Northern Watershed Project (Report 2). Defining Relationships between Landscape Characteristics and Fish Communities in the Notikewin River Basin, Alberta.
Garry Scrimgeour, Paul Hvenegaard, Al Wildeman, John Tchir and Sharon Kendall
Spatial and temporal patterns in stream fish communities in the Notikewin River Basin, Alberta, Canada were evaluated using data from 266 stream reaches sampled between 1995 and 2001. Using these data, we addressed the following five focal questions:
1) How temporally variability is fish presence and density?
2) Is the presence of fish predictable based on stream habitat and watershed variables?
3) Do stream reach and watershed characteristics shape fish assemblage structure?
4) To what extent are relations between fish assemblages and watershed variables applicable to adjacent watersheds? and
5) To what extent are fish assemblages shaped by forest cover attributes.
Focal question 1: How temporal variability is fish presence and fish density?
At the watershed scale, stream fish communities comprised 17 species from 8 familial groups. Arctic grayling, longnose sucker, lake chub, brook stickleback, trout perch and northern pike were the most frequently encountered species and occurred at between 5% and 32% of all 266 sampling sites. Comparisons of relative abundance showed that Arctic grayling, lake chub, brook stickleback, longnose sucker, finescale dace, trout perch and stickleback were numerically dominant and, when combined, accounted for 89% of all fish collected.
Overall density of fish in the Notikewin Basin was low (Mean = 0.92 individuals / 100 m2) and only densities of brook stickleback, Arctic grayling, finescale dace and lake chub typically exceeded of 0.1 individuals / 100 m2. Within taxonomic groups, mean densities of cyprinid minnows (overall mean = 0.34 individuals / 100 m2) exceeded that of salmonids (0.22 individuals / 100 m2) and gasterosteids (0.27 individuals / 100 m2). In general, fish communities were numerically dominated by small bodied fishes (i.e., cyprinids, gasterosteids, percopsids, cottids) compared to larger bodied forms (e.g., percids, salmonids, esocids, and catostomids).
Analysis of fish community data from 27 sites (i.e., stream reaches) sampled on two occasions between 1995 and 2001 showed that the presence of fish was highly temporally variable and overall only 52% of all sites that contained fish in one year also contained fish one to five years later. Concordance in the presence of fish between sampling years increased with stream size from about 33% in first order streams (i.e., small streams) to 67% fifth order stream reaches (i.e., larger rivers). Similarly, concordance in the absence of fish between sampling events was also variable; on average only 39% of all sites that did not contain fish on one sampling occasion were also devoid of fish one to five years later.
Regression analyses indicated that the number of species replacements (i.e., number of species gained or lost between sampling years) was positively rela ted to stream size, measured as stream bankfull width, and multiple regression showed that the relative number of species replacements was negatively related with bankfull width and positively related with time between repeated samples suggesting that the presence of fish was more temporally variable in small compared to larger stream reaches.
Fish density was also temporally variable and varied by about 10–fold between 1995 and 2001 (Highest annual mean±1SE [standard error] = 2.2±1.39 100 m2, lowest annual mean = 0.24±0.13 100 m2). Density of Arctic grayling was also highly variable and differed 20-fold during the six year period (1.82±1.1 to XII 0.088±0.05 100 m2). Densities of brook stickleback and other numerically dominant fish species were also variable among years and varied between 5 and 25 fold during the six year study period.
Focal question 2: Is the presence of fish predictable based on stream habitat and watershed variables?
At the stream-reach-scale, the occurrence of fish was strongly affected by stream size. While only about 40% of first and second order stream reaches and small permanent stream reaches contained fish, the majority (65% to 90%) of third, fourth, fifth order stream reaches and large permanent stream reaches contained fish. Stream size also strongly affected the presence of individual species and species groups. Game fish (all species of game combined as one group), cyprinids, catostomids, and individual species of Arctic grayling occurred relatively infrequently (<35% occurrence) in small streams (1st to 3rd and small permanent streams) but were recorded between 20 and 40 times more frequently in larger streams. Brook stickleback, lake chub, longnose sucker, white sucker, northern pike, and trout perch occurred most frequently in larger streams.
At the landscape-scale, logistic regression analyses showed the presence of fish, game fish (all game fish species combined as one group), cyprinid minnows and Arctic grayling were highly predictable based on stream bankfull width, UTM northing and percent gravel. Reach elevation was the strongest predictor of the presence of suckers, longnose sucker, lake chub and brook stickleback. Single variable models using these predictors explained less variance in the presence of northern pike and resulted in lower classification success. Logistic regression models generally explained between 42% to 74% of the variance in the presence of fish, fish type and individual species and correctly classified sites where fish should be present or absent compared to presence based on electroshocking between 71% and 93% of the time.
Logistic regression models completed at the stream-scale (i.e., for 1st to 5th order reaches and small and large permanent streams) also performed well. In the majority of cases, variables that predicted the presence or absence of fish were measures of stream size or related to measures of stream size. At the stream reach level, stream size measured as bankfull width, water temperature, reach slope and to a lesser extent, elevation and gravel were significant predictors of the presence of fish, game fish, taxonomic groups and individual species. After excluding models which were not calculated due to low sample sizes, the majority of logistic regression models were statistic ally significant at P < 0.05 and typically explained between 51% and 87% of variance in the presence or absence of fish and on average correctly classified about 73% of sites as containing fish or as fish-absent.
The effects of stream classification method on the ability of logistic regression models to predict the presence and absence of fish was evaluated by comparing model fit and classification success between models created using Strahler stream orders and that using Alberta watercourse classes. With two exceptions, the stream classification method had only minor effects on the fit of logistic regression models and overall classification success. However, the ability of logistic regression models to correctly classify large permanent reaches as containing game fish and Arctic grayling were low (classification successes = 53% and 59%) and these models are problematic if the objective of modeling is to correctly predict their presence of these groups.
Focal question 3: Do stream reach and watershed characteristics shape fish assemblage structure?
Canonical correspondence ordinations indicated that instream habitat and watershed variables appear to shape fish assemblage structure and explained 24.6% of the variance in fish densities. When constrained by environmental variables, Axes 1 and 2 explained 80.8% of the variation between species abundances and environmental variables. Increased densities of white sucker, northern pike, sculpin and trout perch XIII were positively related with percent cobble, percent gravel, water depth and bankfull width. Density of Arctic grayling was positively related to elevation whereas increased density of longnose sucker was negatively related with elevation. Densities of dace, shiner, brook stickleback and chub were negatively related with cobble, gravel and water depth and bankfull width.
Multiple regression using forward selection showed that total fish density and density of the dominant groups were only weakly or moderately strongly related to environmental variables. Total density and density of individual species and species groups were typically significantly (P < 0.05) related to cobble, gravel, elevation, water depth and width and elevation. These regression models explained between 5% and 32% of variance in fish densities and often (5 of 9 regression models) accounted for 18 to 32% of variance in the abundance of individual species and species groups.
Density of Arctic grayling was positively related with elevation and negatively related with water depth and forest cover. Density of brook stickleback was negatively related with cobble, bankfull depth and percent gravel within the substratum whereas density of longnose sucker and trout perch were negatively related with elevation and positively related with bankfull width, respectively.
Focal question 4: To what extent are relations between fish assemblages and watershed variables applicable to adjacent watersheds?
Hierarchical cluster analyses, using fish density data, identified three relatively discrete fish assemblage types in the Notikewin River Sub-basin. Assemblage 1 consisted primarily of Arctic grayling, brook stickleback and dace whereas Assemblage 2 consisted of Arctic grayling, chub, white sucker, longnose sucker and northern pike. For these assemblages, mean total density and density of the three most numerically abundant groups ranged from 1.42 to 1.77 / 100 m2 and 0.15 to 0.69 / 100 m2, respectively. In contrast, Assemblage 3 was comprised of high densities of brook stickleback, dace, chub and to a lesser extent longnose sucker. Overall, mean total density and density of the three most abundant species was 12 to 15 and 18 fold higher than that in Assemblages 1 and 2.
We used a discriminant function model to differentiate among the three fish assemblage types using instream and watershed-scale variables. Based on fish density data, the forward selection discriminant analysis identified site elevation, percent gravel and cobble, and reach slope as significant (Wilks’ Lambda, P <0.0001) discriminators among the three fish assemblages. The linear discriminant function model had an overall classification success of 76.2% (i.e., 48 of the 63 sites were classified correctly) and correctly classified 73.3% (i.e., 33 of 45), 75.0% (i.e., 9 of 12), and 100% (6 of 6) of sites into assemblages 1, 2 and 3, respectively.
Stream reaches that supported Assemblage 1 were typically located at higher elevations and had higher amounts of gravel within the substratum compared with those belonging to Assemblages 2 and 3. In contrast, water depths and percent cobble at sites supporting Assemblages 1 and 2 were greater than stream reaches that supported Assemblage 3.
We tested the utility of the discriminant function model (based on density and percent composition) derived for fish communities in the Notikewin River Sub-basin by applying the model to fish communities described at 39 sites in the adjacent Hotchkiss and Meikle River Sub-basins. When applied to the Hotchkiss and Meikle River Sub-basins, the discriminant function models correctly classified 74.4% (i.e., 48 of the 63 sites were correctly classified) of all sites and correctly classified 4 of 5 sites into assemblage 1 (i.e., classification success = 80.0%), 17 of 23 sites into assemblage 2 (classification success = 73.9%) and all 8 of the 11 sites into assemblage 3 (i.e., classification success = 72.7%). These data suggest that the discriminant function model developed from the Notikewin River Sub-basin is a relatively good predictor of fish community types in the adjacent Hotchkiss and Meikle River Sub-basins.
These results suggest that relations between fish communities and watershed attributes derived in the Notikewin Sub-basin are also applicable to that in the Hotchkiss and Meikle River Sub-basins.
Focal question 5: To what extent are fish assemblages shaped by forest cover attributes?
Logistic regression analyses showed that forest cover attributes were typically poor predictors of the presence of fish, game fish, taxonomic groups and individuals species compared with regression models derived using both non-forest physiographic attributes and forest cover attributes.
Overall, only 50% (16 of the 32 models) of empirical models using forest attributes were statistically significant (P< 0.10) predictors of the presence of fish. The remaining models were either statistically non significant or were questionable because maximum likelihood estimates could not be calculated. In contrast, 97% (31 of 32 models) of empirical models using both physiographic and forest cover attributes were significant predictors (P<0.10) of fish presence and in the majority of cases (23 of 32 models) models were statistically significant at an alpha of 0.05. Lastly, forest cover attributes were included within only 12.5% (4 of 32) of best-fit models.
Analyses based on Strahler stream orders showed that fish community structure was typically not statistically related (mean model P = 0.12) to forest cover attributes compared with the predominance of statistically significant models derived using non-forest physiographic (P = 0.04). Empirical models derived solely with forest cover attributes also had lower classification success but only marginally lower explanatory power. Analyses using Alberta watercourse stream classes also showed that forest cover attributes had minimal explanatory power in explaining fish community structure and were less powerful than empirical models derived using non-forest physiographic attributes.
The extent to which forest cover attributes explained variation in fish community structure was evaluated by partitioning the variance in fish communities into that which could be explained by: 1) forest cover attributes and 2) physiographic, non-forest cover attributes using variance partitioning techniques.
For this analysis, forest cover attributes included those measured at the watershed scale and those measured within the three riparian zones areas (i.e., areas of 3.2 ha, 28.1 ha, 78.1 ha) delineated by establishing radii of 100 m, 300 m and 500 m around sampling points.
Multiple canonical correspondence analyses showed that the environmental variables explained a relatively small proportion of the overall variance in fish density. When decomposed, the majority (45.6%) of explained variance in fish communities was attributed to non-forest cover attributes of elevation, mean water depth, reach slope, and the percentage of gravel and cobble within the substratum. Forest cover attributes including percent of watershed forested, and percent conifer, deciduous accounted for 29% of the variance that could be explained with the remaining 25.4% attributed to the interaction of forest cover and non-forest cover attributes.
Synopsis of Management Implications and Challenges
Taken together, our data indicate that logistic regression is a potentially powerful and relatively simple method to predict the presence of fish, game fish and some individual species in the Notikewin watershed. The presence of fish was moderately to highly predictable based on a small suite of recurring variables. The majority of these variables are easily measured either by querying existing GIS layers or by completing field surveys and are moderately robust to measurement errors. Logistic regression may prove to be a useful management technique to predict the presence of fish.
In contrast, empirical models predicting total fish density and density of the numerically dominant species and species groups from instream and watershed variables performed relatively poorly. If the lack of model fit reflects inherent variability, rather than poor selection of predictor variables, these watershedscale models may be of limited value to fisheries biologists and resource managers.
In contrast, community-based models identified three relatively distinct fish assemblages in the Notikewin River Basin; these assemblage types were predictable based on reach elevation and percent gravel. Of equal importance, is that the fish assemblage model developed in the Notikewin Sub-basin was broadly applicable to that in the adjacent Notikewin and Meikle River Sub-basins, suggesting that the model may be broadly applicable at least at the local regional scale. The extent that these models can be applied to larger spatial scales, such as ecoregions remains to be determined, but if successful, could provide fisheries managers with an effective larger-scale predictive tool.
Based on our understanding of current stream fish management practices we identify eight areas where additional information and changes in management practices would benefit the conservation of stream fish communities. In an abbreviated form they include:
1) Assessments of the potential effects of industrial, municipal and agricultural activities on stream fish are seldom completed using spatially accurate and up-to-date information. This paucity of information challenges resource managers because they may not be fully aware of species-complexes present in stream and thus may not be able to identify practices that minimize impacts from industrial activities.
2) Fisheries assessments are typically completed at small spatial scales with little consideration of larger landscape patterns in fish communities. This approach precludes identifying broad-scale patterns in fish abundance or community types that should be included as part of evaluations of ecological risks. Management of stream fish communities would be enhanced if evaluations were based on assessments completed at multiple spatial scales.
3) The absence of an effective stream monitoring program compromises our ability to manage stream fish communities in Alberta. Rigorous monitoring programs are required to: i) understand current trends in fish populations, ii) evaluate the ecological effects of anthropogenic and natural disturbances on stream fish communities, iii) evaluate the effectiveness of restoration measures, and iv) to critically assess the effectiveness of current watershed management practices.
4) Current approaches to stream fish management do not take full advantage of quantitative tools or recognized quantitative relationships. The use of these approaches and tools can assist stream fish management by: i) providing techniques to understand large-scale patterns in fish communities, ii) gaining insights to potential cause-effect relationships that drive fish communities, iii) evaluating environmental impacts, iv) quantifying temporal variance in fish communities, and vi) monitoring the effectiveness of current management actions.
5) A better understanding of the causal mechanisms responsible for the degradation of stream fish communities would enhance stream fish management by: i) identifying the major causes of negative impacts on fish populations, ii) allow scenario modeling exercises to forecast long-term and large scale consequence of alternative management actions and iii) providing an ecologically sound approach to restoration actions.
6) An improved understanding of the cumulative effects of watershed disturbances on stream fish communities is required including the effects of land use conversions (i.e., conversion and loss of forested watersheds to agriculture).