Species distribution models (SDMs) provide quantitative descriptions of species’ distributions based on associations between observational data and environmental predictors (Guisan and Zimmermann 2000, Elith and Leathwick 2009a). SDMs built with count data, called species abundance models (SAMs), quantify indices of abundance or density rather than occurrence (Elith and Leathwick 2009a). Given suitable predictors, SDMs and SAMs can identify key habitat factors for species of interest (Milsom et al. 2000, Morrison et al. 2006), providing insight on autecology and informing conservation decisions. Predictive maps generated from these models provide basic information on species distribution and abundance (Guisan and Zimmermann 2000, Austin 2002). These maps can be used to improve species range delineation (Loiselle et al. 2003), to test biogeographical hypotheses (Leathwick 1998, Anderson et al. 2002), to inform conservation strategies (Ferrier 2002, Ortega-Huerta and Peterson 2004, Moilanen et al. 2005, Elith and Leathwick 2009b), and to predict population responses to climate and land use change (Araújo et al. 2004, Thuiller 2004, Jetz et al. 2007). The focus of this paper is on SAMs for breeding waterfowl in boreal and southern Canada.
The urgency of conservation challenges in the Canadian boreal forest is increasingly recognized (Schindler and Lee 2010, Badiou et al. 2013, Berteaux 2013). To meet these challenges, we must quantify species abundances, distributions, and habitat relationships to target locations for conservation and identify the conflicts between wildlife habitat and intensive human activities (Slattery et al. 2011). The predictive maps generated through SDMs and SAMs may be particularly useful for these purposes in remote or unsurveyed areas, such as much of the boreal regions of northern Canada. A number of such studies are now under way, directed at boreal forest songbirds (Cumming et al. 2014), and waterfowl in eastern Canada (Lemelin et al. 2010, Börger and Nudds 2014) and western Canada (L. Armstrong, J. Devries, D. Howerter, B. Kazmerik, A. Richard, S. Slattery, and B. Tedford, unpublished report; Ducks Unlimited Canada, unpublished report). However, there has been no attempt to model the abundance of individual waterfowl species at national levels.
The lack of national waterfowl distribution models is surprising given the existence of an extensive and long-term aerial survey, the Waterfowl Breeding Population and Habitat Survey (WBPHS; Smith 1995). Each May since 1955, the U.S. Fish and Wildlife Service (USFWS) and Canadian Wildlife Service (CWS) survey much of the breeding distributions of many waterfowl species (Fig. 1). The survey has created a large data set on individual species’ abundances; this information was collected using relatively consistent sampling methods and survey design. The survey was originally designed to provide annual estimates of total duck population to inform hunting regulation (Nichols et al. 1995), although a substantial body of secondary research exists.
Some secondary applications exploit the survey’s time series to model population dynamics, exploring relationships between population size and wetland availability, adult survival, annual productivity, and climate change (Pospahala et al. 1974, Anderson 1975, Johnson and Shaffer 1987, Kaminski and Gluesing 1987, Raveling and Heitmeyer 1989, Podruzny et al. 2002, Drever et al. 2012). More recently, the data set has been used to test new methodological approaches (Jamieson and Brooks 2004, Gimenez et al. 2009, Ross et al. 2012, Lawrence et al. 2013). However, there have been only limited attempts to use these data in predicting species distribution and abundance over large areas. Species abundance patterns have been mapped for several species at low spatial resolution (Johnson and Grier 1988). Annual abundances were correlated with the number of available wetlands (Johnson and Grier 1988), but only for regions with annual pond counts. Spatial interpolation of survey data has been used to predict abundances of Scaup (Aythya affinis, Aythya marila) pairs across their range (Hobson et al. 2009). However, no SAMs and their derived predictive abundance maps have been published for any waterfowl species at national extent.
We used the WBPHS database to develop high spatial resolution, predictive SAMs for 17 waterfowl species or species groups. Our modeling methodology was Boosted Regression Trees (BRTs) with a library of 78 environmental predictor variables selected a priori for their biological or physical effects on waterfowl species. We used the models to develop predictive relative abundance maps for most of Canada, excluding tundra areas where certain hydrological covariates were not available. In this paper, we present an evaluation of the models in terms of their predictive power and bias. We summarize the importance of different classes of environmental covariates: climatic, bioclimatic, hydrological, and landscape, and informally examine whether their relative importances differ by feeding or nesting habit.
We obtained waterfowl count data from the WBPHS (Smith 1995). Observers count adults of common waterfowl species seen within 200 m of fixed-wing aircraft flight transects. Within transects, counts are spatially assigned to segments of about 28.8 km (by 0.4-km width for a segment area of 11.2 km²). For details on survey methodology see Smith 1995, U.S. Fish and Wildlife Service 2012, and Zimmerman et al. 2012.
We modeled segment-level counts of total indicated pairs from the Canadian portion of the WBPHS survey area (2273 segments; Fig. 1). Total indicated pairs are an estimate of the number of pairs present based on the raw counts of actual pairs and lone males detected, taking into account species-specific life history (Dzubin 1969, Smith 1995, U.S. Fish and Wildlife Service 2012). We used 15 years of data from 1995 to 2010; data from 2007 were excluded because of a deviation in that year from the usual survey design (Silverman 2011). We used this temporal subset to capture a relatively static, recent, and representative sample of waterfowl counts.
We created 17 species abundance models, 13 for individual species and 4 for species that are grouped within the survey protocol (Table 1): Goldeneye (Bucephala clangula, Bucephala islandica), Merganser (Mergus merganser, Mergus serrator, Lophodytes cucullatus), Scoter (Melanitta americana, Melanitta fusca, Melanitta perspicillata), and Scaup. These four groups are referred to as species hereafter. To explore how characteristics of fitted models might vary with life history, species were assigned feeding and nesting guilds (Table 1) following Bellrose 1980 and Johnsgard 2010. We classified the Merganser group as cavity nesting, despite the Red-breasted Merganser being a ground nester (Titman 1999), to reflect the more specialized nesting habit of the other two species in the group.
We selected environmental predictor variables based on previous studies of waterfowl habitat selection (Horn et al. 2005, Paszkowski and Tonn 2006, Suhonen et al. 2011) and songbird distribution (Cumming et al. 2014, Stralberg et al. 2014). Selection criteria included a priori hypotheses on likely ecological relationships (Mac Nally 2000, Barry and Elith 2006, Wenger and Olden 2012), constrained by availability and coverage. Where possible, we prioritized variables with more direct and likely causal relationships with waterfowl abundance, although proxy or indirect variables (Austin 2002) were used in many cases. We used a total of 78 climatic, bioclimatic, landscape, and hydrological predictor variables (Appendix 1).
Annual climatic variables and most bioclimatic indices were extracted from interpolated weather station data provided by Natural Resources Canada (McKenney et al. 2011). We calculated 30-year means of maximum and minimum monthly temperature, summarized to season, to capture potential temperature limits imposed on waterfowl abundance through vegetation, invertebrate phenology, and other seasonal influences (Appendix 1). We calculated 30-year means of total monthly precipitation, summarized to season, to capture the influence of precipitation through wetland conditions. We calculated 30-year standard deviations of monthly maximum temperature, minimum temperature, and precipitation as proxies for the probabilities of extreme events (Cumming et al. 2014, Lynch et al. 2014) Thirty-year means of bioclimatic variables capture effects of seasonality in temperature and precipitation, the duration of the breeding season, and other factors (Appendix 1). Gross primary productivity was included as a proxy for food availability, both directly as plant food and indirectly through the provision of food for invertebrate prey.
We included variables for land cover to capture the influence of upland vegetation on waterfowl abundance and distribution, using the 250-m resolution Land Cover Map of Canada (Canada Centre for Remote Sensing 2008). The original 39 land cover classes were reclassified to 16 (Appendix 2). Proportional class areas were calculated per segment. Shannon’s diversity index was calculated from these land cover proportions as an index of habitat heterogeneity. Topographic ruggedness (Integrated Remote Sensing Studio 2010) was included to represent the influence of landform or terrain.
To measure the type and abundance of surface waters, we used Ducks Unlimited Canada's Hybrid Wetland Layer (N. Jones, unpublished report) to calculate the proportional area of wetlands, including swamps, marshes, bogs, and fens, and open water, including shallow open water and deeper water systems, in each segment. Shoreline length and complexity, stream length, and the density and mean size of water bodies were measured from the National Hydrology Network product (Appendix 1; Natural Resources Canada 2007). These variables were included to represent influences of size, depth, and shape of wetlands as surrogates for availability of different foods for waterfowl.
All environmental variables were measured over a 1-km buffer around each segment, to account for imprecision in survey flight lines among years and to capture the influence of local environmental conditions. Spatial analysis relied on the integration of raster and vector data in PostGIS 2.0 (Obe and Hsu 2011).
We built SAMs using BRTs, a machine learning technique that combines the advantages of decision-tree analyses with boosting and bagging algorithms to improve predictive performance (Friedman et al. 2000, Friedman 2002). BRTs have several advantages: they detect important relationships from large sets of predictor variables; they accommodate the complex, nonlinear relationships that often exist between species and habitat (Gaston 2003); they accommodate multiway interactions among predictor variables; they are insensitive to outliers and transformations of the predictor variables (Elith et al. 2008); and they show good predictive performance (Elith et al. 2006, 2008, Oppel et al. 2012).
We assumed that the annual counts of waterfowl pairs per segment were independent Poisson random variables conditional on the environmental predictors. We used BRT models with Poisson error and logarithmic link. The response variable was segment-level total indicated pairs. Because segments were surveyed for variable numbers of years, we summed total indicated pairs over all years sampled and included log(years sampled) as an offset. Under the Poisson assumptions, this correctly accounts for variation in sampling effort among segments (Sólymos et al. 2013, Cumming et al. 2014). The model-fitted values and predictions were interpreted as the expected annual abundance in a segment. To fit models, perform cross-validation analyses, and derive predictions, we used the dismo (Hijmans et al. 2012) and gbm (Ridgeway 2013) packages, running under R 2.15.2 (R Core Team 2012).
Within BRTs, user-selected parameters and constraints customize the complexity of the trees to prevent the overfitting that is common in simpler tree methods such as classification and regression trees (Guisan and Zimmermann 2000, Elith et al. 2008). The tree complexity parameter controls the number of permitted interactions among predictor variables (Elith et al. 2008). Based on preliminary analysis of model performance, we set the tree complexity parameter to four as representing a reasonable trade-off between flexibility and overfitting. Learning rates necessary to obtain approximately 1000 trees using a 10-fold cross-validation procedure detailed by Elith et al. 2008 varied among species, ranging from 0.015 for the Ruddy Duck (Oxyura jamaicensis) to 0.65 for the Mallard (Anas platyrhynchos).
We assessed the predictive performance of our models using repeated random subsampling cross-validation (Lu et al. 2011, section 10.6.1). We built 20 replicate BRT models for each species, using independent random samples of 70% of the segments (n = 1591). Model predictive performance was assessed by comparing predictions with observations for the remaining segments (n = 682). For each replicate model, we calculated four evaluation statistics. The first two statistics were the estimated intercepts and slopes of linear regression models of predictions against observations. The intercept measures the magnitude and direction of bias, with values close to 0 indicating low or no bias. The slope yields information about the consistency in the bias as a function of the mean, with a value of 1 indicating a consistent bias if the intercept is a nonzero value. The third statistic is Spearman’s rank correlation, which measures overall consistency of predictions and observations. Finally, the squared deviance explained, D², is the Poisson equivalent to the Normal R² and quantifies the goodness of fit. Spatial autocorrelation was assessed by comparing Moran’s I calculated using residuals from BRT models with that calculated using residuals of a null model including only the mean. We calculated residual autocorrelation at two scales: the segment and the transect (see Appendix 3 for details). At each scale, we report statistics for both null model and BRT model residuals and their ratio. The ratio measures the degree to which the BRTs reduce spatial autocorrelation by conditioning on the covariates.
We mapped predicted pair abundance by applying BRT models to environmental variables, resampled to a consistent projection, i.e., geographic North American Datum 1983, on a 300-arcsecond grid. Cell areas over the prediction zone ranged from 27.9 to 68.1 km². Predicted values were the expected pair abundance within a representative 11.2-km² section of each grid cell, conditional on the values of environmental predictors within the cell. For mapping and statistical comparison, these values were scaled to densities of pairs per square kilometer. Maps of species’ abundances show means over 20 model replicates. We mapped prediction uncertainty as the coefficient of variation in density over the 20 replicates. This measure is independent of the mean and provides information on variability in the predictions resulting from particular data subsets and subsequent contributions of different environmental predictor variables.
We evaluated patterns of relative importance of different classes of predictor variables across feeding and nesting guilds. The relative importance of a predictor represents its influence within the model based on how often it is selected for splitting a tree, combined with how much it improves the model as a result of the split, averaged over all trees (Friedman 2001, Friedman and Meulman 2003, Elith et al. 2008). Relative importance was calculated for all predictor variables in a BRT model using formulas within the gbm package (Elith et al. 2008, Ridgeway 2013). For each predictor and each species, we calculated the mean ± 1 SD relative importance over the 20 BRT replicates. One caution in examining these models in terms of variable importance is that complex, nonlinear responses are permitted within BRTs. Therefore, simply examining variable importance only captures one element of the results.
To understand patterns in variable importance across species, we identified the class (Appendix 1) of the most influential variable across all species, and for each feeding and nesting guild. We also calculated the cumulative relative importance for each variable class across all species and for each feeding and nesting guild.
Overall, models performed well. Across all species, mean ± SD was 0.78 ± 0.09 for Spearman’s rank coefficient and 0.75 ±0.13 for D², indicating good correspondence between predictions and observations. Model calibration metrics indicated little bias in the predictions. A mean ± SD slope of 0.94 ± 0.03 and an intercept of 0.17 ± 0.08 across all species suggest slightly positively biased predictions, with slightly greater overprediction at low abundances. There was some variability in evaluation metrics among species and guilds (Fig. 2), indicating that models were better for some species than others. The residuals of 13 out of 17 species exhibited significant (p < 0.05) spatial autocorrelation at either the segment or transect scale (Appendix 3). Null ratios were much larger than 1 in all cases, indicating that the BRT models removed most of the spatial autocorrelation. The magnitude of Moran’s I was generally higher at the segment scale that at the transect scale.
Most species had the lowest prediction uncertainty in the prairie-parkland region (Appendix 4), where the density of WBPHS transects is highest. Notable exceptions were the Canvasback (Aythya valisineria) and Redhead (Aythya americana), which showed localized areas of high uncertainty in western Ontario and eastern Manitoba (Appendix 4). The Ruddy Duck showed high uncertainty at the northern limits of the survey area and along the St. Lawrence water way (Appendix 4).
The mean coefficient of variation across all 17 species indicated that the highest uncertainty occurred in the Yukon, and along parts of the Pacific, northern, and Atlantic coasts (Fig. 3). Localized areas of high uncertainty also were scattered throughout the prediction area. These were mostly outside the area covered by the WBPHS (Fig. 1), especially in the tundra or western cordillera. Within the survey area, areas of localized high uncertainty were associated with large lakes or, as in northern Québec, with hydroelectric reservoirs.
The highest total waterfowl density was predicted to occur in the prairie-parkland region (Fig. 4), and in scattered areas throughout the western boreal, including the MacKenzie River Delta, the unsurveyed southwestern Yukon, the Peace-Athabasca Parklands, the Athabasca River Delta, and the Saskatchewan River Delta. Within the survey area, total predicted densities were lower in the eastern than in the western regions. The lowest total predicted densities occurred in the unsurveyed western cordillera, between the Pacific Ocean and the Rocky Mountains, and the Atlantic coastal areas.
Within the WBPHS survey area, predictions of species’ density from BRT analyses closely resembled spatial patterns in the raw observations (Appendix 4). Maximum raw densities exceeded maximum predicted densities for nine species (Green-winged Teal [Anas crecca], Blue-winged Teal [Anas discors], Goldeneye spp., Merganser spp., American Black Duck [Anas rubripes], American Wigeon [Anas americana], Canvasback, Redhead, and Scaup spp.), whereas the opposite was true for two species (Gadwall [Anas strepera] and Scoter spp.). Maximum densities were similar between observations and predictions for the remaining six species. The preponderance of underpredictions at high densities is consistent with mean model calibration slopes being less than 1.0 for all but one species (Fig. 2).
Ground-nesting dabbling species tended to have highest predicted densities in the prairie-parkland (Appendix 4). Exceptions include the American Black Duck, the Mallard, and the Green-winged Teal. The American Black Duck was predicted to occur across the entire eastern portion of our study area, including the boreal, Laurentians, and lowlands. It was predicted to occur at comparable densities on the west coast and in the southern interior of British Columbia. In addition to the prairie-parkland region, the Mallard was predicted to occur in most of the western boreal and the Laurentians and lowlands of Ontario. The Green-winged Teal was predicted to occur at low densities throughout nearly the entire study area. Predicted densities of American Wigeon, Northern Pintail (Anas acuta), and Northern Shoveler (Anas clypeata) were highest in the prairie-parkland region, but substantial abundances were also predicted in parts of the western boreal, including the northern limits of the region.
Cavity-nesting diving ducks were predicted to occur at low densities over most of our study area, excluding the prairies (Appendix 4). The Bufflehead (Bucephala albeola) had the most restricted distribution, covering most of the western boreal and the aspen parkland, but was rare or absent in the eastern half of the study area. The Merganser group had the broadest predicted spatial distribution, occurring almost everywhere except the prairie region, with highest overall predicted abundances occurring in forested areas of Saskatchewan and the eastern boreal. The Goldeneye group showed a pattern similar to that of Mergansers, except for occurring at apparently lower abundances within the lowlands along the St. Lawrence River. The single ground-nesting diving species, the Scoter group, was predicted in highest densities along the northern limit of our study area, covering much of the northern boreal, including the unsurveyed Yukon (Appendix 4).
Overwater-nesting diving ducks tended to have high predicted densities in the prairie-parkland region, particularly in the aspen parkland (Appendix 4). Exceptions included the Ring-necked Duck (Aythya collaris), which according to the model, is very broadly distributed at low densities, and the Scaup group, which was predicted to occur over much of the western boreal and the northern portion of the eastern boreal in addition to the prairie-parkland region.
The most important single variable within models was most frequently hydrological. Hydrological variables were selected more frequently than expected given their representation within the predictor set (Fig. 5). The next most important class was landscape variables, followed by climatic variables. Bioclimatic variables were never the most influential variable. Overwater-nesting ducks never had climatic variables as the top variable. None of the cavity nesters had landscape variables as the top variable.
Within variable classes, some specific variables were frequently important across species. The proportion of open water was most commonly the top hydrological variable, being most important for five out of eight species (Appendix 4), followed by climate moisture index, which was most important for three out of eight species. Proportion of cropland was the top variable for four out of six species. The other top variables were each represented by only a single species (Appendix 4).
Across all species, the class total of variable relative influence was highest for climatic variables, followed by hydrological, landscape, and lastly bioclimatic variables (Fig. 6). This was also the case for all nesting and feeding guilds. For all species and among all guilds, hydrological variables had a disproportionately large share of influence given the number of variables in each class (Fig. 6). Species-specific tables of variable relative importances are presented in Appendix 4.
BRT models of waterfowl relative abundance performed well, according to a suite of standard evaluation metrics. Uncertainty maps show that predictions were most precise within the prairie-parkland region, where WBPHS transect density is highest. On visual inspection, large-scale patterns in predicted species’ abundances were congruent with spatial patterns in the raw data (Appendix 4). A map of total predicted waterfowl abundance identified many areas traditionally recognized for their importance to waterfowl, including the prairie-parkland region, the MacKenzie River Delta, the Peace-Athabasca Parklands, and the Saskatchewan River Delta. The differing magnitudes of residual spatial autocorrelation across scales are consistent with the main sources of unexplained spatial structure being processes acting at scales on the order of 10 km, i.e., among adjacent segments. Possible examples would be population processes such as short-distance natal dispersal or unmeasured environmental factors varying at that scale. In either case, we do not expect parameter estimates or variable selection to be markedly biased.
Species’ abundances followed two main patterns. In the first pattern, high abundances were predicted within a relatively small area, most often the prairie-parkland region. This pattern was most common for ground-nesting dabbling ducks. In the second pattern, comparatively low abundances were predicted evenly across much of Canada, with no well-defined core area. This pattern was most common among the cavity-nesting ducks but also applied to the Green-winged Teal and Ring-necked Duck. Model statistical performance varied consistently among these two patterns, with models for ground-nesting and dabbling species having higher predictive performance than those for other nesting and feeding guilds.
For many species, the most influential predictor variable appeared to correspond to spatial pattern rather than to a clear biological process. Species with most of their population contained within the prairie-parkland region tended to be associated with the amount of cropland, climate moisture index, and water body density. These three variables likely delineate the prairie-parkland region through its hydrological regime or soil composition (Environment Canada 2010), either directly or as reflected in the abundance of farmland in the region. Species with higher densities in the aspen parkland, such as the overwater-nesting diving ducks, showed more relationships with the amounts of open water and shoreline and with the abundance of cropland and cropland-woodland landcover classes. Wetlands in the aspen parkland tend to have less emergent vegetation than those in the prairies (Johnson et al. 2005), so the open-water variable may represent these parkland wetlands. Similarly, the aspen parkland has more wooded areas than the nearby prairie region, while also maintaining large amounts of farmland (Johnson et al. 2005, Wiken et al. 2011), so amounts of cropland and cropland-woodland would help to distinguish this region from the prairies to the south and closed forests to the north. For species not concentrated within the prairie-parkland region, climate means and bioclimatic indices were more important than landcover variables. Species with larger, more homogenous distributions, including the cavity nesters and Ring-necked Ducks, had relatively high associations with variability in seasonal temperatures. Species with geographic distributions not corresponding to the two main patterns, such as Scoters in the North and the American Black Duck in the East, showed unique combinations of variable importance: an overwhelming importance of climate for Scoters and an extremely strong influence of topography for the Black Duck. Our results for waterfowl contrast somewhat with those showing that climate and climate variability are most important in describing or determining songbird distributions (Cumming et al. 2014).
When the importance of environmental correlates is determined primarily by the continental-scale geographic distribution of species abundances, the variables selected as most important may not be causally related to the pattern of interest. For instance, the importance of cropland variables does not suggest that prairie-parkland species prefer to nest within cropland. We are more likely to see noncausal relationships when using indirect environmental predictors. One could improve the biological correspondence of important variables by repeating the analyses at regional scales (e.g., N. Barker, unpublished report). For example, restricting the analysis of prairie-parkland species to segments within the prairie-parkland would likely reduce the apparent importance of amount of cropland and instead yield insight into drivers of finer-scale patterns in abundance. In a regional analysis, it might be possible to use more direct environmental predictors rather than the proxies used here, because of increased data availability. These analyses might therefore yield more insight into biological relationships and habitat selection.
Waterfowl habitat selection has been summarized as “ducks like water” (Pimm 1994). Indeed, the most influential variable in 8 of our 17 species’ models was a hydrological one, and at least 1 hydrological variable was among top 5 most important variables for 16 species. The single exception was the Scoters, for which the top water-related variable was ranked 23rd. Our findings confirm that ducks do, in fact, like water. However, we have also shown that this glib summary is not the whole story. The type or distribution of water, or the way in which water is quantified, differs among species’ models, depending on either life history traits or spatial distributions. For ground-nesting dabbling ducks, the top hydrological variables were usually water body density, climate moisture index, amount of shoreline, and amount of open water. For overwater-nesting ducks, both the amount of open water and the amount of shoreline were important. These variables likely help identify the prairie-parkland region and separate the prairies from the aspen parkland. For the broadly distributed cavity-nesting species and the Ring-necked Duck, the amount of open water was highly important, potentially because it best identified available wetlands outside of the prairie-parkland region. The “amount of wetland” variable was in the top 20 variables for 9 species, but in the top 10 variables for only 2 species. Given the known dependence of many duck species on wetlands and marshes, particularly those with a hemimarsh condition of 50% plants and 50% open water (Murkin et al. 1982), it is surprising that this variable was relatively unimportant. We posit that this result stems from the difficulty of classifying wetlands from satellite imagery. A finer wetland classification system that separates wetlands into classes such as marsh, swamp, bogs, fens, river, and lake might shed more light on actual habitat selection and preference, although such a product has yet to be developed at the national extent. An even finer classification that incorporates detailed wetland attributes such as size, shoreline characteristics, bathymetry, phosphorus, food supply, and presence of fish would likely prove more effective, but no such data are available at national extents.
The primary use of WBPHS data is to estimate annual continental waterfowl populations. Detection of waterfowl from aerial surveys is not perfect. The resulting counts are nearly always underestimates, although overestimates do occur, e.g., for Scaup (Austin et al. 2002). To account for incomplete detection in estimating continental population sizes, raw counts are adjusted by visibility correction factors (VCFs). VCFs are estimated from ground surveys on a subsample of survey segments (Smith 1995). We did not incorporate these VCFs in our models for three reasons: (1) they were generated at the coarser crew area scale (Smith 1995), and this imprecision may lead to inaccurate visibility correction at the segment scale; (2) they were generated by different methods and frequencies for different regions (e.g., boreal vs. prairie-parkland; Smith 1995), casting doubt on the validity of comparing corrected abundances among regions; and (3) none are available for the eastern half of the survey area (Zimmerman et al. 2012), precluding the creation of Canada-wide models. We modeled the uncorrected observed counts. Under the assumption that detectability is constant across habitats, all observed counts are underestimated in the same proportion. Thus, our model predictions could be interpreted as relative abundances.
In fact, detectability does vary among habitats and species, as shown by the variation in VCFs among species and regions: 0.42-12.6 in the boreal, 0.60-29.10 in the prairie-parkland, and 0.92-10.30 in the tundra. Therefore, we expect results of our models to be biased in some degree, both in terms of predicted abundances and in the habitat associations identified. We have conducted some preliminary analyses comparing models with and without VCFs, and found that they yielded visually similar spatial patterns in predicted abundances and fairly similar identification of at least the top three influential habitat associations. Despite these apparent similarities, we encourage caution when comparing relative abundances between regions or among species, because VCFs are variable and our predicted maps are a combination of true abundances and unknown detection probabilities. We expect that our maps are most reliably used for single species within regions with relatively homogenous detectability. If desired, one could adjust our predictions post hoc using known VCFs. Our preliminary analyses suggested that the spatial variation in predictions echoed the variation in VCFs across crew areas. Therefore, one could simply multiply our predicted densities by VCFs where available, yielding a very rough approximation of VCF-corrected models. Our models and resulting predictions could certainly be improved were regionally specific VCFs, obtained at fine spatial scales using consistent methods, to become available.
For some species, surveys are conducted before pairs are settled on the breeding grounds (Austin et al. 2002). As a consequence, late-nesting species such as Scaup and Scoters may be observed outside their breeding ranges (Schummer et al. 2013). However, if migrating in groups, staging and migrating birds would be excluded from the total indicated pair count. We expect the numbers of out-of-range birds to be low relative to the numbers observed on the breeding grounds, even in late-nesting species. Therefore, we expect the influence of these observations on our models to be low. This is one advantage of having used segment-level counts, rather than presence, as the response variable.
There are limits to interpreting a BRT simply in terms of the most important variables. BRTs permit complex, nonlinear responses. Here, we specified a tree complexity of four, allowing several levels of interactions. With enough predictor variables, BRT models can likely match most observed spatial patterns. Therefore, even when the top variables are the same or similar across species, individual response curves and interactions can lead to substantial variation among fitted models and subsequent predictions of species’ abundance. BRTs are therefore better suited for prediction than for biological understanding (Elith et al. 2008).
Extrapolation, or out-of-sample prediction, occurs when we predict for times or locations not represented in the data (Wenger and Olden 2012). Coefficient-of-variation maps (Fig. 3 and Appendix 4) indicated more uncertainty outside the survey area than within, as expected. Within the survey area, we noticed localized areas of particularly high uncertainty near the coasts of very large water bodies, e.g., big lakes or oceans. This uncertainty may further be associated with extrapolation error. Models were built using data sampled near smaller water bodies, but were extrapolated to raster cells adjacent to extremely large water bodies.
We observed a few instances of clear overprediction outside the survey area, notably for the Scoters and the American Black Duck. The northern range of the Scoter group was best explained by climatic variables. These variables typically exhibit gradients, but only part of the range of these covariates was contained within the sampling space. As one consequence, predicted Scoter abundance was high in the far North of our study area, probably further north than they actually occur (Savard et al. 1998). Similarly, climatic and topographic variables explained spatial variation in American Black Duck numbers across the area sampled by the WBPHS. Conditions in Eastern Canada, where the Black Duck occurs, are also found in western coastal and southern interior British Columbia, where the species does not occur except as a few local introductions. Because these western areas are unsurveyed, our data set contained no information to constrain model creation. Therefore, the models predicted high abundances of the species because the habitat is theoretically suitable. One approach when faced with extrapolation errors would be clipping maps to regions of high confidence. We chose not to do this because these errors represent information, either pointing to potentially suitable habitat or to a need for methodological refinement.
Four factors may limit the extrapolation reliability of our models. First, while machine-learning techniques typically have high in-sample prediction reliability, they tend to have somewhat limited extrapolative ability because the models include complex nonlinear relationships (Wenger and Olden 2012). We attempted to restrict overfitting by using relatively simple trees, but this is not always sufficient. One solution is to use simpler models (Wenger and Olden 2012), such as generalized linear models, but these do not readily accommodate the number of predictor variables used here. An a priori reduced covariate set would be necessary, a step we specifically intended to avoid. As suggested by Cumming et al. (2014), it should be possible to use the most influential variables identified in this study as predictor variables in future generalized linear model analyses. Regional analyses may further facilitate the selection of a smaller set of predictors.
Second, model extrapolation is more accurate when predictions are restricted to the range of environmental conditions sampled and when the full relationship of species’ abundance to predictors is captured (Thuiller et al. 2004, Randin et al. 2006, Elith and Leathwick 2009a). The figures within Appendix 1 suggest that unsurveyed portions within our study area present novel environmental conditions. This would account for the extrapolations of Scoters beyond the northern limit of published range maps. There is no easy solution to this problem. Ideally, new data or data from other sources would be collected or otherwise obtained for these regions (Audubon and Cornell Lab of Ornithology, eBird, http://ebird.org/content/ebird/).
Third, the set of environmental variables used in this analysis was limited by availability and extent, and included many indirect or proxy variables. This can degrade model prediction reliability (Mac Nally 2000, Randin et al. 2006). Prediction errors may occur if proxies are not causal or consistent across the study area (McIntire and Fajardo 2009).
Fourth, these models were based on environmental variables only, and did not account for historical or biological factors such as dispersal, evolution, or competition. If current patterns in waterfowl presence or abundance were influenced by such factors, then the resulting habitat model may predict theoretically suitable habitat rather than actual patterns in abundance or distribution. For example, the western limit of the American Black Ducks in central Canada may be because of biogeohistorical factors having nothing to do with current climate or topography. Further methodological research is necessary to explore the incorporation of such historical factors.
These models and the resulting maps suggest where, and for which species, existing range maps are in need of revision. For example, we predicted relatively high extralimital abundances of Ring-necked Duck. Having confirmed with raw survey data that our predictions accurately reflect observations of this species (Appendix 4), the range map of Ring-necked Duck could be updated. For other species, such as American Wigeon and Scoters, we predicted high extralimital abundances beyond the survey area. However, we would not recommend updating existing range maps for these species. Rather, in such cases, our predictions represent hypotheses derived from the existing data that should be tested against independent data.
Our models may also serve as guides for future modeling exercises or field studies. For instance, some future analyses may involve regional subsets of the data, to compare the variables identified as important in continental versus regional models. The prediction and associated uncertainty maps may guide where more research effort is needed, for instance in the southern Yukon where high abundances were predicted in association with high uncertainty. Future modeling focused on this area, particularly if accompanied by ground-truthing surveys, may identify important waterfowl areas for some species, or may simply show the current predictions to be an extrapolation error. In that case, definitive information would lead to reformulation of the models, possibly identifying more mechanistically informative variables. Extralimital predictions provide information for models of nonwaterfowl species as well. These unexpected occurrences can point to directions for future research in poorly surveyed areas of the breeding or wintering ranges.
Our goal was to produce predictive maps of waterfowl relative abundance, in large part to support the waterfowl conservation initiatives of Ducks Unlimited Canada. Although transect-level observations can yield an overall picture of waterfowl distributions and coarsely suggest some areas of high abundance, model prediction maps provide finer-scaled information over larger regions. The maps can also be used in future research to improve or compare modeling methods, to quantify aspects of waterfowl distribution ecology, and to evaluate and execute conservation planning strategies.
The U.S. Fish and Wildlife Service and Canadian Wildlife Service collected and supplied waterfowl data, and we acknowledge their extensive support regarding the database. Dan McKenney and Pia Papadopol of Natural Resources Canada supplied several climatic data sets. We acknowledge the valuable contributions of M. Houle and P. Racine, who performed all PostGIS operations. Spatial analysis and assembly were supported by the Boreal Avian Modeling Project, the Canada Research Chairs program, and the Canada Foundation for Innovation. N. Barker was funded by an Industrial Innovation Scholarship from Natural Sciences and Engineering Research Council of Canada and the Fonds québécois de la recherche sur la nature et les technologies, a Fellowship Grant from the Ducks Unlimited Canada-Institute for Wetland and Waterfowl Research, and a Natural Sciences and Engineering Research Council of Canada Discovery Grant to S. Cumming. We thank S. Slattery, J. Nowak, C. Roy, E. Racine, S. Renard, D. Stralberg, S. Bauduin, E. McIntire, T. Nudds, K. Swiston, an anonymous reviewer, and a subject editor for discussion and/or comments. This study benefited from funding by Ducks Unlimited Canada and Ducks Unlimited Inc.
Anderson, D. R. 1975. Population ecology of the mallard V. Temporal and geographic estimates of survival, recovery, and harvest rates. U.S. Fish and Wildlife Service resource publication 125. U. S. Fish and Wildlife Service, Washington, DC, USA.
Anderson, R. P., A. T. Peterson, and M. Gómez-Laverde. 2002. Using niche-based GIS modeling to test geographic predictions of competitive exclusion and competitive release in South American pocket mice. Oikos 98:3-16. http://dx.doi.org/10.1034/j.1600-0706.2002.t01-1-980116.x
Araújo, M. B., M. Cabeza, W. Thuiller, L. Hannah, and P. H. Williams. 2004. Would climate change drive species out of reserves? An assessment of existing reserve-selection methods. Global Change Biology 10:1618-1626. http://dx.doi.org/10.1111/j.1365-2486.2004.00828.x
Austin, J. E., D. A. Granfors, M. A. Johnson, and S. C. Kohn. 2002. Scaup migration patterns in North Dakota relative to temperatures and water conditions. Journal of Wildlife Management 66:874-882. http://dx.doi.org/10.2307/3803152
Austin, M. P. 2002. Spatial prediction of species distribution: an interface between ecological theory and statistical modeling. Ecological Modelling 157:101-118. http://dx.doi.org/10.1016/S0304-3800(02)00205-3
Badiou, P., R. Baldwin, M. Carlson, M. Darveau, P. Drapeau, K. Gaston, J. Jacobs, J. Kerr, S. Levin, M. Manseau, G. Orians, S. Pimm, H. Possingham, P. Raven, F. Reid, D. Roberts, T. L. Root, N. T. Roulet, J. Schaefer, D. Schindler, J. Stritthold, N. Turner, and J. Wells. 2013. Conserving the world’s last great forest is possible: here’s how. Briefing note. International Boreal Conservation Science Panel. [online] URL: http://borealscience.org/wp-content/uploads/2013/07/conserving-last-great-forests1.pdf
Barry, S., and J. Elith. 2006. Error and uncertainty in habitat models. Journal of Applied Ecology 43:413-423. http://dx.doi.org/10.1111/j.1365-2664.2006.01136.x
Bellrose, F. C. 1980. Ducks geese & swans of North America. Stackpole, Harrisburg, Pennsylvania, USA.
Berteaux, D. 2013. Québec’s large-scale Plan Nord. Conservation Biology 27:242-243. http://dx.doi.org/10.1111/cobi.12018
Börger, L., and T. D. Nudds. 2014. Fire, humans and climate: modeling distribution dynamics of boreal forest waterbirds. Ecological Applications 24:121–141. http://dx.doi.org/10.1890/12-1683.1
Brandt, J. P. 2009. The extent of the North American boreal zone Environmental Reviews 17:101-161. http://dx.doi.org/10.1139/A09-004
Canada Centre for Remote Sensing. 2008. Land cover map of Canada 2005. Natural Resources Canada, Ottawa, Ontario, Canada. [online] URL: ftp://ftp.ccrs.nrcan.gc.ca/ad/NLCCLandCover/LandcoverCanada2005_250m/
Cumming, S. G., D. Stralberg, K. L. Lefevre, P. Sólymos, E. M. Bayne, S. Fang, T. Fontaine, D. Mazerolle, F. K. A. Schmiegelow, and S. J. Song. 2014. Climate and vegetation hierarchically structure patterns of songbird distribution in the Canadian boreal region. Ecography 37:137-151. http://dx.doi.org/10.1111/j.1600-0587.2013.00299.x
Drever, M. C., R. G. Clark, C. Derksen, S. M. Slattery, P. Toose, and T. D. Nudds. 2012. Population vulnerability to climate change linked to timing of breeding in boreal ducks. Global Change Biology 18:480-492. http://dx.doi.org/10.1111/j.1365-2486.2011.02541.x
Dzubin, A. 1969. Assessing breeding populations of ducks by ground counts. Saskatoon Wetlands Seminar. Canadian Wildlife Service Report Series Number 6. Northern Prairie Wildlife Research Center Online. Northern Prairie Wildlife Research Center, Jamestown, North Dakota, USA. [online] URL: http://www.npwrc.usgs.gov/resource/birds/duckcoun/index.htm
Elith, J., C. H. Graham, R. P. Anderson, M. Dudík, S. Ferrier, A. Guisan, R. J. Hijmans, F. Huettmann, J. R. Leathwick, A. Lehmann, J. Li, L. G. Lohmann, B. A. Loiselle, G. Manion, C. Moritz, M. Nakamura, Y. Nakazawa, J. McC. M. Overton, A. T. Peterson, S. J. Phillips, K. Richardson, R. Scachetti-Pereira, R. E. Schapire, J. Soberón, S. Williams, M. S. Wisz, and N. E. Zimmermann. 2006. Novel methods improve prediction of species’ distributions from occurrence data. Ecography 29:129-151. http://dx.doi.org/10.1111/j.2006.0906-7590.04596.x
Elith, J., and J. Leathwick. 2009b. The contribution of species distribution modeling to conservation prioritization. Pages 70-93 in Spatial conservation prioritization: quantitative methods and computational tools. Oxford University Press, Oxford, UK.
Elith, J., and J. R. Leathwick. 2009a. Species distribution models: ecological explanation and prediction across space and time. Annual Review of Ecology, Evolution, and Systematics 40:677-697. http://dx.doi.org/10.1146/annurev.ecolsys.110308.120159
Elith, J., J. R. Leathwick, and T. Hastie. 2008. A working guide to boosted regression trees. Journal of Animal Ecology 77:802-813. http://dx.doi.org/10.1111/j.1365-2656.2008.01390.x
Environment Canada. 2010. Ecozone and ecoregion descriptions. Environment Canada, Ottawa, Ontario, Canada. [online] URL: http://ecozones.ca/english/zone/index.html
Ferrier, S. 2002. Mapping spatial pattern in biodiversity for regional conservation planning: where to from here? Systematic Biology 51:331-363. http://dx.doi.org/10.1080/10635150252899806
Friedman, J. H. 2001. Greedy function approximation: a gradient boosting machine. Annals of Statistics 29:1189-1232. http://dx.doi.org/10.1214/aos/1013203451
Friedman, J. H. 2002. Stochastic gradient boosting. Computational Statistics & Data Analysis 38:367-378. http://dx.doi.org/10.1016/S0167-9473(01)00065-2
Friedman, J., T. Hastie, and R. Tibshirani. 2000. Additive logistic regression: a statistical view of boosting (with discussion and a rejoinder by the authors). Annals of Statistics 28:337-407. http://dx.doi.org/10.1214/aos/1016218223
Friedman, J. H., and J. J. Meulman. 2003. Multiple additive regression trees with application in epidemiology. Statistics in Medicine 22:1365-1381. http://dx.doi.org/10.1002/sim.1501
Gaston, K. J. 2003. The structure and dynamics of geographic ranges. Oxford University Press, Oxford, UK.
Gimenez, O., S. J. Bonner, R. King, R. A. Parker, S. P. Brooks, L. E. Jamieson, V. Grosbois, B. J. T. Morgan, and L. Thomas. 2009. WinBUGS for population ecologists: Bayesian modeling using Markov chain Monte Carlo methods. Pages 883-915 in D. L. Thomson, E. G. Cooch, and M. J. Conroy, editors. Modeling demographic processes in marked populations. Volume 3. Springer US, New York, New York, USA. http://dx.doi.org/10.1007/978-0-387-78151-8_41
Guisan, A., and N. E. Zimmermann. 2000. Predictive habitat distribution models in ecology. Ecological modeling 135:147-186. http://dx.doi.org/10.1016/S0304-3800(00)00354-9
Hijmans, R. J., S. Phillips, J. Leathwick, and J. Elith. 2012. dismo: species distribution modeling. R package version 0.7-17. R Foundation for Statistical Computing, Vienna, Austria. [online] URL: http://CRAN.R-project.org/package=dismo
Hobson, K. A., M. B. Wunder, S. L. Van Wilgenburg, R. G. Clark, and L. I. Wassenaar. 2009. A method for investigating population declines of migratory birds using stable isotopes: origins of harvested lesser scaup in North America. PLoS ONE 4:e7915. http://dx.doi.org/10.1371/journal.pone.0007915
Horn, D. J., M. L. Phillips, R. R. Koford, W. R. Clark, M. A. Sovada, and R. J. Greenwood. 2005. Landscape composition, patch size, and distance to edges: interactions affecting duck reproductive success. Ecological Applications 15:1367-1376. http://dx.doi.org/10.1890/03-5254
Integrated Remote Sensing Studio. 2010. Canadian boreal ruggedness. Data Basin. Conservation Biology Institute, Corvallis, Oregon, USA. [online] URL: http://databasin.org/datasets/14d70746535e4be99aaf66595cc0b677
Jamieson, L. E., and S. P. Brooks. 2004. Density dependence in North American ducks. Animal Biodiversity and Conservation 27:113-128.
Jetz, W., D. S. Wilcove, and A. P. Dobson. 2007. Projected impacts of climate and land-use change on the global diversity of birds. PLoS Biology 5:e157. http://dx.doi.org/10.1371/journal.pbio.0050157
Johnsgard, P. 2010. Waterfowl of North America. Revised edition. University of Nebraska-Lincoln Libraries, Lincoln, Nebraska, USA.
Johnson, D. H., and J. W. Grier. 1988. Determinants of breeding distributions of ducks. Wildlife Monographs 100:3-37.
Johnson, D. H., and T. L. Shaffer. 1987. Are mallards declining in North America? Wildlife Society Bulletin 15:340-345.
Johnson, W. C., B. V. Millett, T. Gilmanov, R. A. Voldseth, G. R. Guntenspergen, and D. E. Naugle. 2005. Vulnerability of northern prairie wetlands to climate change. BioScience 55:863-872. http://dx.doi.org/10.1641/0006-3568(2005)055[0863:VONPWT]2.0.CO;2
Kaminski, R. M., and E. A. Gluesing. 1987. Density- and habitat-related recruitment in mallards. Journal of Wildlife Management 51:141-148. http://dx.doi.org/10.2307/3801645
Lawrence, J. D., R. B. Gramacy, L. Thomas, and S. T. Buckland. 2013. The importance of prior choice in model selection: a density dependence example. Methods in Ecology and Evolution 4:25-33. http://dx.doi.org/10.1111/j.2041-210X.2012.00255.x
Leathwick, J. R. 1998. Are New Zealand’s Nothofagus species in equilibrium with their environment? Journal of Vegetation Science 9:719-732. http://dx.doi.org/10.2307/3237290
Lemelin, L.-V., M. Darveau, L. Imbeau, and D. Bordage. 2010. Wetland use and selection by breeding waterbirds in the boreal forest of Quebec, Canada. Wetlands 30:321-332. http://dx.doi.org/10.1007/s13157-010-0024-z
Loiselle, B. A., C. A. Howell, C. H. Graham, J. M. Goerck, T. Brooks, K. G. Smith, and P. H. Williams. 2003. Avoiding pitfalls of using species distribution models in conservation planning. Conservation Biology 17:1591-1600. http://dx.doi.org/10.1111/j.1523-1739.2003.00233.x
Lu, H. H.-S., B. Schölkopf, and H. Zhao, editors. 2011. Handbook of statistical bioinformatics. Springer, New York, New York, USA. http://dx.doi.org/10.1007/978-3-642-16345-6
Lynch, H. J., M. Rhainds, J. M. Calabrese, S. Cantrell, C. Cosner, and W. F. Fagan. 2014. How climate extremes—not means—define a species’ geographic range boundary via a demographic tipping point. Ecological Monographs 84:131-149. http://dx.doi.org/10.1890/12-2235.1
Mac Nally, R. 2000. Regression and model-building in conservation biology, biogeography and ecology: the distinction between—and reconciliation of—‘predictive’ and ‘explanatory’ models. Biodiversity & Conservation 9:655-671. http://dx.doi.org/10.1023/A:1008985925162
McIntire, E. J. B., and A. Fajardo. 2009. Beyond description: the active and effective way to infer processes from spatial patterns. Ecology 90:46-56. http://dx.doi.org/10.1890/07-2096.1
McKenney, D. W., M. F. Hutchinson, P. Papadopol, K. Lawrence, J. Pedlar, K. Campbell, E. Milewska, R. F. Hopkinson, D. Price, and T. Owen. 2011. Customized spatial climate models for North America. Bulletin of the American Meteorological Society 92:1611-1622. http://dx.doi.org/10.1175/2011BAMS3132.1
Milsom, T. P., S. D. Langton, W. K. Parkin, S. Peel, J. D. Bishop, J. D. Hart, and N. P. Moore. 2000. Habitat models of bird species’ distribution: an aid to the management of coastal grazing marshes. Journal of Applied Ecology 37:706-727. http://dx.doi.org/10.1046/j.1365-2664.2000.00529.x
Moilanen, A., A. M. A. Franco, R. I. Early, R. Fox, B. Wintle, and C. D. Thomas. 2005. Prioritizing multiple-use landscapes for conservation: methods for large multi-species planning problems. Proceedings of the Royal Society B: Biological Sciences 272:1885-1891. http://dx.doi.org/10.1098/rspb.2005.3164
Morrison, M. L., B. Marcot, and W. Mannan. 2006. Wildlife-habitat relationships: concepts and applications. Island Press, Washington, D.C., USA.
Murkin, H. R., R. M. Kaminski, and R. D. Titman. 1982. Responses by dabbling ducks and aquatic invertebrates to an experimentally manipulated cattail marsh. Canadian Journal of Zoology 60:2324-2332. http://dx.doi.org/10.1139/z82-299
Natural Resources Canada. 2007. National hydro network, Canada. Natural Resources Canada, Earth Sciences Sector, Centre for Topographic Information, Sherbrooke, Quebec, Canada.
Nichols, J. D., F. A. Johnson, and B. K. Williams. 1995. Managing North American waterfowl in the face of uncertainty. Annual Review of Ecology and Systematics 26:177-199. http://dx.doi.org/10.1146/annurev.es.26.110195.001141
Obe, R., and L. Hsu. 2011. PostGIS in action. Manning, Greenwich, Connecticut, USA.
Oppel, S., A. Meirinho, I. Ramírez, B. Gardner, A. F. O’Connell, P. I. Miller, and M. Louzao. 2012. Comparison of five modeling techniques to predict the spatial distribution and abundance of seabirds. Biological Conservation 156:94-104. http://dx.doi.org/10.1016/j.biocon.2011.11.013
Ortega-Huerta, M. A., and A. T. Peterson. 2004. modeling spatial patterns of biodiversity for conservation prioritization in North-eastern Mexico. Diversity and Distributions 10:39-54. http://dx.doi.org/10.1111/j.1472-4642.2004.00051.x
Paszkowski, C. A., and W. M. Tonn. 2006. Foraging guilds of aquatic birds on productive boreal lakes: environmental relations and concordance patterns. Hydrobiologia 567:19-30. http://dx.doi.org/10.1007/s10750-006-0053-z
Pimm, S. L. 1994. The importance of watching birds from airplanes. Trends in Ecology & Evolution 9:41-43. http://dx.doi.org/10.1016/0169-5347(94)90264-X
Podruzny, K. M., J. H. Devries, L. M. Armstrong, and J. J. Rotella. 2002. Long-term response of northern pintails to changes in wetlands and agriculture in the Canadian Prairie Pothole Region. Journal of Wildlife Management 66:993-1010. http://dx.doi.org/10.2307/3802932
Pospahala, R. S., D. R. Anderson, and C. J. Henny. 1974. Population ecology of the Mallard II: breeding habitat conditions, size of the breeding populations, and production indices. U.S. Fish and Wildlife Service Resource Publication. U.S. Fish and Wildlife Service, Washington, D.C., USA.
R Core Team. 2012. R: a language and environment for statistical computing. R Foundation for Statistical Computing, Vienna, Austria. [online] URL: http://www.R-project.org/
Randin, C. F., T. Dirnböck, S. Dullinger, N. E. Zimmermann, M. Zappa, and A. Guisan. 2006. Are niche-based species distribution models transferable in space? Journal of Biogeography 33:1689-1703. http://dx.doi.org/10.1111/j.1365-2699.2006.01466.x
Raveling, D. G., and M. E. Heitmeyer. 1989. Relationships of population size and recruitment of pintails to habitat conditions and harvest. Journal of Wildlife Management 53:1088-1103. http://dx.doi.org/10.2307/3809615
Ridgeway, G. 2013. gbm: generalized boosted regression models. R package version 2.1. R Foundation for Statistical Computing, Vienna, Austria. [online] URL: http://CRAN.R-project.org/package=gbm
Ross, B. E., M. B. Hooten, and D. N. Koons. 2012. An accessible method for implementing hierarchical models with spatio-temporal abundance data. PLoS ONE 7:e49395. http://dx.doi.org/10.1371/journal.pone.0049395
Savard, J.-P. L., D. Bordage, and A. Reed. 1998. Surf Scoter (Melanitta perspicillata). The birds of North America online. Cornell Lab of Ornithology, Ithaca, New York, USA. [online] URL: http://dx.doi.org/10.2173/bna.363
Schindler, D. W., and P. G. Lee. 2010. Comprehensive conservation planning to protect biodiversity and ecosystem services in Canadian boreal regions under a warming climate and increasing exploitation. Biological Conservation 143:1571-1586. http://dx.doi.org/10.1016/j.biocon.2010.04.003
Schummer, M. L., A. D. Afton, S. S. Badzinski, S. A. Petrie, G. Olsen, K. Jacobs, M. Mitchell, and S. Jenkins. 2013. Using satellite telemetry to evaluate the effectiveness of the waterfowl breeding population and habitat survey for counting lesser and greater scaup in North America. Oral presentation. Ecology and Conservation of North American Waterfowl, Memphis, Tennessee, USA. [online] URL: http://www.northamericanducksymposium.org/docs/New%20Approaches%20and%20Methods.pdf
Silverman, E. D. 2011. Waterfowl breeding population and habitat survey—traditional survey. Division of Migratory Bird Management, U.S. Fish and Wildlife Service, Laurel, Maryland, USA. [online] URL: https://migbirdapps.fws.gov/mbdc/databases/mas/aboutmas.htm
Slattery, S. M., J. L. Morissette, G. G. Mack, and E. W. Butterworth. 2011. Waterfowl conservation planning: science needs and approaches. Pages 23-40 in J. V. Wells, editor. Boreal birds of North America: a hemispheric view of their conservation links and significance. Studies in Avian Biology No. 41. Cooper Ornithological Society, University of California Press, Oakland, California, USA.
Smith, G. W. 1995. A critical review of the aerial and ground surveys of breeding waterfowl in North America. Biological science report. U.S. National Biological Service, Washington, D.C., USA.
Sólymos, P., S. M. Matsuoka, E. M. Bayne, S. R. Lele, P. Fontaine, S. G. Cumming, D. Stralberg, F. K. A. Schmiegelow, and S. J. Song. 2013. Calibrating indices of avian density from non-standardized survey data: making the most of a messy situation. Methods in Ecology and Evolution 4:1047-1058. http://dx.doi.org/10.1111/2041-210X.12106
Stralberg, D., S. M. Matsuoka, A. Hamann, E. M. Bayne, P. Sólymos, F. Schmiegelow, X. Wang, S. G. Cumming, and S. J. Song. 2014. Projecting boreal bird responses to climate change: the signal exceeds the noise. Ecological Applications, in press. http://dx.doi.org/10.1890/13-2289.1
Suhonen, S., P. Nummi, and H. Poysa. 2011. Long term stability of boreal lake habitats and use by breeding ducks. Boreal Environment Research 16:71-80.
Thuiller, W. 2004. Patterns and uncertainties of species’ range shifts under climate change. Global Change Biology 10:2020-2027. http://dx.doi.org/10.1111/j.1365-2486.2004.00859.x
Thuiller, W., L. Brotons, M. B. Araújo, and S. Lavorel. 2004. Effects of restricting environmental range of data to project current and future species distributions. Ecography 27:165-172. http://dx.doi.org/10.1111/j.0906-7590.2004.03673.x
Titman, R. D. 1999. Red-breasted Merganser (Mergus serrator). The birds of North America online. Cornell Lab of Ornithology, Ithaca, New York, USA. http://dx.doi.org/10.2173/bna.443
U.S. Fish and Wildlife Service. 2012. Waterfowl population status, 2012. Report. U.S. Department of the Interior, Washington, D.C., USA.
Wenger, S. J., and J. D. Olden. 2012. Assessing transferability of ecological models: an underappreciated aspect of statistical validation. Methods in Ecology and Evolution 3:260-267. http://dx.doi.org/10.1111/j.2041-210X.2011.00170.x
Wiken, E., F. Jiménez Nava, and G. Griffith. 2011. North American terrestrial ecoregions—level III. Commission for Environmental Cooperation, Montreal, Quebec, Canada.
Zimmerman, G. S., J. R. Sauer, W. A. Link, and M. Otto. 2012. Composite analysis of black duck breeding population surveys in eastern North America. Journal of Wildlife Management 76:1165-1176. http://dx.doi.org/10.1002/jwmg.351