Mechanisms linking landform, land use, and climate to instream conditions and fish responses are often hypothesized but we have few formal ways of testing mechanisms that operate over such vast areas. The greatest challenges and, therefore, greatest potential opportunities in the field, clearly lie in our ability to identify and test specific mechanisms. Johnson and Gage (1997*) wrote that relationships between the landscape and the stream were not yet well enough defined or quantified as to permit specific predictions of instream responses. This lack of hypotheses has proven a major roadblock to uncovering mechanisms about how streams interact with their surrounding landscapes.
The vast majority of analyses using these data have focused on identifying correlations between landscape conditions and instream conditions, such as fish assemblages, indexes of biological integrity, or water quality parameters. For example, Carlisle et al. (2008) correlated biological indices based on fish with urban and agricultural land uses and Pinto et al. (2006*) found a negative correlation between percent urban area and fish assemblages. Creque et al. (2005*) linked landscape-scale variables such as mean July temperature with spatial variation in the density of sport fish in Michigan, USA. Often, as Creque et al. (2005) did, these analyses compare the relative magnitude of landscape-scale variables versus site-scale variables such as depth, pool distribution, or channel gradient or they compare relationships across similar variables measured at multiple scales (e.g., Feist et al., 2010, 2003; see Table 1 in Durance et al., 2006*, for further examples). Many additional examples of correlative analyses linking landscapes and aquatic systems can be found. Those published before 2004 are summarized in Allan (2004a*) (see Table 2).
Correlative studies run the risk of assuming causality between things that are merely coincident or that share some common yet unmeasured driver. Allan (2004a) commented that correlative analyses are often plagued by (1) covariation between anthropogenic influences and natural landscape gradients; (2) multiple scale-dependent mechanisms; (3) non-linear relationships; and (4) difficulties in untangling impacts of current versus past conditions. GIS software combined with multi-layer remotely sensed data can generate, literally, hundreds of potential predictor variables for any instream response of interest. As these data are dredged, spurious relationships are bound to arise. Results of the MIRR Project in Austria (Schmutz et al., 2007) show that “uncontrolled observations” of the hydromorphological status of streams have led to numerous uncontrolled and cross-correlated variables. In another example, Dauwalter and Jackson (2004) identified counter intuitive relationships that they explained as the result of random observations over limited ranges of land use and water-quality variables.
One sign of our over-reliance on correlative relationships is our inability to make predictions in new areas. Meta-analyses of studies across regions could elucidate generalities and hone hypotheses. Examining the impacts of agriculture on fish assemblages across disparate basins, ecosystems, and even ecotones would begin to test whether there are, in fact, generalizable relationships between landscapes and instream responses. Brown et al. (2009) evaluated the impacts of urbanization on fish assemblages across disparate basins, ecosystems, and ecotones to test for generalizable relationships between landscapes and instream responses, and found few. The development of a standardized GIS framework for collecting, organizing, and sharing riverine and landscape data (Hollenhorst et al., 2007; Brenden et al., 2006) and a standardized framework for effectively incorporating biological data into our digital representations of streams would facilitate such meta-analyses.
Many projects have looked across multiple spatial extents to try and uncover mechanisms linking landscape patterns to instream conditions (e.g., Gido et al., 2006*). The underlying concept of this family of analyses has been to compare the strength of correlations at different scales to identify the scale at which the mechanistic relationship exists, and from that, to develop a stronger theory about which mechanisms are causing the observed patterns. Such analyses have been used to identify relationships between land use or geology and fish assemblages in prairie steams in Kansas, USA (Gido et al., 2006*); between geology or forest composition and instream habitats such as pool distribution in the forested mountains of the Pacific Northwest, USA (Burnett et al., 2007); and impacts of human disturbance on fish assemblages in flat to rolling topography in the Midwest USA (Wang et al., 2006b). Difficulties of this approach for uncovering mechanisms linking landscape patterns to instream conditions include a reliance on correlative patterns, lack of independence between potential predictor variables within and across scales, and arbitrary assignment of particular variables to a particular scale. For example, Gido et al. (2006) assigned stream order to the reach scale whereas some might argue that it could be applied to the catchment or site scale. Despite these limitations, there are many opportunities to improve our mechanistic understanding by incorporating the concept of scale and perhaps expanding it to include both extent and grain and both space and time.
Some progress may be made through application of better and more advanced sampling schemes. Although the technology is not new, advanced statistical sampling designs are increasingly being applied for assessing surface waters in the USA at riverscape (LaVigne et al., 2008a,b), state (Klauda et al., 1998), regional (Ode et al., 2005; Hughes et al., 2004), multi-state (Stoddard et al., 2005; McCormick et al., 2001) and national (USEPA, 2009; Paulsen et al., 2008*) scales. Data from these monitoring programs are then used to link biological conditions and trends with landscape stressors. Further consideration of sampling scales for both instream responses and landscape predictors will make such sampling schemes more efficient.
Elucidating causal mechanisms at landscape scales is likely to remain a challenge. Development of more refined hypotheses should limit potential predictors a priori and reduce spurious effects. But, relationships between landscape conditions and instream responses are inherently noisy and difficult to model. Not all potential causal factors can be incorporated over large extents. For example, downstream factors are often ignored (Hitt and Angermeier, 2006; Pringle, 1997; Osborne and Wiley, 1992). As well, disturbance thresholds may obscure our ability to detect causal mechanisms (Brendan et al., 2008). Large-scale, long-term experiments would be ideal for uncovering mechanisms (Carpenter et al., 1995); however, these are extremely difficult to manage, usually prohibitively costly and, even when treatments can be applied over entire watersheds, identifying appropriate controls remains a challenge (Strayer et al., 2003). Analysis of existing “natural” experiments (space for time substitution, natural disturbances, before and after policy changes) is a promising approach (e.g., Paulsen et al., 2008; Van Sickle and Paulsen, 2008; Brazner et al., 2007).
Human impacts to landscapes occur worldwide (Hughes et al., 2005a; Rinne et al., 2005; Dodge, 1989). Everywhere humans alter landscapes, those landscape are tied to waterbodies that host ecological communities. We are now understanding that human disturbances to river landscapes such as land use (Vitousek et al., 1997, 1986), climate change (Matulla et al., 2007), non-indigenous species (Leprieur et al., 2008b), and morphological and hydrological alterations (Tockner et al., 2009; Nilsson et al., 2005; Tockner and Stanford, 2002; Dynesius and Nilsson, 1994) occur at a global scale. Studies from South America (e.g., Pinto et al., 2006; Tejerina-Garro et al., 2006; Hued and Bistoni, 2003), Asia (e.g., Fausch et al., 2010*; An and Choi, 2003; Ganasan and Hughes, 1998; Houssain et al., 2001), Africa (e.g., Kleynhans, 1999; Toham and Teugels, 1999; Hugueny et al., 1996), New Zealand (e.g., Joy and Death, 2004), and Australia (e.g., Davies et al., 2006) have confirmed that human impacts across the landscape alter fish assemblages.
Many published analyses rely implicitly on the assumption that human activities are randomly distributed across the landscape. But, in fact, human activities are typically constrained by the very same environmental gradients as the biological assemblages of interest. For example, Yates and Bailey (2006) examined relationships between agriculture and landform across 191 basins. They found that agricultural intensity was constrained by geology, in particular drumlin formation and glacial landform type. Covariation between land use and several natural gradients (e.g., geology, soil, slope, elevation, precipitation, temperature) hinders associating land use with biological response (Whittier et al., 2006; Allen et al., 1999). As well, most human disturbances are highly correlated with each other leading to multiple stressors (Fausch et al., 2010) and lack of independence. For example, habitat degradation resulting from hydrological alterations is often accompanied by introductions of non-indigenous fish species that may become invasive, and the effects of these two types of disturbances to river ecosystems are generally impossible to assess separately (Light and Marchetti, 2007; Gurevitch and Padilla, 2004).
Legacy effects of human actions have been well documented in some places (e.g., Harding et al., 1998*). Walter and Merritts (2008*) detail how current stream form and bed load in many eastern US streams is largely a result of streams down-cutting through pond sediments deposited by many thousands of historic mill dams. And, Poissant et al. (2005) found genetic relationships among 12 brook trout populations to better reflect historical hydrologic structure and landscape features than those that were present at the time of the study. Because biological responses can lag behind changes to habitats, poor model fits and spurious results have been linked to a failure to consider historical factors (Van Sickle et al., 2004*; Harding et al., 1998*). Van Sickle et al. (2004) and Harding et al. (1998) both reported that historical agricultural land uses seriously limited current assemblage composition. Humphries and Winemiller (2009) describe the limitations of attempting to design riverine restoration projects without incorporating information about the distribution of once abundant fishes. The few existing studies of legacy effects (e.g., Walter and Merritts, 2008) demonstrate that they have a high potential to form constraints for current conditions, but that they are difficult to detect and quantify.
A major opportunity exists in modeling and quantifying the spatio-temporal structure of human impacts at landscape scales. Rather than considering, for example, “agriculture” simply as a predictive variable in isolation, we could consider its spatio-temporal context. We might consider the relationship of the distribution of agriculture to the distribution of urban areas or to the distribution of particular geologies and climatic conditions. We might also consider the relationship of current land use to past agricultural practices and the current distribution and implementation of agriculture to past patterns of human transportation and settlement. Perhaps riverine landscape ecology can benefit from tools developed in the social sciences to better address the complex nature of human impacts to landscapes.
Johnson and Gage (1997) outlined a series of statistical challenges including (1) skewed data sets, (2) lack of true replication, (3) inherent multi-variate nature of research problems, and (4) colinearity and autocorrelation of landscape metrics. And, Durance et al. (2006*) noted that large-scale research is often plagued by (1) by non-independent sampling, (2) weak inference, (3) poor model testing or (4) model over-extrapolation. She claimed that these challenges prevent fisheries managers from quantifying the importance of large-scale, anthropogenic disturbances. Pyne et al. (2007) and King et al. (2005*) have also identified serious challenges in relating catchment-scale landscape structure to local-scale biotic processes.
Statistical techniques for testing hypotheses rather than “data mining” have generally been applied despite the lack of strong hypotheses about causal links between landscapes and rivers. Beale et al. (2010) used a simulation study approach to assessing the performance of a suite of statistical models and model fitting methods to synthetic datasets designed to capture the range of issues commonly found in spatial data. They found that the generalized least squares family of models and a Bayesian implementation of the conditional auto-regressive model performed best. While the specific results of their study cannot be applied to all spatial datasets, they provide a good example and a reminder that simulation studies can be used to assess the relative performance of various statistical methods for managing the unique challenges of spatial analysis and can therefore help prevent misuse of techniques such as model selection or a priori removal of large-scale spatial trends.
There are also many opportunities for approaching these same data and questions with new or customized statistical tools. Durance et al. (2006), for example, suggest that an increased use of geostatistics, popular in landscape ecology, could improve our understanding of the scale-dependence of landscape-fish relationships and improve our ability to test and develop hypotheses about landscape-scale impacts on fish assemblages. Covariance structure analysis (CSA) enabled Wehrly et al. (2006) to incorporate both direct and indirect landscape influences in models to predict and understand water temperature patterns in stream networks. Zorn and Wiley (2006) also used CSA to untangle the hierarchy of landscape and local influences on fish biomass distribution in streams. Random forest analysis offers a mechanism for teasing out the major predictor variables from large survey data sets (Cutler et al., 2007; Peters et al., 2007). A more customized technique for linking landscapes and rivers is offered by parametric distance weighting (Van Sickle and Johnson, 2008*) which weights landscape and environmental variables by their flowpath distance from the stream. Partitioning variance of the response variable into independent components that reflect spatial variance, environmental variance, noise, and the spatial component of environmental influence is another opportunity (Borcard et al., 1992). Formalized data mining techniques, wavelet analysis, spatial statistics, graph theory (e.g., Schick and Lindley, 2007) and the use of neutral models (e.g., Gardner and Urban, 2007; Gardner et al., 1987) may also help us explore observed patterns and tease out relationships.
To best use results from riverine landscape analyses, we also need to better incorporate and communicate uncertainty into landscape-scale analyses (Burgman et al., 2005). Beginning steps for managing uncertainty include identification of sources of uncertainty, identification (and quantification) of biases that might result from uncertainty, and explicit recognition that management and policy decisions must be robust to known uncertainties. Emerging statistical methods, as described above, may be able to improve the quantification of uncertainty in parameters, in predictions, and across space and time. Mapping methods that can improve our ability to visualize the spatial dimensions of uncertainty will be welcome. The use of alternative scenarios (e.g., Jorgensen et al., 2009; Steel et al., 2008) is also a good tool (Peterson et al., 2003). Alternative scenarios, a method for structured thinking about large-scale patterns in an uncertain and uncontrollable world, are an opportunity both for managing uncertainty and communicating analytical results to inform decision-making.
Landscapes are inherently heterogeneous and dynamic (Pickett and White, 1985). The idea of spatial heterogeneity, integral to landscape ecology, is not new to aquatic ecology (Thorp et al., 2006; Zalewski et al., 1997). Benda et al. (2004) have shown how climatic, hydrologic, and geomorphologic processes mediate the formation and maintenance of dynamic stream habitats. Stream biota, such as salmon (Waples et al., 2008), have evolved in this dynamic landscape. As a result, spatially explicit models have a high potential to address questions about the response of organisms to changing landscape patterns at broad spatial scales (Dunning et al., 1995; Turner et al., 1995). Flitcroft (2007) found hierarchical relationships between dynamic stream habitats and the spatial distributions of coho salmon. However, adequately capturing these dynamic systems with one or two metrics is a formidable challenge (Uuemaa et al., 2009*).
Traditional landscape metrics capture either landscape content (e.g., percent of wetlands within a watershed) or landscape structure (spatial position of land use patches in relation to a waterbody). For instance, does it matter more that urban patches are next to a stream or that there is a large proportion of urban land uses in the watershed relative to forest? Careful application of landscape structure metrics has led to advances in our understanding of many ecological issues (Uuemaa et al., 2009, see Table 1). Examples in riverine ecology include Goetz and Fiske (2008); Van Sickle and Johnson (2008); King et al. (2005); Hunsaker and Hughes (2002) and Allan and Johnson (1997). Kearns et al. (2005) suggested that the lack of metrics that capture biologically relevant components of spatial pattern is, at least in part, responsible for the mixed conclusions of so many analyses attempting to link land use and instream responses. Therefore, future research should focus on metrics that are able to differentiate landscapes by their configuration as well as heterogeneity and patchiness. Patch density, size, and shape are measures of spatial complexity that could be tested for their ability to better develop linkages between landscape configuration and aquatic responses. Careful development of response metrics is also essential.
There are examples in which the careful development of novel metrics or new indices has advanced our understanding of landscape impacts to rivers. The landscape development index (LDI) of Brown et al. (Brown and Vivas, 2005; Brown and Moyle, 2005) uses energy use per unit area to estimate the cumulative impacts of human dominated activities over large extents. By lumping human activities, they have sidestepped the issues in Section 5.2 related to non-independence of human impacts across the landscape. Another example of the use of novel metrics is Scott (2006). He modeled the effects of a trajectory of forest cover change over time on the ratio of endemic specialists to broad-ranged fishes. Using this approach, he quantified the loss of endemism as a result of urbanization.