ECOGRAPHY 27: 350
/360, 2004

Evaluating alternative data sets for ecological niche models of birds in the Andes Juan L. Parra, Catherine C. Graham and Juan F. Freile

Parra, J. L., Graham, C. C. and Freile, J. F. 2004. Evaluating alternative data sets for ecological niche models of birds in the Andes.
/ Ecography 27: 350
/360. Ecological niche modeling (ENM) is an effective tool for providing innovative insights to questions in evolution, ecology and conservation. As environmental datasets accumulate, modelers need to evaluate the relative merit of different types of data for ENM. We used three alternative environmental data sets: climatic data, remotesensing data (Normalized Difference Vegetation Index), and elevation data, to model the distribution of six bird species of the genus Grallaria in the Ecuadorian Andes. We assessed the performance of models created with each environmental data set and all possible combinations by comparing the geographic predictions of our models with detailed maps developed by expert ornithologists. Results varied depending on the specific measure of performance. Models including climate variables performed relatively well across most measures, whereas models using only NDVI performed poorly. Elevation based models were relatively good at predicting most sites of expected occurrence but showed a high over-prediction error. Combinations of data sets usually increased the performance of the models, but not significantly. Our results highlight the importance of including climatic variables in ENM and the simultaneous use of various data sets when possible. This strategy attenuates the effects of specific variables that decrease model performance. Remote-sensing data, such as NDVI, should be used with caution in topographically complex regions with heavy cloud-cover. Nonetheless, remote-sensing data have the potential to improve ENM. Finally, we suggest a priori designation of modeling purposes to define specific performance measures accordingly. J. L. Parra ([email protected]) and C. C. Graham, Museum of Vertebrate Zoology, 3101 VLSB, Univ. of California, Berkeley, CA 94720-3160 USA.
/ J. F. Freile, Fundacio´n Numashir para la Conservacio´n de Ecosistemas Amenazados, Casilla Postal 17-12-122, Quito, Ecuador.

Ecological niche models have been used to study issues in evolution (Peterson 2001, Hugall et al. 2002), ecology (Anderson et al. 2002), and conservation (Godown and Peterson 2000, Sa´nchez-Cordero and Martı´nez-Meyer 2000, Peterson and Robins 2003). These methods (e.g. BIOCLIM-Nix 1986, Busby 1991; GARP-Stockwell and Noble 1991; DOMAIN-Carpenter et al. 1993) combine geographic locations of a given species with spatial surfaces of environmental data to identify suitable parameters for a given species and then map this information to predict the species geographic distribution. Typically, interpolated climate data (e.g. Berry et al.

2002, Joseph and Stockwell 2002); or environmental data obtained through remote sensing (e.g. Fuentes et al. 2001, Oindo 2002, Zinner et al. 2002) are used to build models. To date, there has been no assessment of the relative performance of models created by these different datasets. In this article, we assess the utility of using climate, remotely-sensed Normalized Difference Vegetation Index (NDVI) data, elevation, and a combination of these variables to predict distributions for six bird species inhabiting the Ecuadorian Andes. Interpolated climate data are derived from direct measurements of climate at weather stations (New et

Accepted 29 December 2003 Copyright # ECOGRAPHY 2004 ISSN 0906-7590

350

ECOGRAPHY 27:3 (2004)

al. 1999). These data usually cover a considerable span of time (e.g. 30 yr) and are sensitive to error occurring during data processing and archiving. Coverage of weather stations is not uniform in space or time, and while there is information for some large areas with high resolution (e.g. Canada, Australia), information is still scarce for other areas (e.g. Amazonia). NDVI is a measure of the reflectance of earth’s surface vegetation (Lillesand and Kiefer 2000), and is representative of the leaf area index (Asrar et al. 1984) and vegetation productivity (Chong et al. 1993). While NDVI data has many positive attributes for modeling (i.e. global coverage at a 1.1 km2 resolution), it is susceptible to interference from moisture and/or background noise (Huete 1989). Elevation data is often used in models as a covariate of an environmental variable (Hutchinson 2000), or as a descriptor of a physical characteristic such as slope (Berry et al. 2002). Remote sensing programs such as the Shuttle Radar Topography Mission ( B/http://www.jpl.nasa.gov/srtm//) are further increasing the accuracy and resolution of digital elevation surfaces. Climate, NDVI, and elevation data sets are often related (Henricksen and Durkin 1986, Hutchinson 1998, Richard and Poccard 1998, Ichii et al. 2002, Oindo 2002); for example, in South America, ‘‘greener’’ areas have high rainfall and low temperatures (Ichii et al. 2002). Likewise, elevation shows a close relationship to temperature. Nevertheless, the nature of these relationships is variable in space and time (Ichii et al. 2002). Assuming a correlation between NDVI values and climate parameters (Fjeldsa˚ et al. 1997, Oindo 2002), we expect model performance by each data set to be similar in terms of the correctly predicted occurrences, in relation to 1) all expected occurrences (sensitivity) and 2) all predicted occurrences (positive predictive powerPPP). We also expect models based on each of these variables to have similar spatial predictions of species distributions. However, if data sets are not correlated, predictions should be different and possibly complementary, resulting in an overall improvement of model performance when datasets are combined. While climate and NDVI are similar in terms of the biotic and abiotic factors they measure, there are also several differences. These differences include temporal, spatial coverage, and sampling effort. NDVI data are available for short periods within the last 20 yr while climate data span up to 30 yr. In addition, at the time of this study, there was no global climate coverage available at a 1 km2 resolution, as with NDVI. NDVI data are collected such that a systematic reading is taken for each pixel. In a climate surface, pixel values are interpolated from weather station data. Hence, NDVI data should reflect spatial and temporal variation
/ for the space and time covered
/ at a higher resolution than climate data. Based on the consistent and recent coverage by remote ECOGRAPHY 27:3 (2004)

sensing, we predict that NDVI based models will be more representative of current distribution boundaries. Within the range map of a given species, NDVI based models should have less over-prediction (commission error), or in other words, higher specificity
/ higher probability of correctly predicting a cell as absent. Exclusion of old locality records which reflect available habitat in previous times that has recently been disturbed, should improve the performance of NDVI models. In order to assess the utility for modeling of climate, NDVI, and elevation data sets, we compared predicted versus expected distributions of six bird species in the Ecuadorian Andes. Predictive models were evaluated with expert knowledge based maps developed by Krabbe et al. (1998) for the highlands (/1200 m a.s.l.) of Ecuador. The high levels of endemism and species richness characteristic of the Ecuadorian Andes, combined with a long history of anthropogenic use that has left many species threatened through habitat alteration, makes it an area of high conservation priority (Renjifo et al. 1997, Krabbe et al. 1998, Sierra et al. 2002). We hope this intent will encourage similar modeling studies in other countries and eventually at large biogeographic scales.

Methods: data sets Seven data sets were used to develop ecological niche models: 1) climate data set compiled and developed by C. Graham for Ecuador, 2) NDVI data set available from the ‘‘Land Processes Distributed Active Archive Center’’ (B/http://edcdaac.usgs.gov/1KM/1kmhomepage. html/), 3) elevation data set developed by Sierra (1999), 4) climate/elevation, 5) NDVI/elevation, 6) NDVI/climate, and 7) NDVI/climate/elevation. All grid layers were scaled to 1 km2 grid cell size. Climate surfaces for Ecuador were created using an interpolation technique based on thin plate smoothing splines (ANUSPLIN, Hutchinson 2000). Monthly climate surfaces were extrapolated for total precipitation and average temperature. Thin plate smoothing splines are a generalization of standard multi-variate linear regression in which the parametric model is replaced by a suitable smooth non-parametric function (Hutchinson 1995). We included three independent spline variables: longitude, latitude, and elevation to interpolate climate surfaces. Elevation was incorporated because temperature and rainfall are often highly correlated with elevation, and inclusion of elevation in the model reduced statistical error (Hutchinson 1991). We used a digital elevation model developed for Ecuador at a 0.16 km2 resolution by Sierra (1999). We included 264 rainfall points spread across Ecuador. The standard predictive errors ranged from 11.3% to 16.5%, which is 351

similar to errors in other topographically complex areas (Faith et al. 2001). The number of temperature data points was smaller than rainfall but the strong dependence of temperature on elevation enabled us to create accurate surfaces. We used data from 163 climate stations to create maximum and minimum monthly temperature surfaces. The standard predictive errors varied between 2.6% to 3.1% and 2.8% to 3.3% for maximum and minimum temperatures; respectively. Small errors such as these are typical for temperature surfaces (Hutchinson 1991, 2000). We used ANUCLIM ver. 5.1 package (Houlder et al. 2000), to create bioclimatic parameters that are biologically meaningful combinations of monthly climate variables (Nix 1986). These surfaces included annual mean temperature, annual total rainfall, monthly coefficients of variation for temperature and rainfall, and dry quarter-three consecutive months with the least total precipitation (grids available upon request from C. Graham). All climate grids were resampled to 1 km2 resolution. NDVI measurements used in this study were captured by the Advanced Very High Resolution Radiometer (AVHRR) carried by the National Oceanic and Atmospheric Administration’s (NOAA) satellite. NDVI is calculated as the difference between the reflectance readings in the near infrared (channel 2) and visible (channel 1) light spectrum, and normalized over the sum of both readings: (Channel 2 /Channel 1)/(Channel 1/ Channel 2). Raw values range from /1.0 to 1.0; positive high values indicate highly vegetated areas (almost all visible spectrum is absorbed plus a high reflectance in the near infrared spectrum), and lower negative values non-vegetated features or cloud-covered areas (all visible spectrum is reflected; Holben 1986). Raw readings are processed according to international agreements ( B/http://edcdaac.usgs.gov/1KM/paper.html /) and final NDVI products are available online. Values are rescaled to a 0
/200 scale, where values B/100 are considered non-vegetated. We used NDVI data from two years, April 1992 to March 1993 and February 1995 to January 1996. Data are downloadable as maximum readings from 10-day composites (maximum NDVI

value obtained during a 10 day time lapse). To eliminate additional noise by cloud interference or soil reflectance, maximum monthly values out of the two years of data were used to calculate measures representative of productivity and seasonality. We calculated the following monthly layers based on monthly maximum values: overall maximum NDVI, overall minimum NDVI, mean annual NDVI, coefficient of variation, maximum and minimum quarters (three consecutive months with the maximum or minimum mean value respectively), annual seasonality (100/[overall maximum/overall minimum]/overall maximum, Hurlbert and Haskell 2003), and dense and medium ‘‘greenness’’ (number of months where each pixel had a value higher than 150 or between 109 and 150 respectively, Holben 1986). We tested for correlation among variables within and between data sets by plotting 1000 random points within Ecuador and extracting associated environmental values. Pearson’s correlations were performed among all variables. We excluded variables that had a coefficient of correlation /0.7 (Green 1979) within any data set (i.e. among climate variables). When datasets were used in conjunction (e.g. climate and elevation) we included all variables even if they were correlated (Table 1). Four climatic variables (total annual rainfall, annual monthly mean temperature, coefficient of variation in monthly temperature and monthly rainfall), three NDVI derived variables (overall maximum NDVI, annual seasonality, and medium ‘‘greenness’’), and the elevation data were used as predictor variables.

Point localities We modeled the distribution of six species of antpittas (Grallaria quitensis, G. ruficapilla , G. rufula , G. nuchalis, G. squamigera , and G. hypoleuca , Passeriformes, Formicariidae) for which expert opinion distribution maps were published (Krabbe et al. 1998), and /40 different locality points were available for Ecuador (Freile 2000). Locality points were gathered from major collections and published records. Bird specimen data were gathered

Table 1. Pearson’s correlation coefficients among variables used in modeling excercise. Variables within a dataset with a correlation coefficient /0.7 were not included in the analysis.

1 2 3 4 5 6 7 8 9

Environmental variables

1

2

3

4

5

6

7

8

9

Annual monthly mean temperature CV monthly temperature Total annual rainfall CV monthly rainfall Dry quarter Elevation Maximum NDVI Annual seasonality Medium greenness

1.00 0.38 0.43 0.24 0.27 0.99* 0.20 0.15 0.43

1.00 0.42 0.35 0.36 0.41 0.04 0.03 0.01

1.00 0.66 0.88* 0.45 0.12 0.16 0.15

1.00 0.87* 0.27 0.22 0.27 0.35

1.00 0.30 0.23 0.26 0.33

1.00 0.19 0.14 0.41

1.00 0.35 0.35

1.00 0.35

1.00

* Pearson’s correlation coefficient /0.7.

352

ECOGRAPHY 27:3 (2004)

from the Academy of Natural Sciences of Philadelphia (ANSP), American Museum of Natural History (AMNH), British Museum of Natural History (BMNH), Carnegie Museum of Natural History (CMNH), Field Museum of Natural History (FMNH), Louisiana State Univ. Museum of Zoology (LSUMZ), Museo Ecuatoriano de Ciencias Naturales (MECN), Museo de Zoologı´a, Pontificia Univ. Cato´lica del Ecuador (QCAZ), Museo de Zoologı´a, Escuela Polite´cnica Nacional del Ecuador (EPN), U.S. National Museum of Natural History (NMNH), and the Western Foundation of Vertebrate Zoology (WFVZ). JFF verified identification and locality of published and museum records for birds. Points were checked for concordance between the reported and projected elevation correcting or removing points with an incongruence /500 m. No duplicate point localities were included in the models. We used a recent vegetation map developed by Sierra et al. (1999) to identify points within currently deforested areas and measure the change in NDVI based model performance when excluding those points from the analysis. We assumed deforested or heavily degraded areas were not suitable for any of the bird species included in the analyses (Freile 2000).

Model We used BIOCLIM as the modeling technique to simulate species distributions. BIOCLIM is a simple ‘‘envelope’’ or ‘‘profile matching’’ technique where a rule set, based on the variables included in the dataset is developed (Busby 1991). After a bioclimatic profile is obtained, the program determines all possible locations

on the grid with similar environmental characteristics to produce a map of the predicted potential distribution of a species in realized geographic space. The predicted distribution was based on the environmental envelope included between the 5th and 95th percentiles of the total environmental space; in other words, points that fell within the 5 tail percentiles of the total environmental space (10% of the total environmental space) were removed.

Model performance The performance of models should be measured relative to the particular research objective (Manel et al. 2001). In our case, the purpose was to closely predict geographic areas that were ‘‘suitable’’ or where we expected a species to occur, based on environmental correlates. In order to obtain a measure of the accuracy of the output, we compared the predictions from each dataset versus expert distribution maps limiting our analyses to the area encompassed within the proposed range map of each species (Fig. 1). We limited our comparison to cells within range maps (obtained from R. S. Ridgely as digital files) to create relatively objective and consistent confusion matrices (Appendix). Ecological niche models often over predict to areas where the organism is absent due to ecological (i.e. there is an extra dimension in the model, such as a competing species, that was not included) or historical reasons (i.e. a geographic barrier has impeded a species to colonize). This over prediction is difficult to evaluate, especially when areas of predicted occurrence are not close to current proposed ranges. By taking only areas within the range map, we restrict the

Fig. 1. Expert knowledge and predictive BIOCLIM distribution maps of Grallaria nuchalis in Ecuador. Only areas within the range map were included in the analysis. Black areas represent 5
/95 percentiles of BIOCLIM’s prediction.

ECOGRAPHY 27:3 (2004)

353

possible interpretations and assume that the absence of a species from any area within the range map is due to niche unsuitability. We used the predicted model’s results within range maps to build a confusion matrix taking as true positives (a) the number of correctly predicted occurrence cells (within the range map), false positives (b) the number of grid cells predicted within the range but outside the expert map, false negatives (c) the number of grid cells incorrectly predicted as absent, and true negatives (d) as the cells that were correctly predicted absent. From the confusion matrix we derived: sensitivity (proportion of correctly predicted occurrence cells in relation to all expected cells, a/[a/c]), specificity (proportion of cells correctly predicted absent cells in relation to all absent cells, d/[b/d]), positive predictive power- PPP (proportion of correctly predicted cells in relation to all predicted occurrence cells, a/[a/b]), and Kappa (defined as the accuracy of the prediction in relation to that expected by chance alone, Farber and Kadmon 2003, [(a/ d) / (((a/ c)(a/b)/ (b/d)(c/d))/N)]/[N / (((a/ c)(a/b)/(b/d)(c/d))/N)], where N is the sum of all cases, Fielding and Bell 1997). To test for statistical differences in the average of each performance measure among models based on different data sets, we used analysis of variance (ANOVA) with a repeated measures design (Zar 1996). Data were transformed using arcsinâ, which is the suggested transformation for proportions (Zar 1996). Non-parametric tests (Friedman’s test) were performed when normality or homogeneity of variance was violated. If a significant ANOVA was obtained, we used Tukey post-hoc tests, or their non-parametric equivalent, to evaluate which datasets were different. We used paired T-tests to evaluate significant changes in Kappa values of NDVI models when excluding points in deforested areas. All data were checked for normality and homogeneity of variances. Statistical tests were performed using statgraphics plus 5.1 software, except for the Friedman’s test, which were done by hand following Zar (1996).

Results Model performance Average performance of models based on different data sets was statistically different in terms of PPP (x2r(0.05,7,6) /31.142, pB/005). Models including elevation and climate performed better in terms of PPP, relative to models based only on elevation or NDVI (Fig. 2). Elevation based models had a higher sensitivity
/ predicted a higher proportion of the true area of occurrence
/ than any of the other models (F0.05(2)6,30 /55.337, pB/0.05, Tukey post hoc tests were all significant). Average specificity
/ the proportion of correctly predicted absence
/ was significantly 354

lower for both elevation and NDVI only based models (x2r(0.05,7,6) /33.5, pB/0.05), but not different from each other (q0.05,8,7 /0.378, p/0.05). In terms of Kappa values, the most accurate models were based on combined datasets of climate and elevation (x2r(0.05,7,6) / 21.571, pB/0.05). Nonetheless, differences among datasets were not significant except in relation to NDVI based models, which performed significantly worse (various q0.05,8,7 /4.17, pB/0.05) (Fig. 2). To better understand how Kappa values were related to the other performance measures assessed, we created pair-wise plots relating two specific performance measures for each species prediction with their respective Kappa value (Fig. 3). There was no consistent trend for all performance measures. When PPP was plotted against specificity or sensitivity, high PPP values often reflected high Kappa values, whereas there was no trend for specificity or sensitivity. When specificity was plotted against sensitivity, usually intermediate values of both measurements reflected high Kappa values, except for elevation based models which showed high Kappa values but low specificity. The overall accuracy (Kappa) of NDVI based models did not improve (maximum increase of Kappa/0.08, T0.05(1),5 / /0.083, p/0.05) when points in currently deforested areas (sensu Sierra et al. 1999) were left out. The only two species for which the Kappa values slightly increased (G. quitensis and G. ruficapilla ), were the ones where most points were removed from the model (38 and 48% of total points available, respectively).

Predicted areas Area predicted by elevation and NDVI based models within range maps was always larger than area predicted by any other model (Table 2). This trend was consistent when total predicted area was analyzed. On average, elevation based models predicted two times more area than climate based models (range 1.47
/3.01) when results only within range maps were analyzed, and 2.8 times more when total area was considered (range 1.41
/ 6.68). NDVI based models showed a similar trend, predicting even total larger areas in some instances (Table 2).

Discussion One of the main goals of ecological niche modeling is to accurately predict the distribution of a species at the present time. Climate data is an obvious choice but of limited availability and quality in some areas. Remotely sensed data, such as NDVI are being increasingly used for this purpose (Fuentes et al. 2001, Berry et al. 2002, Zinner et al. 2002). Global coverage at ca 1-km ECOGRAPHY 27:3 (2004)

Fig. 2. Box plots summarizing results of measures of performance for each dataset used. Median values (line across box), range excluding outliers (error bars), interquartile range containing 50% of values (box), and outliers (circles) from results of six species for each performance measure evaluated across data sets (E/elevation, C /climate, N/NDVI). Untransformed values were used.

resolution and public accessibility are two characteristics that make the NDVI data set attractive. However, in the present study we found that models based solely on NDVI generally performed poorly in comparison to those created using climate and elevation data. Models based on our seven data sets varied in terms of the performance measures evaluated (PPP, specificity and sensitivity), specific agreement (Kappa), and area that they predicted. No obvious congruence or tendency was found across different measures of performance and we could not easily identify a set of environmental layers that consistently performed well across all performance measures. This result indicates that choice of environmental variables can be complex and the measure of performance should be specific to the type of errors, which should be minimized for a given application (Loiselle et al. 2003, Fielding and Bell 1997). Given that relative model performance was not consistent across evaluation criteria, it may be necessary to assign prior cost values to specific types of errors based on the particular objectives of the modeling procedure (Fielding and Bell 1997). For example, conECOGRAPHY 27:3 (2004)

servationists often prioritize models with high sensitivity and PPP because failure to predict actual areas of occurrence may be more costly than falsely predicting them (Fielding 2002, Loiselle et al. 2003). In essence, by maximizing these two measures, the probability of omission error
/ not including an area where a species occurs in the prediction
/ is decreased, and the accuracy of occurrence predictions is increased, as long as commission error (false occurrence predictions) is maintained. In our case, PPP did not vary significantly across most models; however, models built only with elevation data had higher sensitivity than models based on any other dataset. Interestingly, elevation based models achieve a high sensitivity (no areas of occurrence are omitted) by sacrificing specificity through a high overprediction. Hence, final decisions on which models are the most reliable, should take all types of error into account and how each error could compromise the specific research objective (Fig. 3). Application of ecological niche models in ecology and evolution has raised the importance of measuring and interpreting commission errors (Anderson et al. 355

Fig. 3. Bubble charts showing relationships between specific measures of performance (PPP, sensitivity and specificity) represented by the axes, and Kappa values, represented by the surface area of the circles.

2003)
/ errors of over-prediction. In principle, if environmental variables included in the model are appropriate, over-prediction can indicate that other other factors may be determining geographic distributions such as historical connectivity or competitive exclusion by closely related or ecologically similar taxa. Specificity, which measures the proportion of correctly predicted absent cells to all expected absent cells (Fielding 2002), was highest in climate-based models. Models able to differentiate suitable and non-suitable areas within the range map should be the most reliable in terms of over-predictions. Range maps are large-scale expectations of where a species can be found and areas not inhabited within range maps represent inappropriate habitat (Ridgely and Tudor 1994) rather than other factors. Thus, we argue that climate based models did a fairly good job determining unsuitable areas within the range map and also achieve a relatively high PPP. Elevation and NDVI based models had significantly lower specificities than any of the other models, again 356

highlighting the fact that these models attain a high sensitivity with large over predictions. When elevation and NDVI datasets were combined, their specificity increased significantly, suggesting datasets were predicting distinct areas of expected absence, and thus when joined, were complementary. Elevation data generally were able to delineate the range of a given species but they did not detect smaller-scale changes within boundaries of the range maps. However, NDVI is more likely to capture small isolated patches where human intervention or any type of disturbance has caused the vegetation to change. Hence, elevation and NDVI might be detecting different areas based on the particular nature of each dataset. Finally, performance measures, such as Kappa, make use of all the information in a confusion matrix and give us a picture of the specific agreement of the model over chance (Fielding and Bell 1997). NDVI based models had the lowest Kappa value. This is likely a result of its particularly low specificity and PPP. In each of the other datasets, the average Kappa values are similar, reflecting in most cases intermediate values of sensitivity and specificity, except in the elevation-only based models, which had both significantly high sensitivity and low specificity (Fig. 3). In general, average Kappa values were relatively low (B/0.4, Landis and Koch 1977) suggesting that model predictions do a slightly better job than chance. Various aspects of the model such as the quality of raw data, the predictor variables and modeling procedure employed, could be contributing to a general low performance. The quality of the environmental data sets can be evaluated through their estimated error. Temperature interpolations had a low error estimate in relation to rainfall, which had a relatively high error because of the complex topography of the Andes. Nonetheless, species ecological niche models have been successfully created with similar error rates in other areas (Faith et al. 2001). Unfortunately, we have no error estimates for NDVI or elevation surfaces; estimation of these errors is out of our reach. One of the potential sources of error in remotely-sensed data such as NDVI is interference with clouds (Asner 2001). We controlled for cloud interference in NDVI data by including only the maximum monthly values from both years. Previous studies using NDVI data have been performed at high elevations in the Andes and suggest that a signal can be captured out of the data (Fuentes et al. 2001). Distribution models may perform poorly if we chose environmental predictor variables that are not biologically meaningful for the study subject. NDVI and climate data are two of the most commonly used predictors of species distributions at a regional scale because of their presumed importance as limiting factors for species distributions (Guisan and Zimmermann 2000, Peterson et al. 2002, Hurlbert and Haskell 2003). ECOGRAPHY 27:3 (2004)

37020 20689 29758 19075 15063 10450 9608 44854 25492 40110 23028 19166 13186 11975 38059 22485 36109 18112 19648 16925 13413 49010 30064 56353 24034 26057 22940 18219 44816 24276 43608 21708 25631 18278 16433 64000 35374 55998 32106 35585 24517 22407 30001 20322 31298 19172 22339 17151 16128 41120 29029 56436 25819 29097 22991 20865 11774 5316 11616 4529 7716 4078 3492 51250 15199 61610 12966 33907 11384 9992 12848 4268 7832 4199 5937 2432 2387 34064 5094 39494 5014 14331 2921 2869 Elevation Climate NDVI Elev/Cli Elev/NDVI Cli/NDVI Elev/Cli/NDVI

Cut Total Total Cut Total

Cut

Total

Cut

Total

Cut

Total

Cut

G. squamigera G. rufula G. ruficapilla G. quitensis G. nuchalis G. hypoleuca

Table 2. Predicted areas of occurrence based on different data sets. Predicted areas within each species’ range map (under column labeled ‘‘cut’’) and total predicted areas within Ecuador (under column labeled ‘‘total’’). Units are squared meters.

ECOGRAPHY 27:3 (2004)

Antpittas are a distinctive group of new world forest dwelling birds that are shy and are not believed to fly long distances (Ridgely and Tudor 1994). These traits combined with historical fragmentation cycles of cloud forests due to pleistocene glaciations can lead to high, localized adaptation (Holt 2003). Grallaria distributions are often restricted to narrow altitudinal ranges along the Andes (Ridgely and Tudor 1994) delineated by narrow environmental variables. These variables should be reflected in the environmental data sets we employed. In the case of NDVI-only based models, we expected performance to improve when we excluded points in currently deforested areas. Nonetheless, there were no apparent changes in terms of Kappa values. The fact that we had a small sample size (40
/140 point localities), makes the exclusion of just a few points out of the model a compromise that could lead to decreased model performance. A final factor, which could contribute to a general low performance is the modeling method employed. Although BIOCLIM is a simple straightforward method
/ one of its main strengths
/ it has some flaws (i.e. giving an equal weight to all variables, and sensitivity to outliers, Farber and Kadmon 2003). However, there is no consensus in the literature as to what method is ‘‘best’’ (Guisan and Zimmermann 2000) and initial testing between BIOCLIM, GARP, and DOMAIN methods indicated that results from each modeling type were similar (results not shown here). Our objective in this research was not to test among modeling methods but to test among environmental datasets. Modeling species geographic distributions is a useful exercise to better understand potential limiting factors or correlates of species’ distributions (O’Connor 2002). Nonetheless, modeling procedures can be misleading if the performance of the models and the types of environmental variables used are not evaluated. The results presented here show that performance has many faces, each representing a cost depending on the purpose of the model (Fielding and Bell 1997, Guisan and Zimmermann 2000). Lack of thorough spatial sampling and the need for effective conservation networks are the usual scenario for many tropical humid forest dwelling birds. Niche modeling can be a useful aid to their conservation (Peterson et al. 2002). As detailed natural history information on these species accumulates, we will be able to relate landscape variables (such as mean patch size or proximity of patches) to species-specific information on tolerance of habitat alteration or minimum patch size requirements to refine and improve our models. Acknowledgements
/ We would like to thank N. Krabbe and R. Ridgely for his collaboration with the maps. Robert Hijmans and Wayne Heiser provided technical advice. Craig Moritz provided funding for JLP work in the MVZ. Robert Hijmans and Craig Moritz provided useful advice on earlier drafts. JFF would like to thank Programa de Becas de Investigacio´n para la Conservacio´n, Fundacio´n EcoCiencia, and William Belton

357

Donations Program, American Bird Conservancy, for funding research on the distribution and conservation of Grallaria species. The Museum of Vertebrate Zoology (MVZ), NCEAS working group and an NSF grant provided funding for all modeling analyses. All computer software used was provided through the MVZ.

References Anderson, R. P., Gomez-Laverde, M. and Peterson, A. T. 2002. Geographical distributions of spiny pocket mice in South America: insights from predictive models.
/ Glob. Ecol. Biogeogr. 11: 131
/141. Anderson, R. P., Lew, D. and Peterson, A. T. 2003. Evaluating predictive models of species distributions: criteria for selecting optimal models.
/ Ecol. Modell. 162: 211
/232. Asrar, G. et al. 1984. Estimating absorbed photosynthetic radiation and leaf area index from spectral reflectance in wheat.
/ Agr. J. 76: 300
/306. Asner, G. P. 2001. Cloud cover observation of the Brazilian Amazon.
/ Int. J. Remote Sens. 22: 3855
/3862. Berry, P. M. et al. 2002. Modelling potential impacts of climate change on the bioclimatic envelope of species in Britain and Ireland.
/ Glob. Ecol. Biogeogr. 11: 453
/462. Busby, J. R. 1991. BIOCLIM
/ A bioclimate analysis and prediction system.
/ In: Margules, C. R. and Austin, M. P. (eds), Nature conservation: cost effective biological surveys and data analysis. CSIRO, pp. 64
/68. Carpenter, G., Gillison, A. N. and Winter, J. 1993. DOMAIN: a flexible modeling procedure for mapping potential distributions of plants and animals.
/ Biodiv. Conserv. 2: 667
/680. Chong, D. L. S., Mougin, E. and Gastellu-Etchegorry, J. P. 1993. Relating the global vegetation index to net primary productivity and actual evapotranspiration over Africa.
/ Int. J. Remote Sens. 14: 1517
/1546. Faith, D. P. et al. 2001. The Biorap biodiversity assessment and planning study for Papua New Guinea.
/ Pacific Conserv. Biol. 4: 279
/288. Farber, O. and Kadmon, R. 2003. Assessment of alternative approaches for bioclimatic modeling with special emphasis on the Mahalanobis distance.
/ Ecol. Modell. 160: 115
/ 130. Fielding, A. H. 2002. What are the appropiate characteristics of an accuracy measure?
/ In: Scott, J. M. et al. (eds), Predicting species occurrences: issues of accuracy and scale. Island Press, pp. 271
/279. Fielding, A. H. and Bell, J. F. 1997. A review for the assessment of prediction errors in conservation presence/absence models.
/ Environ. Conserv. 24: 38
/49. Fjeldsa˚, J. et al. 1997. Are biodiversity ‘hotspots’ correlated with current ecoclimatic stability? A pilot study using the NOAAAVHRR remote sensing data.
/ Biodiv. Conserv. 6: 401
/ 422. Freile, J. F. 2000. Patrones de distribucio´n y sus implicaciones en la conservacio´n de los ge´neros Grallaria y Grallaricula (Aves: Formicariidae) en Ecuador. PhD thesis, Dept of Biology, Pontificia Univ. Cato´lica del Ecuador, Quito, Ecuador. Fuentes, M. V., Malone, J. B. and Mas-coma, S. 2001. Validation of a mapping and prediction model for human fasciolosis transmission in Andean very high altitude endemic areas using remote sensing data.
/ Acta Tropica 79: 87
/95. Godown, M. E. and Peterson, A. T. 2000. Preliminary distributional analysis of US endangered bird species.
/ Biodiv. Conserv. 9: 1313
/1322. Green, R. H. 1979. Sampling design and statistical methods for environmental biologists.
/ Wiley.

358

Guisan, A. and Zimmermann, N. E. 2000. Predictive habitat distribution models in ecology.
/ Ecol. Modell. 135: 147
/ 186. Henricksen, B. L. and Durkin, J. W. 1986. Growing period and drought early warming in Africa using satellite data.
/ Int. J. Remote Sens. 7: 1583
/1608. Holben, B. N. 1986. Characteristics of maximum-value composite images from temporal AVHRR data.
/ Int. J. Remote Sens. 7: 1417
/1434. Holt, R. T. 2003. On the evolutionary ecology of species’ ranges.
/ Evol. Ecol. Res. 5: 159
/178. Houlder, D. et al. 2000. ANUCLIM user’s guide. Centre for Resource and Environmental Studies (CRES).
/ Australian National Univ., Canberra, Australia. Huete, A. R. 1989. Soil influences in remotely sensed vegetationcanopy spectra.
/ In: Asrar, G. (ed.), Theory and applications of optical remote sensing. Wiley, pp. 107
/141. Hugall, A. et al. 2002. Reconciling paleodistribution models and comparative phylogeography in the Wet Tropics rainforest land snail Gnarosophia bellendenkerensis (Brazier 1875).
/ Proc. Natl. Acad. Sci. USA 99: 6112
/6117. Hurlbert, A. H. and Haskell, J. P. 2003. The effect of energy and seasonality on avian species richness and community composition.
/ Am. Nat. 161: 83
/97. Hutchinson, M. F. 1991. Climatic analyses in data sparse regions.
/ In: Muchow, R. C. and Bellamy, J. A. (eds), Climatic risk in crop production. CAB, pp. 55
/71. Hutchinson, M. F. 1995. Interpolating mean rainfall using thin plate smoothing splines.
/ Int. J. GIS 9: 305
/403. Hutchinson, M. F. 1998. Interpolation of rainfall data with thin plate smoothing splines: II analysis of topographic dependence.
/ J. Geogr. Inf. Dec. Anal. 2: 168
/185. Hutchinson, M. F. 2000. Anusplin, ver. 4.1. User guide.
/ Centre for Resource and Environmental Studies, The Australian National Univ., Canberra, Australia. Ichii, K., Kawabata, A. and Yamaguchi, Y. 2002. Global correlation analysis for NDVI and climatic variables and NDVI trends: 1982
/1990.
/ Int. J. Remote Sens. 23: 3873
/ 3878. Joseph, L. and Stockwell, D. 2002. Climatic modeling of the distribution of some Pyrrhura parakeets of Northwestern South America with notes on their systematics and special reference to Pyrrhura caeruleiceps Todd, 1947.
/ Ornitol. Neotrop. 13: 1
/8. Krabbe, N. et al. 1998. An Atlas of avian diversity in the Ecuadorian Andes.
/ Centre for Research on Cultural and Biological Diversity of Andean Rainforests (DIVA), DIVA Tech. Rep. no. 4. Quito, Ecuador. Landis, J. R. and Koch, G. C. 1977. The measurement of observer agreement for categorical data.
/ Biometrics 33: 159
/174. Lillesand, T. M. and Kiefer, R. W. 2000. Remote sensing and image interpretation, 4th edn.
/ Wiley. Loiselle, B. A. et al. 2003. Avoiding pitfalls of using species distribution models in conservation planning.
/ Conserv. Biol. 17: 1591
/1600. Manel, S., Williams, H. C. and Ormerod, S. J. 2001. Evaluating presence-absence models in ecology: the need to account for prevalence.
/ J. Appl. Ecol. 38: 921
/931. New, M., Hulme, M. and Jones, P. 1999. Representing twentieth-century space-time climate variability. Part I: development of a 961-90 mean monthly terrestrial climatology.
/ J. Climate 12: 829
/856. Nix, H. A. 1986. A biogeographic analysis of Australian elapid snakes.
/ In: Anon. (ed.), Atlas of Australian elapid snakes. Bureau Flora Fauna, Canberra, Australia, pp. 4
/15. O’Connor, R. J. 2002. The conceptual basis of species distribution modeling: time for a paradigm shift?
/ In: Scott, J. M. et al. (eds), Predicting species occurences: issues of accuracy and scale. Island Press, pp. 25
/33. Oindo, B. O. 2002. Predicting mammal species richness and abundance using multitemporal NDVI.
/ Photo. Engin. Rem. Sens. 68: 623
/629. ECOGRAPHY 27:3 (2004)

Peterson, A. T. 2001. Predicting species geographic distributions based on ecological niche modeling.
/ Condor 103: 599
/ 605. Peterson, A. T., Stockwell, D. R. B. and Kluza, D. A. 2002. Distributional prediction based on ecological niche modeling of primary occurrence data.
/ In: Scott, J. M. et al. (eds), Predicting species occurences: issues of accuracy and scale. Island Press, pp. 617
/623. Peterson, A. T. and Robins, C. R. 2003. Using ecological niche modeling to predict barred owl invasions with implications for spotted owl conservation.
/ Conserv. Biol. 17: 1161
/ 1165. Renjifo, L. M. et al. 1997. Patterns of species composition and endemism in the northern Neotropics: a case for conservation of montane avifaunas.
/ Ornit. Monographs 48: 577
/ 594. Richard, Y. and Poccard, I. 1998. A statistical study of NDVI sensitivity to seasonal and interannual variations in Southern Africa.
/ Int. J. Remote Sens. 19: 2907
/2920.

ECOGRAPHY 27:3 (2004)

Ridgely, R. S. and Tudor, G. 1994. The birds of South America. vol. I: the oscine passerines.
/ Univ. of Texas Press. Sa´nchez-Cordero, V. and Martı´nez-Meyer, E. 2000. Museum specimen data predict crop damage by tropical rodents.
/ Proc. Natl. Acad. USA 97: 7074
/7077. Stockwell, D. R. B. and Noble, I. R. 1991. Induction of sets of rules from animal distribution data: a robust and informative method of data analysis.
/ Math. Comp. Sim. 32: 249
/ 254. Sierra, R. 1999. Propuesta preliminar de un sistema de clasificacio´n de vegetacio´n para el Ecuador continental.
/ Proyecto Inefan/Gef-Birf y EcoCiencia. Quito, Ecuador. Sierra, R., Campos, F. and Chamberlin, J. 2002. Assessing biodiversity conservation priorities: ecosystem risk and representativeness in continental Ecuador.
/ Landsc. Urban Plann. 59: 95
/110. Zar, J. H. 1996. Biostatistical analysis, 3rd ed.
/ Prentice Hall. Zinner, D., Pela´ez, F. and Tokler, F. 2002. Distribution and habitat of grivet monkeys (Cercopithecus aethiops aethiops ) in eastern and central Eritrea.
/ Afr. J. Ecol. 40: 151
/158.

359

360

ECOGRAPHY 27:3 (2004)

Predicted

Elev/Cli/NDV (7)

Cli/NDV (6)

Elev/NDV (5)

Elev/Clim (4)

NDVI (3)

Climate (2)

0

Elevation (1)

1

0

1

0

1

0

1

0

1

0

1

0

1

Cut

Data sets

Species

/ / / / / / / / / / / / / / / / / / / / / / / / / / / /

9408 751 7250 631 3616 6543 2987 4894 5302 4857 3930 3951 3600 6559 2971 4910 4883 5276 3653 4228 2043 8116 1673 6208 2031 8128 1661 6220

/ 24656 47167 5598 3004 1478 70345 1281 7321 34192 37631 3902 4700 1414 70409 1228 7374 9448 62375 2284 6318 878 70945 759 7843 838 70985 726 7876

/

G. hypoleuca

11293 318 5783 224 6637 4974 3053 2954 8834 2777 4646 1361 6498 5113 2976 3031 8371 3240 4325 1682 4983 6628 2360 3647 4880 6731 2296 3711

/

G. nuchalis

39957 30414 5991 4090 8562 61809 2263 7818 52776 17595 6970 3111 6468 63903 1553 8528 25536 44835 3391 6690 6401 63970 1718 8363 5112 65259 1196 8885

/

Observed

18545 738 14662 620 14553 4730 11676 3606 16125 3158 13147 2135 14257 5026 11397 3885 14722 4561 11950 3332 12430 6853 10243 5039 12151 7132 9978 5304

/

G. quitensis

22575 40124 15339 8934 14476 48223 8646 15627 40311 22388 18151 6122 11562 51137 7775 16498 14375 48324 10389 13884 10561 52138 6908 17365 8714 53985 6150 18123

/ 24059 124 18875 46 17268 6915 12816 6105 16742 7441 13696 5225 16936 7247 12556 6365 15869 8314 12969 5952 12355 11828 9735 9186 12164 12019 9554 9367

/ 39941 17858 25941 14550 18106 39693 11460 29031 39256 18543 29912 10579 15170 42629 9152 31339 19716 38083 12662 27829 12162 45637 8543 31948 10243 47556 6879 33612

/

G. ruficapilla

15184 967 12949 656 10102 6049 8270 5335 11560 4591 9657 3948 8938 7213 7412 6193 9170 6981 7728 5877 7369 8782 6054 7551 6559 9592 5434 8171

/

G. rufula

33826 32005 25110 13901 19962 45869 14215 24796 44793 21038 26452 12559 15096 50735 10700 28311 16887 48944 11920 27091 15571 50260 10871 28140 11660 54171 7979 31032

/

5015 225 4018 156 3733 1507 3025 1149 3167 2073 2424 1750 3619 1621 2926 1248 2839 2401 2215 1959 2278 2962 1810 2364 2207 3033 1749 2425

/

39839 36903 33002 23537 21759 54983 17664 38875 36943 39799 27334 29205 19409 57333 16149 40390 16327 60415 12848 43691 10908 65834 8640 47899 9768 66974 7859 48680

/

G. squamigera

Appendix. Confusion matrices of models based on each dataset used. Numbers correspond to counts of square kilometer cells. Confusion matrices for predictions within each species’ range map are coded under 1 in the column ‘‘cut’’; matrices for total predictions within Ecuador are coded as 0. Positive ‘‘/’’ sign refers to presence and negative ‘‘/’’ to absence.