Predictability of In-Stream Physical Habitat for Wisconsin ... · correlation analyses of fish assemblage metric data with both predicted and observed values. Prediction of physical

Research Article • DOI: 10.2478/remc-2013-0003 • REMC • 2013 • 11–24

RipaRian Ecology and consERvation

11

* E-mail: [email protected]

Predictability of In-Stream Physical Habitat for Wisconsin and Northern Michigan Wadeable Streams Using GIS-Derived Landscape Data

1Institute for Fisheries Research, Michigan Department of Natural Resources and University of Michigan, 1109 N University, Ann Arbor, MI 48109

Current Address: International Joint Commission, Great Lakes Regional Office, P.O Box 32869, Detroit, MI 48232

2Department of Fisheries and Wildlife, Michigan State University, 480 Wilson Road, East Lansing, MI 48824

3Wisconsin Department of Natural Resources, 2801 Progress Road, Madison, WI 53716

Lizhu Wang1*, Travis Brenden2, John Lyons3, Dana Infante2

Received 30 July 2012Accepted 12 December 2012

AbstractQuantifying spatial patterns of physical and biological features is essential for managing aquatic systems. To meet broad-scale habitat assessment and monitoring needs, we evaluated the feasibility of predicting 25 in-stream physical habitat measures for wadeable stream reaches in Wisconsin and northern Michigan using geographic information system (GIS) derived stream network and landscape data. Using general additive modeling and boosting variable selection, predictions of reasonable accuracy were obtained for 10 widely used in-stream habitat measures, including bankfull depth and width, conductivity, substrate size, sand substrate, thalweg water depth, wetted width, water depth, and width-to-depth ratio. Biased predictions were obtained for habitat measures such as bank erosion, large woody debris, fish cover, canopy shading, and substrate embeddedness. Model predictions for many commonly-used habitat variables were judged acceptable based on several criteria, including correspondence between prediction errors and observed inter-annual and inter-site variability in habitat measures and agreement in correlation analyses of fish assemblage metric data with both predicted and observed values. Prediction of physical habitat variables from widely available GIS datasets represents a potentially powerful and cost-effective approach for broad-scale (e.g., multi-state, national) assessment and monitoring of in-stream conditions, for which direct measurement is largely impractical because of resource limitations.

KeywordsPhysical habitat • Modeling • Landscape • Stream network • Habitat prediction • Regional database

© Versita Sp. z o.o.

1. Introduction

Hierarchical frameworks that describe rivers and streams as

nested series of habitats ranging from local- (in-stream) to

basin-level (stream network, catchment) scales have become

relatively commonplace due to wide acceptance that stream

biota are concomitantly influenced by in-channel structure

and processes, stream hydrological network descriptors, and

landscape characteristics [1-3]. Stream network and catchment

characteristics can affect aquatic biota directly by influencing

network connectivity and hydrology, sediment, and thermal

regimes, as well as indirectly by influencing water quality, energy

source, substrate composition, and channel morphology [4-6].

As a result, complete understanding of stream-biota habitat

relationships necessitates the assessment of fluvial habitat

across multiple spatial scales.

The increased availability of regional geographic information

system (GIS) databases and technological advancements has

enhanced our ability to remotely capture stream network and

catchment scale information [7]. Among the notable achievements

that have resulted from this increased capacity has been the

development of the Great Lakes regional river database and

classification system (GLRRDACS), which includes all streams

and rivers in Illinois, Michigan, and Wisconsin [8], and the National

Hydrography Dataset plus (NHDPlus), which includes all streams

and rivers in the conterminous United States [9]. These databases

divide stream networks into confluence-to-confluence stream

reaches, with each reach having delineated local and network

catchments. Reaches are the smallest spatial unit for these stream

networks, but they can be easily combined into segments that

possess similar physicochemical and biological characteristics

[10,11]. The local and network catchments provide the basis for

attributing landscape-scale information to the streams, which

can be used to create complex, multi-scale, spatially-explicit

datasets that are essential for comprehensive evaluation of stream

conditions [7].

Brought to you by | Michigan State UniversityAuthenticated | 35.13.44.22

Download Date | 4/30/14 9:17 PM

L. Wang et al.

12

aforementioned GLRRDACS and NHDPlus frameworks provide

a potential solution to this problem as a result of their using

confluence-to-confluence stream reaches as the basic spatial

unit. Past research has established that a sampling site on a

specific confluence-to-confluence stream reach can reasonably

be considered representative of conditions throughout the reach

[6,10,11,22]. Thus, the GLRRDACS and NHDPlus frameworks

provide a simple basis for forecasting in-stream physical habitat

conditions from models developed using landscape scale data

to entire stream networks by virtue of their underlying structure,

in particular their reliance of stream reaches as the basic spatial

and the delineation of local and network catchments for all

reaches.

The overall goal of this study was to evaluate the feasibility

of modeling major in-stream physical habitat measures for

wadeable streams in Wisconsin and northern Michigan. We

defined wadeable streams as stream reaches with network

catchment areas < 1,600 km2 or stream orders < 5th order [23].

The specific objectives of the project were to (1) develop models

for predicting major in-stream physical habitat measures that

are commonly used by stream managers; (2) evaluate the fit and

prediction performance of the developed models by comparing

difference between modeled and observed data from the model-

development dataset and between modeled and observed data

from a model-validation dataset that had temporal and spatial

variation measures; and (3) evaluate the usefulness of the model

predictions by comparing habitat and fish assemblage metric

relationships between predicted and observed habitats.

2. Materials and Methods

2.1 Great Lakes Regional River Database and Classification System (GLRRDACS)

We conducted our study using the Michigan and Wisconsin

stream network databases, which are part of the GLRRDACS.

Streams identified from the 1:100,000 scale National Hydrography

Dataset (NHD) were divided into individual stream reaches

defined from headwater to the first confluence, confluence to

confluence, confluence to lake/reservoir, or confluence to the

Great Lakes or the Mississippi River. The Michigan database

included 28,889 reaches and 77,972 kilometers of streams and

rivers, while the Wisconsin database included 35,799 reaches

and 87,053 kilometers of streams and rivers. For each reach,

local (i.e., all land areas draining directly into a stream reach)

and network (i.e., all upstream areas draining into a stream

reach by either overland or waterway routes) catchments were

delineated using a 1-arc second resolution National Elevation

Dataset available for the Great Lakes region. Additionally, we

delineated local and network buffers for each reach, where

buffers were defined as 75-m horizontal distances on either side

of each stream. See [8] for descriptions of how local and network

catchments and buffers were delineated.

A suite of landscape and stream network variables known

to influence local habitat conditions and fish assemblages were

attributed to each of the stream reaches. Landscape descriptors,

Despite the wide availability of stream network and

catchment datasets, the cost-effective measurement of

local habitat across entire states or multi-state regions

remains a challenge. Traditionally, in-stream physical habitats

have been assessed through on-site sampling of channel

geomorphic measures, hydrological and thermal regimes,

water characteristics (width, depth, and velocity), substrate

composition, channel hydraulics, pool-riffle complexity, in-

stream cover for fish, canopy shading, and bank and riparian

conditions [12-14]. Such direct assessment, however, can be

time consuming and expensive, which prohibits its application

across large areas. As a consequence, many regional stream

and river databases do not include in-stream physical habitat

measurements and this incompleteness can seriously affect their

overall usefulness. In-stream physical habitat measurements

are valuable for examining biota–habitat relationships (e.g.,

identification of limiting habitat for particular organisms),

assessing deviations in habitat conditions from natural states

due to human perturbation, measuring improvements that have

resulted from habitat enhancement and restoration activities, and

predicting distribution and/or abundance of aquatic organisms

[3,5,8]. In-stream physical habitat measures are also needed for

conducting stream classifications for formulating management

policies and enacting regulations [1,2,10,11].

One promising approach for obtaining local habitat

information across large regions is through predictive modeling

based on readily available landscape data [6]. Predictive

modeling for the purpose of quantifying in-stream physical

habitat is supported by the riverscape concept of landscape

factors constraining local habitat conditions [1,15] and by

numerous empirical studies that have demonstrated the

predictability of local habitat variables from stream network

and catchment characteristics [16-20]. Although there are many

examples in the scientific literature of in-stream physical habitat

models being developed, rarely have such models been used to

comprehensively predict habitat conditions across large regions.

In some cases, models have been developed from relatively

few observations [17,18,20], which may not be appropriate

for predicting habitat conditions across large regions because

measures may not be representative of all regional habitat types.

In other cases, the underlying purpose of the modeling efforts

has not been to assess physical habitat predictions; rather, it has

been to develop stream habitat classifications or to assess how

much variability in local habitat conditions can be partitioned to

different spatial scales [16,21].

One explanation as to why in-stream physical habitats have

not been routinely predicted across large regions is that many

studies treat individual sampling sites as the basic spatial unit.

Although the measurement of physical habitat and the landscape

data associated with an individual sampling site (usually

100-1,000 m in length) is likely an accurate representation of

the site’s condition, forecasting conditions beyond the sampled

stream sites is difficult because it is not clear how to scale the

predictions to the remainder of the stream network or how to

attribute landscape data to the remainder of the network. The



Prediction of stream physical habitat from landscape data

13

a good cross-section of conditions relative to the model

development dataset and thus were appropriate for verification

of model accuracy. These sampling reaches were not part of the

model-development dataset. Among these reaches, 58 sites

from 47 reaches were sampled in multiple years ( x = 4.7 years;

range 2 to 10 years) for assessing temporal variation, and 14

reaches were sampled at multiple sites (2-3 sites per reach) for

assessing within reach spatial variation. Bankfull depth, bankfull

width, and large woody debris measurements were not available

for the model-validation dataset.

2.4 Physical habitat model developmentWe used generalized additive modeling and boosting variable

selection to fit the local habitat prediction models. Generalized

additive modeling is a semi-parametric regression approach for

generating nonlinear response curves between dependent and

independent variables. Because it is not necessary to specify a

particular model equation, this modeling approach is useful in

cases where there is little information available describing how

independent variables are influenced by dependent variables

[27]. Boosting is an algorithmic approach for variable selection

that is rooted in the field of machine learning but which has

recently proven effective for fitting regression models in cases

where there are large sets of candidate predictor variables

[28-30]. Such situations (i.e., large numbers of candidate

predictor variables) frequently arise with geo-referenced

datasets, because of the ease of attributing information at a

variety of spatial scales.

The generalized additive models were fit in R (R Development

Core Team, 2010, R Foundation for Statistical Computing,

Vienna, Austria. http://R-project.org) using the “mboost”

package (Hothorn, T., Bühlmann, P., Kneib, T., Schmid, M.,

Hofner, B., 2010, Mboost: Model-Based Boosting, version

2.0-4, http://CRAN.R-project.org/package=mboost). All of the

153 landscape and stream network variables that were attributed

to the stream sites were used as candidate variables for fitting the

including catchment area, soil type and permeability, surficial

geology formation and texture, bedrock type and depth,

20-year July mean air temperature, 20-year mean precipitation,

catchment slope, and land use and cover within each of

the spatial scales were attributed to the reaches based on

available data for each state [8]. Stream channel descriptors,

including Shreve linkage numbers for each reach and for the

downstream reach that each reach flows into, reach gradient,

reach elevation, sinuosity, stream order, total upstream stream

length, and distances from upstream most headwaters and from

the Great Lakes or Mississippi River, dam density for upstream

network catchment, dam density per up- or down-stream stream

length and distance to the first upstream or downstream dam

were also calculated using ArcInfo functionalities (PC ARC/

GIS Version 8.2. Environmental System Research Institute,

Redlands, California, http://www.esri.com/software/arcgis).

July mean stream temperatures and stream flow exceedances

were predicted for each stream reach using statistical models

developed from measurements made at a subset of the reaches

[24,25]. See [8] for additional details regarding methods for

stream reach identification, spatial boundary delineation, source

data acquisition, and variable attribution to the stream reaches.

2.2 Model-development datasetStream sites representing all wadeable stream conditions

across the entire state of Wisconsin and the northern part of

Michigan were sampled for physical habitat during 1996-2003

(Figure 1). All stream sites were sampled between early June and

late September under low flow conditions. The length of each

site was approximately 35- to 40-times mean stream width or

a minimum distance of 100 m, which is of sufficient length to

characterize fish assemblages and generally encompasses

around three meander sequences [26]. Sites ranged in length

from 100 to 837 m ( x = 273 m). At each site, a suite of habitat

variables were measured or visually estimated following

established sampling protocols [14]. The lengths of riffles, pools,

and runs were measured over the entire length of the site. Stream

width and depth, bottom substrate composition, availability of

fish cover, bank conditions, and riparian vegetation cover were

measured along 10 to 20 transects spaced 2.0- to 3.5-times

mean stream widths apart. Altogether, 29 local habitat variables

were measured (Table 1). Statistical models were not developed

for four of the habitat variables (% algae substrate, % bedrock

substrate, % clay substrate, % macrophyte substrate) because

of a lack of contrast in observed values (i.e., most values were 0).

For stream reaches where multiple sites were sampled, the mean

of each habitat variable was used for each stream reach; this

resulted in a total of 286 stream reaches for which local habitat

measurements were available.

2.3 Model-validation datasetThe model-validation dataset consisted of 54 stream reaches

from southern Wisconsin that were sampled at multiple sites per

stream reach or in multiple years. Although the validation sites

were limited to southern Wisconsin, the streams represented

Figure 1. Maps of Michigan and Wisconsin showing stream sites for model development (filled circles) and sites for model validation (filled triangles).



L. Wang et al.

14

models to percent data, a binomial distribution was specified

as the family distribution. For all other local habitat measures,

a Gaussian distribution was specified as the family distribution.

For those habitat measures that were fit assuming a Gaussian

distribution, Box-Cox transformations of the dependent variables

local habitat prediction models. A random effect corresponding

to measurement year was also included as a candidate variable

for the models. The effects of the candidate variables, with the

exception of measurement year, were modeled using P-splines

with 20 interior knots and 4 degrees of freedom [30]. When fitting

Habitat variables Mean Range Median 1st Quartile 3rd Quartile

Calculated

Catchment area (km2) 134 1-1943 39 17 137

Channel gradient (m/1000m) 2.7 0-31 2 1 4

Channel sinuosity (ratio) 1.3 1.0-4.1 1.2 1.1 1.4

Linkage number (#) 14.3 1-273 4 2 14

Stream order (#) 2.4 1-5 2 2 3

Observed

Bankfull depth (cm) 119 31-316 109 81 136

Bankfull width (m) 12 1-64 9 6 16

Bank erosion (%) 16 0-88 11 3 24

Buffer vegetation (%) 89 0-100 100 91 100

Canopy shading (%) 37 0-99 33 12 58

Conductivity (µS/cm) 461 1-1902 401 169 680

Dissolved oxygen (mg/l) 8.4 1.7-15 8.7 7.6 9.2

Fish cover (%) 13 0-96 9 4 16

Large substrate (%) 40 0-99 35 11 65

Large woody debris (%) 3 0-25 2 0 4

Pool (%) 11 0-100 5 0 15

Riffle (%) 12 0-67 6 0 20

Run (%) 77 0-100 82 65 96

Sediment depth (cm) 10 0-89 6 2 13

Small substrate (%) 55 0-100 55 30 78

Substrate algae (%)* 4 0-8 0 0 3

Substrate bedrock (%)* 1 0-48 0 0 0

Substrate boulder (%) 4 0-33 1 0 5

Substrate clay (%)* 4 0-98 0 0 3

Substrate cobble (%) 14 0-69 8 1 23

Substrate detritus (%) 5 0-53 2 0 5

Substrate embeddedness (%) 59 1-100 64 33 88

Substrate gravel (%) 22 0-72 21 8 33

Substrate macrophyte (%)* 6 0-81 0 0 7

Substrate sand (%) 36 0-100 29 14 52

Substrate silt (%) 14 0-96 7 2 19

Thalweg water depth (cm) 49 10-121 45 31 62

Wetted width (m) 8.7 1-58 6 4 11

Water depth (cm) 37 8-92 35 23 48

Width-to-depth ratio 18 3-86 14 10 21

Table 1. Summary statistics for the streams that were used to develop the local-scale habitat prediction models. Summary statistics are for GIS-calculated variables and the local-scale habitat measures for which prediction models were developed. An * indicates a variable was removed from further analyses because of the preponderances of zero values (Median=0).




15

observed and predicted habitat values for each habitat variable

using the model-development dataset and the model-validation

dataset separately using the equation

(S(|Observed-Predicted|/Observed)*100)/N, (1)

where N equals the number of equations. We also calculated

the mean absolute relative error between the mean of the

observed habitat values among sampling sites of a stream reach

and the values from each sampling site for the same reach of

the model-validation dataset (i.e., spatial variation) using the

equation

(S(|Mean-Observed|/Mean)*100)/N. (2)

We also calculated the mean absolute relative error between

the mean of the observed habitat values from different years of a

sampling site and the values for each sampling year for the same

site of the model-validation dataset (i.e., temporal variation)

using equation 2.

The second approach was to compare relationships

between the predicted and observed local habitat measures with

fish assemblage metrics. We obtained fish data for 213 stream

reaches that were part of our model-development and validation

datasets from the Departments of Natural Resources of Michigan

and Wisconsin. These data were collected using backpack or

tow-barge electro-fishing units from late May to late September

between 1997 and 2002 on stream sites where physical habitat

data were also collected. The lengths of streams sampled

ranged from 100 to 960 m with larger streams having longer

sampling distances. Fish data were collected using single-pass

electrofishing to collect all fish observed, and all captured fish

were identified, enumerated, and weighed in the field. From the

collected fish data, we calculated five metrics that are known to

be affected by physical habitat characteristics: species richness,

total fish abundance, percent of lithophilic spawning individuals,

percent of intolerant individuals, and darter species richness

[5,26]. Prediction performance of the local habitat models was

evaluated by comparing Spearman’s rank correlation coefficients

calculated between the five fish metrics and the predicted and

observed local habitat measures. Because of the number of

correlation analyses that were conducted (5 fish metrics × 25

local habitat measures = 125 correlation analyses), a Bonferroni

correction was used to help ensure that the experiment-wise

error rate for the significance tests was maintained at 0.05.

3. Results

Of the 153 stream network and landscape variables that

were available as candidate predictors for the local habitat

prediction models, 143 were included in at least one of the

fitted models. The number of predictor variables included in

each of the models ranged from 1 to 43 with an overall mean

of 20 variables per model (Table 2). Fit varied considerably

among the prediction models (Table 3, Figure 2). Based on the

were used to help the data meet assumptions of normality. The

number of boosting iterations used in fitting each of the local

habitat prediction models was determined using 10-fold cross

validation [30].

2.5 Evaluation of model fits and prediction performanceFits of the local habitat models were evaluated by using simple

linear regression to regress observed versus predicted habitat

measures [31,32]. The slopes and shifted intercepts from these

regression models were then tested against values of 1.0 and

the mean of the predicted values, respectively, as a check of

the similarity of predicted and observed means and individual

values [33]. Testing of the slopes and shifted intercepts from

the observed versus prediction regressions was accomplished

using a two one-sided test strategy (TOST), which is a form of

equivalence testing [33]. Equivalence tests are commonly used in

the biomedical fields for situations such as vaccination coverage

in different demographic groups [34] and are starting to be used

as a method for validating model predictions [33,35]. Testing of

the slopes and intercepts followed the TOST strategy of using

bootstrapping to construct the two one-sided 95% confidence

intervals around the regression slope and shifted intercept [33].

One of the key aspects of equivalence testing for model validation

is the specification of a region of equivalence or indifference for

the shifted intercepts and slopes, which distinguishes practical

equivalence from scientifically relevant differences [31,33].

It is this region of equivalence that the confidence intervals

for the shifted intercept and slope estimates are compared. If

the confidence intervals for the slope or shifted intercept lie

entirely within their respective regions of equivalence, then the

null hypothesis of dissimilarity is rejected [33]. We chose to set

the equivalence region at a fairly liberal rate ( y ± y ×0.45% for

the shifted intercept and 1.0 ± 0.45 for the slope) given that we

were attempting to model in-stream habitat features that are

oftentimes dynamic and difficult to measure accurately. The

equivalence testing was conducted in R using the “equivalence”

package (Robinson, A., 2010, Equivalence: provides tests and

graphics for assessing tests of equivalence, version 0.5.6,

http://CRAN.R-project.org/package=equivalenc).

We also calculated Theil’s partial inequality coefficients for

each of the modeled local habitat measures, which separate

total error into three components: Ubias, Uslope, and Uerror [31,32].

Ubias represents the proportion of total error associated with

mean differences between observed and predicted values. Uslope

represents the proportion of total error associated with deviance

of the slope from the 1:1 line. Uerror represents the proportion of

total error associated with the unexplained variance [31,32].

The prediction performance of the local habitat models was

evaluated using two approaches. The first approach was to

compare percent differences between observed and predicted

habitat values for the model-development dataset with the

percent differences between observed and predicted values for

the model-validation dataset, and with the percent differences in

temporal and spatial variations for the model-validation dataset.

We first calculated the mean absolute relative error between the



L. Wang et al.

16

from the observed versus predicted regressions could not be

rejected for any of the in-stream habitat variables, while the

null hypothesis of dissimilarity for the shifted intercepts was

rejected for only 9 of the 15 variables (Table 3). The R2 of the

observed versus predicted regressions for the models that

were considered to have poor fits was generally less than 0.45

with slope estimates greater than 1.4.

Based on the calculated Theil’s partial inequality

coefficients, most (>70%) of the error between observed and

predicted values for each of the local habitat measures was

due to unexplained variance (Uerror) (Table 3). The next largest

source of error differed among the local habitat measures.

For variables such as bankfull depth, large woody debris,

and detritus substrate, most of the remaining error was due

to differences between the slope of the fitted model and the

1:1 line (Uslope), whereas for variables such as pool habitat, riffle

habitat, run habitat, and gravel substrate, most of the remaining

regression of observed versus predicted values and the tests

of equivalence of the regression slopes and shifted intercepts,

the models for bankfull depth and width, conductivity, large

and small substrates, sand substrate, thalweg water depth,

wetted width, water depth, and width-to-depth ratio were

considered to yield satisfactory fits (Table 3). For these

models, the null hypotheses of dissimilarity for both the slopes

and shifted intercepts from the observed versus predicted

regressions were rejected (Table 3). For most of these models,

the R2 of the observed versus predicted regressions was

greater than 0.50 with slope estimates between 1.0 and 1.3

(Table 3). Conversely, the models for bank erosion, vegetative

buffer, canopy shading, large woody debris, cover for fish,

riffle, pool, run, sediment depth, substrate embededdness,

boulder substrate, cobble substrate, detritus substrate,

gravel substrate, and silt substrate were considered to have

poor fits. The null hypothesis of dissimilarity for the slopes

Table 2. Number of predictor variables by variable type that were selected through the boosting algorithm for each of the final local-scale habitat prediction models.

Landscape variable categories

Habitat variableBedrock

depthBedrockgeology

ClimateLandcover

Soilgeology

Streamnetwork

Surficialgeology

Totalvariables

Bankfull depth (cm) 6 7 5 14 0 3 8 43

Bankfull width (m) 3 6 5 8 0 1 5 28

Bank erosion (%) 4 0 5 3 3 2 8 25

Buffer vegetation (%) 3 3 1 9 0 0 3 19

Canopy shade (%) 0 0 0 2 0 2 0 4

Conductivity (µS/cm) 4 4 4 8 5 2 6 33

Fish cover (%) 6 1 2 3 0 1 5 18

Large substrate (%) 2 4 1 3 3 4 4 21

Large woody debris (%) 7 5 2 7 3 2 531

Pool (%) 1 4 4 3 1 3 11 23

Riffle (%) 0 1 0 2 2 3 1 9

Run (%) 2 5 3 3 2 2 5 22

Sediment depth (cm) 0 0 0 0 0 0 1 1

Small substrate (%) 3 2 3 3 3 2 7 23

Substrate boulder (%) 3 4 2 6 1 4 6 26

Substrate cobble (%) 3 3 0 2 2 2 3 15

Substrate detritus (%) 3 6 3 6 2 4 11 35

Substrate embeddedness (%) 1 2 1 3 3 4 115

Substrate gravel (%) 2 4 2 3 3 2 3 19

Substrate sand (%) 0 3 0 2 3 1 3 12

Substrate silt (%) 4 5 5 11 1 4 9 39

Thalweg water depth (cm) 4 4 2 5 2 3 1030

Wetted width (m) 3 5 3 11 1 1 2 26

Water depth (cm) 3 5 3 5 1 3 9 29

Width-to-depth ratio 3 6 1 7 2 3 5 27




17

errors between predicted and observed values for both the

model development and model validation datasets among the

variables that were judged to be predicted satisfactorily and those

judged to be predicted unsatisfactorily. For those variables that

were predicted satisfactorily, the mean absolute relative errors

between predicted and observed values ranged from 15 to 38%

for the model development dataset and 15 to 47% for the model

validation dataset (Table 4). Conversely for variables that were

judged to be predicted unsatisfactorily, the mean absolute relative

errors between predicted and observed values ranged from 11 to

90% for the model development dataset and 20 to 101% for the

model validation dataset (Table 4).

For the correlation analyses that evaluated prediction

performance of the local habitat models, there was a good

error was due to mean differences between observed and

predicted values (Ubias).

For the habitat variables that were judged to be predicted

satisfactorily (Table 4, Figure 2), the mean absolute relative

errors between the predicted and observed values for the model

development dataset were greater than the year-to-year and site-

to-site variation for the observed values of the model validation

dataset. However, when the temporal and spatial variability was

assessed together, the mean absolute relative errors for temporal-

spatial test, predicted versus observed for validation dataset, and

predicted versus observed for model-development dataset were

similar (Table 4). Similar results were obtained for those habitat

variables that were predicted unsatisfactorily (Table 4, Figure 3).

There were noticeable differences in the mean absolute relative

Table 3. Regression parameter estimates (Intercept [Int.] and Slope) and coefficients of determination for the regressions of observed versus predicted local habitat measures for the data used to fit the local habitat prediction models. An * next to the intercept and slope estimates indicates that the null hypothesis of dissimilarity for the shifted intercept and slope estimates from the regressions of observed versus predicted values was rejected at an a=0.05. Also shown are Theil’s partial inequality coefficients, which partitions the error between observed and predicted values into the proportion associated with mean differences between observed and predicted values (Ubias), the proportion associated with differences between the slope of the fitted model and the 1:1 line (Uslope), and the proportion associated with unexplained variance (Uerror).

Habitat variable Int. Slope R2 Ubias Uslope Uerror

Habitat variables predicted satisfactorily

Bankfull depth (cm) -0.3* 1.3* 0.72 0.036 0.129 0.835

Bankfull width (m) 0.4* 1.0* 0.86 0.047 0.007 0.946

Conductivity (µS/cm) -38.4* 1.2* 0.81 0.038 0.073 0.888

Large substrate (%) -0.1* 1.3* 0.48 0.064 0.044 0.891

Small substrate (%) -0.1* 1.2* 0.50 0.010 0.024 0.966

Substrate sand (%) -0.0* 1.1* 0.48 0.028 0.012 0.960

Thalweg water depth (cm) -0.1* 1.3* 0.66 0.029 0.074 0.897

Wetted width (m) 0.5* 1.0* 0.84 0.050 0.005 0.945

Water depth (cm) -0.1* 1.3* 0.61 0.026 0.070 0.905

Width-to-depth ratio -0.3* 1.2* 0.58 0.068 0.025 0.907

Habitat variables predicted unsatisfactorily

Bank erosion (%) -0.0 1.7 0.35 0.084 0.075 0.842

Buffer vegetation (%) -0.9* 1.9 0.35 0.071 0.095 0.834

Canopy shade (%) -0.0* 2.5 0.42 0.070 0.109 0.820

Large woody debris (%) -0.0* 2.4 0.47 0.004 0.224 0.772

Fish cover (%) -0.0* 1.6 0.39 0.049 0.077 0.874

Pool (%) 0.0 1.5 0.38 0.107 0.054 0.839

Riffle (%) -0.0 2.0 0.20 0.143 0.053 0.804

Run (%) -0.5* 1.5 0.31 0.128 0.038 0.833

Sediment depth (cm) -1.2* 13.0 0.09 0.000 0.081 0.919

Substrate boulder (%) -0.0* 2.5 0.48 0.020 0.244 0.735

Substrate cobble (%) -0.0 1.7 0.41 0.117 0.096 0.788

Substrate detritus (%) -0.1 2.9 0.44 0.026 0.241 0.733

Substrate embeddedness (%) -0.3* 1.4 0.45 0.075 0.058 0.867

Substrate gravel (%) -0.1* 1.8 0.25 0.125 0.051 0.824

Substrate silt (%) -0.0 1.8 0.57 0.072 0.184 0.744



L. Wang et al.

18

4. Discussion

Our results demonstrate that important and commonly-used

local physical habitat variables, including bankfull depth and

width, water depth, thalweg depth, wetted width, large and

small substrates, conductivity, sand substrate, and width-to-

depth ratio, can be predicted satisfactorily from stream network

and landscape-scale catchment measures. It is expected

that these variables can be predicted from landscape-scale

datasets because they are strongly linked with factors such as

geology type, soil structure, topography, land cover, and climate

conditions [5,6,19]. Many of the variables for which satisfactory

fits were obtained are among those that are considered to be

the most influential physical habitat measures governing the

distribution and abundance of biological assemblages [5,36],

which underscores the potential utility of these habitat models

for broad-scale assessment and monitoring. Models for other

important and commonly used in-stream physical habitat

agreement in the relationships between fish metrics vs. observed

habitat values and the fish metrics vs. predicted habitat values

for habitat variables predicted satisfactorily (Table 5). As

expected, not all physical habitat measures correlated with all

selected fish metrics. Of the 125 pairs of correlations that were

considered, there were 57 significant correlations between

fish metrics and observed and/or predicted habitat measures.

Among those significant correlations, 75% of the correlation

pairs were in agreement between observed and predicted

habitat measures in their relationship with fish metrics indicated

by correlation significance and correlation directions. About

25% of the correlation pairs were in disagreement in that only

observed habitat measures were significantly correlated with fish

metrics. Among those correlations that were not in agreement

between fish metrics-observed and fish metric-predicted habitat

relationships, 12% were for habitat measures judged to be

predicted satisfactorily and 35% were for habitat measures

judged to be predicted unsatisfactorily.

Figure 2. Plots of observed versus predicted habitat values for models with satisfactory predictions for the model-development dataset.




19

has previously been demonstrated in several smaller-scale

studies, which together with our results suggest that statistical

modeling is a viable approach for assessment and monitoring of

habitat condition for a variety of system types. In a study of 53

southeastern Queensland streams, 5 of 21 local habitat variables

(width, sand, cobble, rocks, and large woody debris) were

predicted with R2s ranging from 0.22 and 0.65 using elevation,

stream order, distance from source, and longitude as predictor

variables [18]. The authors concluded that the predictive

performances of fish distribution models were similar when

using observed or predicted local habitat variables, although

fish distribution predictions were clearly biased at the extremes

of the habitat variable ranges. In a study of 76 New Hampshire

streams, 11 catchment and regional variables were used to

variables, such as bank erosion, large woody debris, fish cover,

canopy shading, substrate embeddedness, and riffle-pool

measures, were found to have poor predictive performances,

suggesting that they would be of limited utility for assessment

and monitoring. It is perhaps not surprising that variable such

as bank erosion, large woody debris, and fish cover could not

be adequately predicted solely from catchment-level and stream

network characteristics since these variables are strongly

influenced by riparian conditions [5,37]. We anticipate that the

performance of prediction models for these variables likely could

be improved by including measures from riparian-scale GIS

datasets among the candidate predictor variables.

The ability to accurately predict in-stream physical habitat

measures from stream network and catchment variables

Figure 3. Plots of observed versus predicted habitat values for models with unsatisfactory predictions for the model-development dataset.



L. Wang et al.

20

to acknowledge that there can be considerable measurement

error when direct sampling of local-scale habitats is conducted.

Accurate sampling of in-stream physical habitats is difficult

because habitat characteristics vary both laterally and

longitudinally. As well, in-stream physical habitat measures are

affected by a wealth of landscape-level factors, including climate

which fluctuates daily, seasonally, and annually. With standard

sampling protocols [12-14], individual measurements can deviate

from sample means by as much as 50% [38], as can inter-

annual variability in measurement as demonstrated in this study.

According to [38], the overall accuracy and seasonal variability of

in-stream physical habitat measures is strongly influenced by the

heterogeneity of habitat composition, with the greatest variability

observed for the most heterogenous environments. Thus, the

predict 64 local chemical and physical habitat variables using

several modeling approaches; the best performing models had

an error of 27% of the mean index value [20]. It was suggested

that other catchment variables that at the time were unavailable

to the authors could improve prediction performance. In a

study of 51 streams in the Upper Murrumbidgee catchment in

southeastern Australia, catchment area, stream length, relief

ratio, alkalinity, percentage of volcanic rocks, percentage of

metasediments, dominant geology, and dominant soil type were

found to provide sufficient information to classify 69% of sites

into appropriate site groups developed by direct sampling of

local stream habitat measures [17].

Although the use of prediction models impart some

uncertainty on the physical habitat measures, it is important

Table 4. Mean absolute relative errors in habitat values among years for the model-validation dataset (Temporal-test), among sampling sites within a reach for the validation dataset (Spatial-test), the sum of Temporal-test and Spatial-test (Temporal-spatial together) for the validation dataset, between predicted and observed for the validation dataset (Predicted-test), and between predicted and observed for the model development dataset (Predicted-observed). Bankfull depth and width and large woody debris were not measured for the validation data set.

Habitat variableTemporal-test

(n=58)Spatial-test (n=14)

Temporal-spatial together

Predicted-test (n=54)

Predicted-observed (n=286)


Bankfull depth (cm) -- -- -- -- 15

Bankfull width (m) -- -- -- -- 24

Conductivity (µS/cm) 6 6 12 15 28

Large substrate (%) 13 18 31 37 38

Small substrate (%) 12 15 27 30 26

Substrate sand (%) 20 24 44 47 38

Thalweg water depth (cm) 9 11 20 26 26

Water depth (cm) 10 12 22 24 27

Wetted width (m) 6 9 15 29 29

Width-to-depth ratio 9 13 22 24 31


Bank erosion (%) 30 33 66 80 53

Buffer vegetation (%) 12 13 25 36 11

Canopy shading (%) 12 24 36 78 57

Fish cover (%) 22 23 45 20 21

Large woody debris (%) -- -- -- -- 85

Pool (%) 45 51 95 71 90

Riffle (%) 23 31 54 87 74

Run (%) 9 11 20 21 19

Sediment depth (cm) 26 32 58 58 55

Substrate boulder (%) 35 47 82 101 83

Substrate cobble (%) 22 27 49 75 62

Substrate detritus (%) 48 56 104 89 85

Substrate embeddedness (%) 23 25 48 35 32

Substrate gravel (%) 21 25 46 43 46

Substrate silt (%) 30 33 63 62 72




21

from landscape datasets may have significant biases at that

scale. From a regional assessment and monitoring perspective,

we are of the opinion that the reach scale is the best spatial scale

for developing and applying prediction models. The reach scale

represents a section of a stream that is relatively homogenous

in chemical and biological characteristics and is closer to the

scale at which local management decisions are often made.

Accounting for natural spatial and temporal variation at reach

scale is also a component of established sampling protocols

[12-14,39]. Admittedly, assessing and monitoring physical habitat

at reach scales will result in some information being sacrificed,

particular at micro-scale levels which are important determinants

uncertainty that results by predicting in-stream physical habitat

measures does not necessarily lead to additional complexity

from a management perspective, as regardless there is a need

to consider how prediction or measurement biases may affect

management outcomes.

In-stream physical habitat variables can be measured or

predicted at a variety of spatial scales, from individual rocks

to entire stream networks. The accuracy of stream habitat

measurements and predictions is anticipated to be strongly

influenced by the spatial scale of sampling or modeling. At small

spatial scales, direct sampling of habitats will likely yield very

accurate measurements, while prediction models developed

Table 5. Spearman correlation coefficient estimates between selected fish metrics and predicted (Pred.) and observed (Obs.) local habitat measures. Coefficient estimates presented only for those that were significantly different from zero using a Bonferroni-corrected a of 0.05/125 = 0.0004. Where correlations between fish metrics and observed habitat measures were not significant, coefficient estimates for correlation between fish metrics and predicted habitat measures were not presented. Numbers in bold indicate that fish metrics were correlated with only observed, but not predicted habitat variables.

Species richness Darter speciesIntolerant

individuals (%)Lithophilous

individuals (%)Abundance

(#/100m)

Habitat variable Obs. Pred. Obs. Pred. Obs. Pred. Obs. Pred. Obs. Pred.


Bankfull depth (cm) -- -- -- -- -- -- -- -- -0.40 -0.28

Bankfull width (m) 0.49 0.37 0.38 0.20 -- -- 0.46 0.31 -- --

Conductivity (µS/cm) -- -- -- -- -0.48 -0.46 -- -- -- --

Large substrate (%) 0.44 0.32 0.30 0.21 -- -- 0.36 0.25 0.22 0.26

Small substrate (%) -0.32 -- -0.25 -- -- -- -0.34 -- -- --

Substrate sand (%) -- -- -- -- 0.28 0.51 -- -- -0.26 -0.21

Thalweg water depth (cm) -- -- -- -- -- -- -- -- -0.24 -0.35

Wetted width (m) 0.50 0.38 0.28 0.20 -- -- 0.38 0.30 -- --

Water depth (cm) -- -- -- -- -- -- -- -- -0.23 -0.34

Width-to-depth ratio 0.52 0.51 0.40 0.34 -- -- 0.36 0.33 -- --


Bank erosion (%) -- -- -- -- -- -- -- -- -- --

Buffer vegetation (%) -- -- -- -- -- -- -- -- -- --

Canopy shade (%) -- -- -0.22 -0.21 -- -- -0.27 -0.31 -- --

Large woody debris (%) -- -- -- -- 0.27 0.28 -- -- -0.23 -0.39

Fish cover (%) -- -- -- -- -- -- -- -- 0.21 --

Pool (%) -- -- -- -- -- -- -- -- 0.21 --

Riffle (%) 0.29 -- 0.21 -- -- -- 0.31 -- 0.21 0.29

Run (%) -0.23 -0.19 -- -- -- -- -0.21 -0.24 -0.28 -0.34

Sediment depth (cm) -0.35 -- -0.28 -- -- -- -0.29 -- -0.23 --

Substrate boulder (%) 0.24 0.23 -- -- -- -- -- -- -- --

Substrate cobble (%) 0.38 0.20 0.23 -- -- -- 0.31 0.24 0.27 0.22

Substrate detritus (%) -0.26 -- -0.27 -0.23 -- -- -0.36 -0.33 -- --

Substrate embeddedness (%) -0.47 -0.25 -0.35 -0.21 -- -- -0.38 -0.23 -0.22 -0.24

Substrate gravel (%) 0.43 0.30 0.35 0.23 -- -- 0.28 0.21 0.25 0.25

Substrate silt (%) -- -- -- -- -0.22 -0.37 -- -- 0.30 0.30



L. Wang et al.

22

influenced by anthropogenic disturbances (e.g., urban land

cover, agriculture land cover) to zero. Although such estimated

potential conditions are relatively coarse due to the predictive

models not excluding all disturbance influences, this approach

has been considered useful and used for assessing regional

scale environmental impairment in several studies [52-54]. The

models that we constructed for this research were an attempt

to meet the management needs for the states of Michigan and

Wisconsin by filling in in-stream physical habitat data gaps for the

GLRRDACS. As previously mentioned, the spatially referenced

database associated with GLRRDACS includes entries for every

stream reach in a three-state regions, and consists primarily of

data attributed using GIS processes. The database also includes

water temperature [24] and flow discharge [25] predictions and

the results from an integrated biotic-abiotic stream classification

analysis [55,56]. Our in-stream physical habitat predictions are

an effort to complete the hierarchical spatial data coverage from

stream reach to network and to catchment, which is urgently

needed by aquatic resource managers and policy makers.

Before the model predictions are integrated with the GLRRDACS

database, additional evaluations of prediction accuracy,

particularly for streams in southern Michigan, will need to be

conducted. As well, additional attempts at improving models for

those in-stream physical habitat variables for which predictions

were found in this research to be unsatisfactory will need to be

conducted, possibly by incorporating riparian-scale measures

among the set of candidate predictor variables.

Acknowledgement

We thank Paul Seelbach, Jana Stewart, Arthur Cooper, Stephen

Aichele, and Edward Bissell for their effort in the development of

the GLRRDACS, which ultimately made this research possible.

We also thank Paul Kanehl, Edward Baker, and many seasonal

technicians for collecting the on-site physical habitat data that

were used for our analyses and Minako K. Edgar for preparing

the sampling site map. We are grateful to Dr. Yong Cao and an

anonymous reviewer who provided very helpful suggestions

that improved the manuscript. This project was partially

supported by Federal Aid in Sport Fishery Restoration Program,

Project F-80-R, through the Fisheries Division of the Michigan

Department of Natural Resources and Project F-95-P, through

the Wisconsin Department of Natural Resources.

of localized site selection of many aquatic organisms [40,41].

From a regional management perspective, however, we believe

that the benefits of reach-scale assessments for formulating

policy and enacting regulations outweigh the loss of information

at micro-scale levels.

The need for in-stream physical habitat prediction models

largely stems from aquatic managers having to understand

biota-habitat relationships and to assess and monitor natural

and human-induced variations in habitat conditions at broad

scales. One of the important applications of understanding

biota–habitat relationships is the identification of factors that

limit distribution and abundance of important fish species

[42,43]. Presently, regional scale fish distribution and abundance

models rely generally on stream network and catchment factors

[44,45] because in-stream physical habitat data are not available

at regional scales. However, habitat conditions at local scale

(e.g., channel morphology, substrate, and water conditions)

can have strong effects on localized fish distribution patterns in

streams [5,46,47]. Although landscape factors constrain local-

scale conditions, without direct measures of in-stream physical

habitat measures the ability to explain variation in aquatic

communities and assemblages may be seriously compromised.

Also, using landscape factors to approximate local characteristic

in identifying limiting factors, one will not be able to pinpoint

local factors that limit fish distribution for targeting management

activities. Our previous study has shown that local habitat alone

explained 36-46% and local habitat interaction with other factors

explained additional 18-41% of variances for fish presence,

abundance, and integrity metrics [19]. Not taking into account

local habitat in fish distribution prediction or identifying limiting

habitat factors could result in an incomplete understanding of

biota-habitat relationships at a regional scale.

Rivers and streams are considered among the world’s most

imperiled ecosystems and biodiversity loss in these systems

has been high due to a myriad of factors, including intensive

land use practices, damming, pollution, and exotic species

[48-51]. Thus, the importance of assessing and monitoring

natural and human-induced variations and perturbations in in-

stream physical habitats cannot be over-stressed. Using the

modeling approach described herein, physical habitat conditions

can be predicted across very legions. As well, constructed

models can be used to predict potential physical habitat

conditions by setting land cover variables that are strongly

[1] Frissell, C.A., Liss, W.J., Warren, C.E., Hurley, M.D., 1986,

A hierarchical framework for stream habitat classification:

viewing streams in a watershed context, Environ. Manage.

10: 199-214.

[2] Hawkins, C.J., Kerschner, J.L., Bisson, P.A., Bryant, M.D.,

Decker, L.M., Gregory, S.V., et al., 1993, A hierarchical

approach to classifying stream habitat features, Fisheries 18:

3-11.

[3] Poff, N.L., 1997, Landscape filters and species traits:

towards mechanistic understanding and prediction in stream

ecology, J. N. Am. Benthol. Soc. 16: 391-409.

[4] Richards, C., Johnson, L.B., Host, G.E., 1996, Landscape-

scale influences on stream habitat and biota, Can. J. Fish.

Aquat. Sci. 53(Suppl. 1): 295-311.

[5] Wang, L., Lyons, J., Rasmussen, P., Kanehl, P., Seelbach,

P.W., Simon, T., et al., 2003, Watershed, reach, and riparian

References




23

variables predicted from catchment scale characteristics

are useful for predicting fish distribution, Hydrobiologia

572: 59-70.

[19] Brenden, T.O, Wang, L., Seelbach, P.W., Clark, R.D., Lyons,

J., 2007, Comparison between model-predicted and field-

measured stream habitat measures for evaluating fish

assemblages-habitat relationships, T. Am. Fish. Soc. 136:

580-592.

[20] Frappier, B., Eckert, R.T., 2007, A new index of habitat

alteration and a comparison of approaches to predict stream

habitat conditions, Freshwater Biol. 52: 2009–2020.

[21] Goldstein, R.M., Carlisle, D.M., Meador, M.R., Short, T.M.,

2007, Can basin land use effects on physical characteristics

of streams be determined at broad geographic scales?,

Environ. Monit. Assess. 130: 495-510.

[22] Wang, L., Brenden, T.O., Seelbach, P.W., Cooper, A.R., Allan,

J.D., Clark, et al., 2008, Landscape based identification of

human disturbance gradients and references for streams in

Michigan, Environ. Monit. Assess. 141: 1-17.

[23] Wilhelm, J.G.O., Allan, J.D., Wessell, K.J., Merritt, R.W.,

Cummins, K.W., 2005, Habitat assessment of non-wadeable

rivers in Michigan, Environ. Manage. 35: 1-19.

[24] Wehrly, K.E., Brenden, T.O., Wang, L., 2009, A comparison

of statistical approaches for predicting stream temperatures

across heterogeneous landscapes, J. Am. Water Resour.

Assoc. 45: 986-997.

[25] Seelbach, P.W., Hinz, L.C., Wiley, M.J., Cooper, A.R., 2010,

Using multiple linear regression to estimate flow regimes for

all rivers across Illinois, Michigan, and Wisconsin. Michigan

Department of Natural Resources and Environment, Fisheries

Research Report No. 2095, Ann Arbor.

[26] Lyons, J., 1992, The length of stream to sample with a towed

electrofishing unit when fish species richness is estimated, N.

Am. J. Fish. Manage. 12: 198-203.

[27] Venables, W.N., Dichmont, C.M., 2004, GLMs, GAMs, and

GLMMs: an overview of theory for applications in fisheries

research, Fish. Res. 70: 179-193.

[28] Tutz, G., Binder, H., 2006, Generalized additive modelling

with implicit variable selection by likelihood based boosting,

Biometrics 62: 961-971.

[29] Bühlmann, P., Hothorn, T., 2007, Boosting algorithms:

regularization, prediction and model fitting, Stat. Sci. 22:

477-505.

[30] Schmid, M., Hothorn, T., 2008, Boosting additive models

using component-wise P-splines, Comput. Stat. Data An.

53: 298-311.

[31] Smith, E.P., Rose, K.A., 1995, Model goodness-of-fit analysis

using regression and related techniques, Ecol. Model. 77:

49-64.

[32] Piñeiro, G., Perelman, S., Guerschman, J.P., Paruelo, J.M.,

2008, How to evaluate models: observed vs. predicted or

predicted vs. observed?, Ecol. Model. 216: 316-322.

[33] Robinson, A.P., Duursma, R.A., Marshall, J.D., 2005, A

regression-based equivalence test for model validation:

shifting the burden of proof, Tree Physiol. 25: 903-913.

influences on stream fish assemblages in the Northern Lake

and Forest Ecoregion, U.S.A., Can. J. Fish. Aquat. Sci. 60:

491-505.

[6] Wang, L., Seelbach, P.W., Hughes, R.M., 2006, Introduction

to landscape influences on stream habitats and biological

assemblages, In: Hughes, R.M., Wang, L., Seelbach, P.W.

(Eds.), 2006, Landscape influences on stream habitats

and biological assemblages, American Fisheries Society,

Bethesda.

[7] Fisher, W.L., Rahel, F.J., Geographic information systems

applications in stream and river fisheries, In: Fisher, W.L.,

Rahel, F.J. (Eds.), 2004, Geographic information systems in

fisheries, American Fisheries Society, Bethesda.

[8] Brenden, T.O., Clark, R.D., Cooper, A.R., Seelbach, P.W.,

Wang, L., Aichele, S.S., et al., A GIS framework for collecting,

managing, and analyzing multi-scale landscape variables

across large regions for river conservation and management,

In: Hughes, R.M., Wang, L., Seelbach, P.W. (Eds.), 2006,

Landscape influences on stream habitats and biological

assemblages, American Fisheries Society, Bethesda.

[9] Wang, L., Infante, D., Esselman, P., Cooper, A.R., Wu, D.,

Taylor, W., et al., 2011, A hierarchical spatial framework

and database for the national river fish habitat condition

assessment, Fisheries 36: 436-449.

[10] Brenden, T.O., Wang, L., Seelbach, P.W., Clark, R.D., Wiley,

M.J., Sparks-Jackson, B.L., 2008, A spatially constrained

clustering program for river valley segment delineation from

GIS digital river networks, Environ. Modell. Softw. 23: 638-

649.

[11] Warrner, S.S., Fischer, R.U., Holtrop, A.M., Hinz, L.C.,

Movak, J.M., 2010, Evaluating the Illinois stream valley

segment model as an effective management tool, Environ.

Manage. 46: 761-770.

[12] Plafkin, J.L., Barbour, M.T., Porter, K.D., Gross, S.K.,

Hughes, R.M., 1989, Rapid bioassessment protocols for

use in streams and rivers: benthic macroinvertebrates and

fish, U.S. Environmental Protection Agency, Office of Water,

Publication EPA/444/4-89-001, Washington, D.C.

[13] Meador, M.R., Hupp, C.R., Cuffney, T.F, Gurtz, M.E., 1993,

Methods for characterizing stream habitat as part of the

national water-quality assessment program. U.S. Geological

Survey, Open-File Report 93-408, Raleigh.

[14] Simonson, T.D, Lyons, J., Kanehl, P.D., 1994, Guidelines

for evaluating fish habitat in Wisconsin streams. U.S. Forest

Service, General Technical Report NC-164, St. Paul.

[15] Hynes, H.B.N., 1975, The stream and its valley, Verh. Int. Ver.

Theor. Ang. Limnol.19: 1-5.

[16] Jeffers, H.N.R., 1998, Characterization of river habitats and

prediction of habitat features using ordination techniques,

Aquat. Conserv. 8: 529-540.

[17] Davies, N.M., Norris, R.H., Thomas, M.C., 2000, Prediction and

assessment of local stream habitat features using large-scale

catchment characteristics, Freshwater Biol. 45: 343–369.

[18] Mugodo, J., Kennard, M., Liston, P., Nichols, S., Linke,

S., Norris, R.H., et al., 2006, Local stream habitat



L. Wang et al.

24

[45] Steen, P.J., Zorn, T.G., Seelbach, P.W., Schaeffer, J.S.,

2008, Classification tree models for predicting distributions of

Michigan stream fish from landscape variables, T. Am. Fish.

Soc. 137: 976–996.

[46] Gorman, O.T., Karr, J.R., 1978, Habitat structure and stream

fish communities, Ecology 59: 507-515.

[47] Neebling, T.E., Quist, M.C., 2010, Relationships between

fish assemblages and habitat characteristics in Iowa’s non-

wadeable rivers, Fisheries Manag. Ecol. 17: 369-385.

[48] Benke, A.C., 1990, A perspective on America’s vanishing

streams, J. N. Am. Benthol. Soc. 9: 77-88.

[49] Allan, J.D., Flecker, A.S., 1993, Biodiversity conservation in

running waters, Bioscience 43: 32-43.

[50] Rinne, J.N., Hughes, R.M., Calamusso, B., 2005, Historical

changes in large river fish assemblages of the Americas,

American Fisheries Society, Bethesda.

[51] Dudgeon, D., Arthington, A.H., Gessner, M.O., Kawabata,

Z.-I., Knowler, D.J., Lévêque, C., et al., 2006, Freshwater

biodiversity: importance, threats, status and conservation

challenges, Biol. Rev. 81: 163-182.

[52] Wiley, M.J., Seelbach, P.W., Wehrly, K.E., Martin, J.S., 2003,

Regional ecological normalization using linear models: a

meta-method for scaling stream assessment indicators,

In: Simon, T.P. (Ed.), 2003, Biological response signatures:

indicator patterns using aquatic communities, CRC Press,

Boca Raton.

[53] Baker, E.A., Wehrly, K.E., Seelbach, P.W., Wang, L., Wiley,

M.J., Simon, T., 2005, A multimetric assessment of stream

conditions in the Northern Lakes and Forests Ecoregion

using spatially explicit statistical modeling and regional

normalization, T. Am. Fish. Soc. 134: 697-710.

[54] Riseng, M.C., Wiley, M.J., Seelbach, P.W., Stevenson, R.J.,

2010, An ecological assessment of Great Lakes tributaries in

the Michigan Peninsulas, J. Great Lakes Res. 36: 505-519.

[55] Brenden, T.O., Wang, L, Seelbach, P.W., 2008, A river valley

segment classification of Michigan streams based on fish

and physical attributes, T. Am. Fish. Soc. 137: 1621-1636.

[56] Wang, L., Brenden, T., Cao, Y., Seelbach, P., 2012,

Delineation and validation of river network spatial scales

for water resources and fisheries management, Environ.

Manage. 50: 875-887.

[34] Barker, L.E., Luman, E.T., McCauley, M.M, Chu, S.Y., 2002,

Assessing equivalence: an alternative to the use of difference

tests for measuring disparities in vaccination coverage, Am.

J. Epidemiol. 156: 1056-1061.

[35] Robinson, A.P., Froese, R.E., 2004, Model validation using

equivalence tests, Ecol. Model. 176: 349-358.

[36] Lyons, J., 1996, Patterns in the species composition of fish

assemblages among Wisconsin streams, Environ. Biol. Fish.

45: 329-341.

[37] Lyons, J., Trimble, S.W., Paine, L.K., 2000, Grass versus

trees: managing riparian areas to benefit streams of central

North America, J. Am. Water Resour. Assoc. 36: 919-930.

[38] Wang, L., Simonson, T.D., Lyons, J., 1996, Accuracy and

precision of selected stream habitat estimates, N. Am. J.

Fish. Manage. 16: 340-347.

[39] Ramsey, M.H., Thompson, M., Hale, M., 1992, Objective

evaluation of precision requirements for geochemistry

analysis using robust analysis of variance, J. Geochem.

Explor. 44: 23-36.

[40] Geist, D.R., Dauble, D.D., 1998. Redd site selection and

spawning habitat use by fall Chinook salmon: the importance

of geomorphic features in large rivers, Envrion. Manage. 22:

655-669.

[41] Lapointe, N.W.R., Thorson, J.T., Angermeier, P.L., 2010,

Seasonal meso- and microhabitat selection by the northern

snakehead (Channa argus) in the Potomac river system,

Ecol. Freshw. Fish 19: 566-577.

[42] Wall, S.S., Berry, C.R., Jr., Blausey, C.M. Jenks, J.A.,

Kopplin, C.J., 2004, Fish-habitat modeling for gap analysis

to conserve the endangered Topeka shiner (Notropis topeka),

Can. J. Fish. Aquat. Sci. 61: 954-973.

[43] Sindt, A.R., Pierce, C.L., Quist, M.C., 2012, Fish species

of greatest conservation need in wadeable Iowa streams:

current status and effectiveness of aquatic gap program

distribution models, N. Am. J. Fish. Manage. 32: 135-146.

[44] Steen, P.J., Passino-Reader, D., Wiley, M., Modeling brook

trout presence and absence from landscape variables using

four different analytical methods, In: Hughes, R.M., Wang,

L., Seelbach, P.W. (Eds.), 2006, Landscape influences on

stream habitats and biological assemblages, American

Fisheries Society, Bethesda.



Documents

Predictability of In-Stream Physical Habitat for Wisconsin ... · correlation analyses of fish assemblage metric data with both predicted and observed values. Prediction of physical