity and provides a practical means of addressing important ecological and management problems by breaking continua into manageable units with known characteristics and interrelationships (Busch and Sly, 1992; Higgins et al., 2005; Seelbach et al., 2006; Sowa et al., 2007; Riseng et al.,

J.E. McKenna Jr., C. Castiglione / Journal of Great Lakes Research 36 (2010) 757–771

767

Fig. 6. Stacked bar charts of predicted species composition of each FCM type within selected COZs of (a) Sandusky Bay, (b) Crane Creek, and (c) islands. Predicted minimum species diversity of each FCM assemblage is indicated by the solid line.

768

J.E. McKenna Jr., C. Castiglione / Journal of Great Lakes Research 36 (2010) 757–771

Fig. 7. Example of nested scales within the Western Lake Erie Ecological Lake Unit. A local site and Fisheries Conservation and Management unit are highlighted from the islands coastal or open water zone. The blue dots within the example FCM patch represent the centroids of the constituent local cells. Each Pie chart specifies the mean assemblage associated with each scale of the hierarchy.

2008). Our hydrospatial framework is built on an objectively defined set of generic spatial containers (with inherent location) that comprise the entire study area. These basic building blocks are relevant to, but generally independent of, specific biotic or habitat conditions and completely nest within the next larger scale. One benefit of this generic framework is that a “reasonably” manageable numbers of spatial units are available to characterize conditions at any given scale from the local cell to the Great Lake and larger. For example, 11 COZs (plus deepwater) comprise the Western Lake Erie ALU. One of these, the islands COZ, is composed of 102 AHAs. The most extensive AHA within the islands COZ is composed of 1562 local sites; the AHA shown in Fig. 7 consists of 12 local sites. These relatively small numbers are in contrast to the N56,000 spatial cells comprising the Western Lake Erie ALU and the numerous combinations of spatial units and scales. This general framework allows for generation of a variety of classifications by populating the spatial containers with biotic or abiotic information (or both) for specific applications, such as aquatic habitat quality, biodiversity, or the fish-based classification described above. Ancillary data not used to define the classification can also be associated with each spatial unit and used to assess conditions or deviations from potential conditions in any area. We have highlighted fetch, distance to the nearest large river, presence of SAV, and water depth in the above examples (because they

contribute to the definition of the AHAs), but also used other habitat variables (that contributed to the NN models) to describe each spatial unit type. Geospatial disturbance data could be applied as attributes, as well. In fact, any congruent geospatial data can be attributed to these units and examined. In the fish-based classification, in which the fish distributions are used to interpret how the lakescape is divided up, the resulting FCMs and AESs may serve as management and planning units for fish-related objectives (Higgins et al., 2005; Sowa et al., 2007); associated habitat conditions can be identified and described. The C_Gap system is adaptive and its elements can be redefined as more accurate and higher resolution data become available. Landscape ecology principals are embodied in this habitat classification system. The hydrospatial framework and resulting classifications manifest as patchwork mosaics of habitat units. Those habitat patches are different sizes and shapes, defined by the rules of the hydrospatial framework and the particular classification implemented. Together, the habitat patches comprise all space at each spatial scale. Characteristics at finer scales can then be aggregated into the larger spatial scales, making clearer the relationships and possible interactions with or effects on neighboring areas and vice versa. This system allows analysis of habitat contrasts, multi-scale context, connectivity, and other important aspects of a complete lakescape of habitat units (Kotliar and Wiens, 1990; Wiens, 2002).

J.E. McKenna Jr., C. Castiglione / Journal of Great Lakes Research 36 (2010) 757–771

769

Potential applications Fish predictions The NN models performed well and provided estimates of general abundance of 39 fish species at any location within the Western Lake Erie area, based on diurnal catch and habitat conditions at multiple spatial scales. The model predictions may be used to examine distribution patterns of optimal and marginal habitats for the predicted species at any scale from the local site to the ecological lake unit. The inherent difficulty of effectively modeling rare species (Thompson, 2004) prevented us from including them in our predicted species assemblages. Thus, the minimum species diversity index based on the predicted assemblages (H′min) is less rich than in the natural system. However, fish abundance predictions were made for all modeled species at each local site throughout the ALU, making the richness component a reflection of minimum beta (or possibly gamma) diversity and generally richer than would be expected at any given site. Despite this mixing of diversity scale, H′min provides a valuable index of potential biodiversity that is comparable from one location to another and across spatial scales. Prediction of two of the most abundant and widespread species presented a different problem. The Morone species catches were sometimes reported as the generic Morone spp. Because the presence of one species or the other (or both) was not known in all cases, we included the generic Morone spp. predictions along with those of white perch and white bass for each fish assemblage. This could inflate the estimate of the abundance that is likely to be supported. However, our log abundance scale masks this redundancy by generalization into broad classes, especially in the largest classes. The effect on minimum diversity was also minor, in most cases. Only 14 FCM types changed by ≥10% when the generic Morone spp. predictions were combined with those of white perch and white bass. All of these FCM types were rare and local (occupying b1 km2), except one (FCM type: 2). Our conclusions are generally unaffected by this redundancy, but it becomes an important consideration if one focuses on any of these rare assemblage types. Fig. 8. Signatures of habitat conditions in selected COZs ((a) Sandusky Bay, (b) Crane Creek, and (c) Islands), based on the suite of habitat variables used in the neural network models. Bars indicate the standardized deviations from the mean value for each variable within the Western Lake Erie Aquatic Lake Unit. Variable code names are defined in Table 1.

Natural resource managers are responsible for all habitats within their districts, but tools and data are rarely available to cope with all habitats. The inclusive nature of the C_Gap system satisfies that need for nearshore habitats. Deepwater is, of course, a critical habitat for the Great Lakes ecosystem and the natural extension of our classification system is into that realm. Future work is planned to adapt the C_Gap system to deepwater habitats.

Fig. 9. Signature of mean habitat conditions within spatial units of the three selected COZs expected to support high abundances of smallmouth bass (N 90th percentile). The ordinate shows standardized deviations from mean conditions for each variable, relative to the entire data set. Variable code names are defined in Table 1.

Biodiversity preserves Numerous efforts have been made to identify high priority habitat areas for protection or restoration (Nehlson, 1997; Higgins et al., 2005; Levin and Stunz, 2005; and Sowa et al., 2007, for example), but few such examples exist for the Great Lakes (Hartig et al., 1997). Examination of the distribution of our model-predicted minimum fish diversity (H′min) can reveal habitats expected to support particularly rich and diverse fish assemblages. At the AES level (equivalent to the framework COZ), the characteristic predicted fish assemblage of the islands AES was only moderately diverse (H′min = 1.3). However, examination at the FCM scale revealed that specific small areas around the islands harbored highly diverse fish assemblages. For example, the fish assemblage of FCM type 19 had a diversity of 2.66 (Table 5). While white perch was the most dominant member of the assemblage, the abundances of 20 other species were more evenly represented than in most other assemblages. This assemblage was also predicted to include the least common of the modeled species, such as, johnny darter, lake whitefish, black crappie, and orangespotted sunfish Lepomis humilis. That fish assemblage is predicted to occur in only three FCMs occupying an area 1.2% of the islands AES (around North Bass Island). This relatively rare habitat might be a good candidate for protection, which may take many forms (e.g., enhances harvest regulations, restricted habitat modification, or water usage in these or nearby units), depending on the valuable characteristics of these units and how they affect or are affected by adjacent habitats. Comparisons of biotic or abiotic conditions (i.e., modeled) and potential conditions among the FCMs may also help to prioritize areas in need of restoration. Uniqueness of habitat types, expected abundance of any particular species or group of taxa, or

770

J.E. McKenna Jr., C. Castiglione / Journal of Great Lakes Research 36 (2010) 757–771

other measures of habitat value may be similarly applied to assess units or groups of units within the lakescape habitat patchwork at any spatial scale. Fish management areas Smallmouth bass and yellow perch are two of the most important game fish in Lake Erie (Ryan et al., 2003). The model-predicted distributions of these species showed numerous areas of high and low abundance within the Western Lake Erie ELU. Several FCM types supported relatively high smallmouth bass abundances (N85th percentile) and were often distributed in fairly discreet patches or closely associated with islands. Yellow perch were predicted to occur in highest abundances in large patches, particularly within several open water FCM types (e.g., 11 and 20 [N90th percentile]) and among the islands. There was a patchy distribution of predicted smallmouth bass abundances, but the highest occurred in shallow coastal areas, including some of the islands units (e.g., Crane Creek types 1 and 3, islands type 14 and 29, and Sandusky Bay type 35 [N90th percentile]). Alewife is an abundant and important forage fish whose highest abundances were predicted to occur in bands of open water areas offshore from the Crane Creek AES and near some of the southernmost islands. Any of the other modeled species, such as walleye, white perch, or black crappie, may be examined in the same way. Fisheries may benefit from special management action in these areas and knowledge of optimum habitat locations is, in general, valuable for planning. Conclusions The C_Gap system includes a flexible multi-scale hydrospatial framework and associated models that represent a substantial step forward in our ability to organize and evaluate habitat in large aquatic systems. The hierarchical nature of the framework means that the characteristics of any spatial unit may be compared, contrasted, and related to any other in the system. The fish-based classification and associated models provide the capability to estimate the abundances and distributions of fishes throughout an extensive nearshore area and organize fish habitat into practical scale-sensitive units. However, our model predictions are for a snapshot in time and a particular season (open water). The predictions also represent the potential for any given habitat unit to support particular fish species or assemblages. This benchmark of habitat potential is unlikely to change in response to acute anthropogenic influences and provides a good measure against which to monitor changes associated with degradation or restoration efforts. The fish-based classification is just one of many potential applications of the C_Gap system. Any georeferenced biotic (or abiotic) group may be used to generate different classifications of aquatic habitat that would be appropriate for different ecological or management problems. As a decision-support tool, applications may include developing fisheries management plans, identifying effective biodiversity reserves, setting priorities for restoration efforts, and evaluating survey effort allocation. Our demonstration project is geographically limited to Western Lake Erie and focused on nearshore habitats. However, these methods should be applicable throughout the Great Lakes and may also be appropriate for our marine coasts. The framework can be extended to the entire Great Lakes once some basic habitat data are collected and computed. However, the availability of well-organized, georeferenced fish assemblage data is likely to be the greatest limitation to widespread application of the fish-based classification. Acknowledgements We are grateful to R. Thoma and OH EPA and R. Knight and OH DNR for making Lake Erie fish data available to us and to R. Haas MI DNR for providing additional fish data. We are indebted to S. Morrison, B.

Greer, M. Henry, and M. Slattery for their assistance with data collection and processing. We also wish to thank the editors and reviewers of this manuscript and the many others that have supported this effort. This is contribution no. 1605 of the USGS Great Lakes Science Center. References Allen, J.D., 1995. Stream ecology, structure and function of running waters. Chapman & Hall, London. Bain, M.B., Meixler, M.S., 2008. A target fish community to guide river restoration. River Res. Appl. 24, 453–458. Bolsenga, S.J., Herdendorf, C.E., 1993. Lake Erie and Lake St. Clair Handbook. Wayne State University Press, Detroit, MI. Busch, W.D.N., Sly, P.G., 1992. The development of an aquatic habitat classification system for lakes. CRC Press, Boca Raton, FL. Christie, W.J., 1974. Changes in the fish species composition of the Great Lakes. J. Fish. Res. Board Can. 31, 827–854. Clements, F.E., 1936. Nature and structure of the climax. J. Ecol. 24, 252–284. Connor, D.W., Allen, J.H., Golding, N., Howell, K.L., Lieberknecht, L.M., Northen, K.O., Reker, J.B., 2004. The marine habitat classification for Britain and Ireland. Version 04.05. JNCC, Peterborough1 861 07561 8. (Internet version);. www.jncc. gov.uk/MarineHabitatClassification. Cowardin, L.M., Carter, V., Golet, F.C., LaRoe, E.T., 1979. Classification of wetlands and deepwater habitats of the United States. FWS/OBS-79/31, December 1979, Reprinted 1992. Davis, L.S., Henderson, J.A., 1978. Many uses and many users: some desirable characteristics of a common land and water classification system. Classification, inventory and analysis of fish and wildlife habitat. US Fish and Wildlife Service, Washington, D.C.. FWS/OBS-78/76. Environment Canada, 1997. Lake Erie—Western Basin Substrate Types. Ontario Region Geomatics Unit—Atmospheric Environment Branch. Available from: Contact: Ian Gillespie. Ontario Ministry of Natural Resources. Gido, K.B., Falke, J.A., Oakes, R.M., Hase, K.J., 2006. Fish–habitat relationships across spatial scales in prairie streams. In: Hughes, R.M., Wang, L., Seelbach, P.W. (Eds.), Influences of landscape on stream habitats and biological assemblages: American Fisheries Society, Symposium 48, Bethesda, Maryland, pp. 265–285. Gleason, H.A., 1939. The individualistic concept of the plant association. Am. Midl. Nat. 21, 92–110. Goodyear, C.D., Edsall, T.A., Ormsby, D.M., Moss, G.D., Polanski, P.E., 1982. Atlas of the spawning and nursery areas of Great Lakes fishes, Volume IX Lake Erie. FWS/OBS82/52. US Department of the Interior, Fish, Wildlife Service, Washington, D.C. Great Lakes Commission (GLC), 2004. Great Lakes Coastal Wetland Inventory. . Available from: the Great Lakes Commission Wetland Consortium. Ric Lawson, Eisenhower Corporate Park, 2805 S. Industrial Hwy, Suite 100, Ann Arbor, MI 48104. (734)971-9135. http://www.glc.org/wetlands/inventory.html. Hartig, J.H., Zarull, M.A., Reynoldson, T.B., Mikol, G., Harris, V.A., Randall, R.G., Cairns, V.W., 1997. Quantifying targets for rehabilitating degraded areas of the Great Lakes. Environ. Manage. 21, 713–723. Herdendorf, C.E., 1992. Lake Erie coastal wetlands: an overview. J. Great Lakes Res. 18, 533–551. Higgins, J.M., Bryer, V.M., Lammert, M., FitzHugh, T., 2005. A freshwater classification approach for biodiversity conservation planning. Conserv. Biol. 19, 432–445. Hudson, P.L., Griffiths, R.W., Wheaton, T.J., 1992. Review of habitat classification schemes appropriate to streams, rivers, and connecting channels in the Great Lakes drainage basin. In: Busch, W.D.N., Sly, P.G. (Eds.), The Development of an aquatic habitat classification system for lakes. CRC Press, Boca Raton, FL, pp. 73–89. Knauss, J., 1978. Introduction to physical oceanography. Prentice-Hall, Inc., Englewood Cliffs, NJ. Kotliar, N.B., Wiens, J.A., 1990. Multiple scales of patchiness and patch structure: a hierarchical framework for the study of heterogeneity. Oikos 59, 253–260. Levin, P.S., Stunz, G.W., 2005. Habitat triage for exploited fishes: can we identify essential “Essential Fish Habitat?”. Estuar. Coast. Shelf Sci. 64, 70–78. Li, C.-Y., Kiser, K.M., Rumer, R.R., 1975. Physical model study of circulation patterns in Lake Ontario. Limnol. Oceanogr. 20, 323–337. Madden, C.J., Grossman, H.D., Goodin, K.L., 2005. Coastal and Marine Systems of North America: Framework for an Ecological Classification Standard: Version II. NatureServe, Arlington, Virginia. Matthews, W.J., 1998. Patterns in freshwater fish ecology. Chapman & Hall, New York. Maxwell, J.R., Edwards, C.J., Jensen, M.E., Paustian, S.J., Parrott, H., Hill, D.M., 1995. A hierarchical framework of aquatic ecological units in North America (NearcUc Zone). Gen. Tech. Rep. NC-176. St. Paul. MN: U.S. Department of Agriculture. Forest Service. North Central Forest Experiment Station. McKenna Jr., J.E., 2003. An enhanced cluster analysis program with bootstrap significance testing for ecological community analysis. Environ. Modell. Softw. 18, 205–220. McKenna Jr., J.E., 2005. Application of neural networks to prediction of fish diversity and salmonid production in the Lake Ontario basin. Trans. Am. Fish. Soc. 134, 28–43. McKenna Jr., J.E., McDonald, R.P., Castiglione, C., Morrison, S., Kowalski, K., PassinoReader, D., 2006. A broadscale fish–habitat model development process: Genesee Basin, New York. In: Hughes, R.M., Wang, L., Seelbach, P. (Eds.). Proceedings of the Symposium on Landscape Influences on Stream Habitats and Biological Assemblages, American Fisheries Society National Meeting, Madison, WI, 2005. pp. 533–554.

J.E. McKenna Jr., C. Castiglione / Journal of Great Lakes Research 36 (2010) 757–771 Mills, E.L., Casselman, J.M., Dermott, R., Fitzsimons, J.D., Gal, G., Holeck, K.T., Hoyle, J.A., Johannsson, O.E., Lantry, B.F., Makarewicz, J.C., Millard, E.S., Munawar, I.F., Munawar, M., O'Gorman, R., Owens, R.W., Rudstam, L.G., Schaner, T., Stewart, T.J., 2003. Lake Ontario: food web dynamics in a changing ecosystem (1970–2000). Can. J. Fish. Aquat. Sci. 60, 471–490. Minns, C.K., Meisner, J.D., Moore, J.E., Greig, L.A., Randall, R.G., 1995. Defensible methods for pre- and post-development assessment of fish habitat in the Great Lakes. A prototype methodology for headlands and offshore structures: Canadian Manuscript Report of Fisheries and Aquatic Sciences 2328. Munawar, M., 2003. State of Lake Ontario: past, present, and future. Ecovision World Monograph Series. Goodword Books Pvt. Ltd., New Delhi, India. National Fish Habitat Science and Data Committee (NFHSDC), 2008. A Framework for Assessing the Nation's Fish Habitat National Fish Habitat Action Plan. October 2008. National Oceanic and Atmospheric Administration (NOAA), 1994. Great Lakes Bathymetry, Daily Sea Surface Temperatures, ice cover. Available from: NOAAGLERL, 4840 S. State Rd., Ann Arbor, MI 48108. (313) 741-2120. http://www.glerl. noaa.gov/data/bathy/bathy.html. National Oceanic and Atmospheric Administration (NOAA), 1999. Great Lakes and St. Lawrence River Medium Resolution Vector Shoreline Data. . Available from: NOAAGLERL, 2205 Commonwealth Blvd., Ann Arbor, MI 48105. (313) 741-2148;. http:// www.glerl.noaa.gov/data/char/glshoreline.html. Nehlson, W., 1997. Prioritizing watersheds in Oregon for salmon restoration. Restor. Ecol. 5, 25–33. Ohio Department of Natural Resources (ODNR), 2003. Ohio Division of Wildlife, Lake Erie Fisheries 2002, Annual Status Report, Federal Aid in Fish Restoration F-69p, Ohio Department of Natural Resources, Division of Wildlife, Lake Erie Fisheries Unit. Ohio Environmental Protection Agency (OEPA), 2003. Biological Monitoring Data: Lake Erie. Ohio Environmental Protection Agency. Available from: Division of Surface Water, 50 West Town Street, Suite 700, Columbus, Ohio. Olden, J.D., Jackson, D.A., 2001. Fish–habitat relationships in lakes: gaining predictive and explanatory insight by using artificial neural networks. Trans. Am. Fish. Soc. 103, 878–897. Olden, J.D., Jackson, D.A., 2002. Illuminating the “black box”: a randomization approach for understanding variable contributions in artificial neural networks. Ecol. Model. 154, 135–150. Olden, J.D., Jackson, D.A., Peres-Neto, P.R., 2002. Predictive models of fish species distributions: a note on proper validation and chance predictions. Trans. Am. Fish. Soc. 131, 329–336. Olson, D.M., Dinerstein, E., Wikramanayake, E.D., Burgess, N.D., Powell, G.V.N., Underwood, E.C., D'amico, J.A., Itoua, I., Strand, H.E., Morrison, J.C., Loucks, C.J., Allnutt, T.F., Ricketts, T.H., Kura, Y., Lamoreux, J.F., Wettengel, W.W., Hedao, P., Kassem, K.R., 2001. Terrestrial ecoregions of the world: a new map of life on Earth. Bioscience 51, 933–938. Peterson, M., 2003. A conceptual view of environment–habitat–production linkages in tidal river estuaries. Res. Fish. Sci. 11, 291–313. Pickett, R.L., 1977. The Observed Winter Circulation of Lake Ontario. J. Phys. Oceanogr. 7, 152–156. Ricklefs, R.E., Miller, G.L., 2000. Ecology, 4th ed. W.H. Freeman and Company, New York.

771

Ripley, B.D., 1996. Pattern recognition and neural networks. Cambridge University Press, Cambridge. Riseng, C.M., Wang, L., Wiley, M.J., Rutherford, E., Brenden, T., 2008. State-of-the-Art Approaches for Assessment of Great Lakes Nearshore and Large River Fish Habitat. Final Report to the Great Lakes Fishery Trust, United States Environmental Protection Agency, and Great Lakes Fishery Commission. Ryan, P.A., Knight, R., MacGregor, R., Towns, G., Hoopes, R., Culligan, W., 2003. Fish– community goals and objectives for Lake Erie. Great Lakes Fishery Commission Special Publication 03-02. Schertzer, W.M., 2003. Physical limnology and hydrometerological characteristics of Lake Ontario with consideration of climate impacts. In: Munawar, M. (Ed.), State of Lake Ontario, past, present and future. Goodword Books Pvt. Ltd., New Delhi, India, pp. 3–57. Scott, W.B., Crossman, E.J., 1973. Freshwater fishes of Canada. Bulletin 184. Fisheries Research Board of Canada, 966 pp. Seelbach, P.W., Wiley, M.J., Baker, M.E., Wehrly, K.E., 2006. Initial classification of river valley segments across Michigan's Lower Peninsula. In: Hughes, R.M., Wang, L., Seelbach, P. (Eds.). Proceedings of the Symposium on Landscape Influences on Stream Habitats and Biological Assemblages, American Fisheries Society National Meeting, Madison, WI, 2005. pp. 25–48. Shannon, C.E., Weaver, W., 1949. The mathematical theory of communication. University of Illinois Press, Urbana, Illinois. Sly, P.G., Busch, W.D.N., 1992. A system for aquatic habitat classification of lakes. In: Busch, W.D.N., Sly, P.G. (Eds.), The Development of an Aquatic Habitat Classification System for Lakes. CRC Press, Boca Raton, FL, pp. 15–26. Sowa, S.P., Diamond, D.D., Abbitt, R., Annis, G., Gordon, T., Morey, M.E., Sorensen, G.R., True, D., 2005. A Gap Analysis for Riverine Ecosystems of Missouri. Final Report, submitted to the USGS National Gap Analysis Program. Sowa, S.P., Annis, G., Morey, M.E., Diamond, D.D., 2007. A gap analysis and comprehensive conservation strategy for riverine ecosystems of Missouri. Ecol. Monogr. 77, 301–334. Spalding, M.D., et al., 2007. Marine ecoregions of the world: a bioregionalization of coastal and shelf areas. Bioscience 57, 573–583. ter Braak, C.J.F., 1995. Ordination. In: Jongman, R.H.G., ter Braak, C.J.F., Van Tongeren, O.F.R. (Eds.), Data Analysis in Community and Landscape Ecology. Cambridge University Press, Cambridge, United Kingdom, pp. 91–173. ter Braak, C.J.F., Smilauer, P., 2002. CANOCO Reference Manual and CanoDraw for Windows User's Guide: Software for Canonical Community Ordination (version 4.5). Available from Microcomputer Power, Ithaca, New York. Thompson, W.L., 2004. Sampling Rare or Elusive Species: Concepts, Designs, and Techniques for Estimating Population Parameters. Island Press, Washington, D.C. U.S. Geological Survey (USGS), 2004. National hydrography dataset: August, 2004. Available from: http://nhd.usgs.gov/. Ward Systems, 2000. NeuroShell2, Release 4.0. Ward Systems Group, Inc., Frederick, Maryland. (http://www.wardsystems.com). Wetzel, R.G., 2001. Limnology, 3rd ed. Academic Press, San Diego, CA. Wiens, J.A., 2002. Riverine landscapes: taking landscape ecology into the water. Freshwat. Biol. 47, 501–515.