The goal of our pilot project was to demonstrate the feasibility and utility of the gap analysis methodology for predicting biodiversity distribution at the watershed scale. We illustrated this through the creation of automated and calibrated geographic information system models for prediction of diversity at the community level. The use of a landcover map in the prediction of water quality served as a possible connection to the gap analysis efforts on terrestrial ecosystems. The utility of the model in conservation planning was demonstrated through the identification of gaps in the protection of waters with relatively rich biological diversity. Examination of the pattern of gaps helped to identify options for enhanced conservation of aquatic biodiversity. Finally, we promoted the aquatic gap analysis methodology through presentations, documentation on a world wide web site, and this final report.

Prediction accuracy from the automated model used on the Allegheny River watershed varied among the five landscape attributes. The automated model succeeded in using cumulative drainage area to characterize bankfull width with significant accuracy. Further, field observations of baseflow depth were well correlated with predicted bankfull depth values obtained from cumulative drainage area relations. This lends further support that drainage area at any point is closely correlated with many size characteristics (Dunne and Leopold 1978) and can be used as a general measure of stream size. In hydrologically homogenous regions, channel dimensions are so consistent with drainage area that deviations can be interpreted as the magnitude of the effect of urbanization (Dunne and Leopold 1978).

Habitat quality and riparian forest cover predictions were significantly different from observed data. However, the accuracy of automated predictions for these two attributes was compromised by possible misclassification of landuse during interpretation of the satellite data from which the digital maps were derived (Cherrill et al. 1995, Tucker et al. 1997). These limitations are inherent in the use of satellite data for ecological modeling. Further, habitat quality and riparian forest cover characterizations were constant for an entire stream segment from confluence to confluence while the observed data were subjective and often only reflective of a point on the stream most easily accessible. Thus, habitat quality and riparian forest cover characterization is difficult to test and may not be predictive of actual stream conditions at all points along the stream segment.

Stream gradient acts as a surrogate for substrate by separating organisms which favor sand, silt and clay (low gradient streams) from those which favor cobble, boulders and rock (high gradient streams). Observed median gradients plotted against dominant substrate in the Allegheny River watershed leant support for the placement of the classification criterion used to separate sites with dominant fine sediment substrate from those with dominant coarse substrate (gravel, pebble and cobble). Thus, the classification criteria used here is successful at separating sites based on substrate composition in the Allegheny River watershed. Automated gradient predictions from 1:250,000 DEM digital data were very well correlated with observed data, indicating that automated procedures could be confidently used to generate significantly accurate gradient values.

The nonpoint source pollution load screening model (Adamus and Bergman 1995) was formerly used to predict total nitrogen, total phosphorous, suspended solids and biological oxygen demand in Florida. In New York, the model proved unable to accurately predict pollution loads for all parameters but total nitrogen. The model was chosen for its coarse, simple GIS approach in the interest of developing a basic water quality characterization from easily available data. For this reason, the model does not incorporate several important factors impacting surface water quality such as point-source pollution, regional geology, elevation, watershed size, shape, and topography (Soranno et al 1996). Misclassifications errors in the creation of the digital data (Cherrill et al. 1995, Tucker et al. 1997), errors in the sampling of the water, incorrect regional runoff coefficients and pollutant concentration values, or incorrect classification criteria for water quality categorization may have further contributed to the inaccuracy of the model. Since water quality is an influential factor in both fish and macroinvertebrate diversity predictions, inaccuracies in water quality prediction may contribute to inaccuracies in both the fish and macroinvertebrate diversity predictions.

The distribution and amount of predicted modified and highly altered habitat was considerable and widespread throughout the three watersheds while far fewer stream segments existed with predicted degraded water quality. Degradations in water quality had a larger effect on predicted fish species diversity than alterations in habitat quality, however, due to its prevalence, degraded habitat quality was the more dominant factor limiting the number of habitats with high predicted fish diversity. Differences in stream size had the smallest effect on predicted fish diversity, however, higher fish diversity was predicted in large streams and small rivers which were very scarce. The relative scarcity of stream segments with intact habitat quality was a consequence of agricultural farming concentrated in the stream valleys and from density of primary and secondary roads. This scarcity combined with the shortage of large streams and small rivers caused few stream segments to have the necessary qualifications for high predicted fish diversity in the automated model.

Macroinvertebrates are very sensitive to changes in water quality. As with fish diversity prediction, stream segments with good water quality were predicted to support a higher diversity of species. Stream segments with predicted degraded water quality were scarce in all watersheds and the effect of gradient and riparian forest cover on diversity was minimal. This accounts for the abundance of stream segments predicted to be high in macroinvertebrate diversity in all watersheds.

Sites with superior water quality and habitat quality were predicted to provide suitable habitat for both tolerant and intolerant fish species forcing the diversity predictions from the automated model for these habitat types to be large in value (ex. 62-100 species). Field observations of this magnitude were less likely though gradations in diversity would be represented in the observations. Thus, predicted fish diversity from the automated unranked and ranked predictions (with seining sites removed from the analysis) were significantly correlated but not in agreement with observed fish diversity in the Allegheny River watershed. Predictions for Fall and French Creek watersheds were less optimistic with only ranked predictions in Fall Creek significantly correlated with observed data. Fish collections in the Allegheny River watershed were designed for use specifically to test automated diversity predictions from this project and were aimed at providing good coverage in all habitat types and consistent field collection methods. Collection data from the Fall Creek and French Creek watersheds were obtained from unrelated studies with datasets often less stratified among habitat types and with varying field methods, thus possibly lowering the success of these two datasets at assessing the accuracy of the automated model. Further, fish diversity is difficult to predict through habitat-species relations since fish are flexible and opportunistic with an ability to change habitat use with need (Bain in press). Fish diversity is also difficult to measure since collection methods are more selective for certain species types and the mobility of fish may easily take them beyond the range of equipment. Further, in this project, select areas were used as representative samples of the entire stream segment, some of which were several miles long. Therefore, the observed data may not accurately characterize the diversity of the entire stream segment being testing. Despite these drawbacks, the automated model was successful at providing some evidence for correlation between predicted and measured fish diversity.

The calibrated model was somewhat effective at predicting fish diversity in the Allegheny River watershed with seining sites included in the observed data. The decrease in prediction accuracy when seining sites were removed from the analysis may be explained by the fact that the original set of data used in discriminant analysis was a diverse compilation of data from varying observers, using different methods, covering a period of many years, and itself may have included seining sites. Fish diversity predictions in French Creek were quite dissimilar from observed data. The French Creek predictions were performed using a set of linear equations developed from data in the Allegheny River watershed so this result may prove that predictions made on the 1:100,000-scale may not be applicable to stream segments at the 1:24,000-scale. Unfortunately, the applicability of the Allegheny River watershed linear equations in other watersheds of equal scale could not be tested since predicted data in test sites of the Fall Creek watershed all resided in a single category.

Despite the fact that the calibrated model predicted many more sites with high fish diversity than the automated model in the Allegheny River watershed, many of the sites were in the same location and the percentages of protected kilometers were almost equal. The percentages of gaps (47% and 48%, automated and calibrated models) indicate somewhat effective conservation of high fish diversity stream segments under current methods, however, more directed efforts would be helpful. Protection levels were quite high in the Fall Creek watershed (94%) using high fish diversity predictions obtained from the automated model. This was likely due to the large amount of the Fall Creek watershed under the protection of the regulated freshwater wetlands. The scarcity of gaps in the Fall Creek watershed was an indication that the type and location of protection methods are effective at conserving high fish diversity stream segments. No high fish diversity stream segments in French Creek, predicted using the automated model, were under any form of protection. Almost all of the protected stream segments in the French Creek watershed are tributaries to the mainstem. The high fish diversity segments were predicted to be small rivers and large streams which lay exclusively on the mainstem and thus, fall outside of the region of protection. Predictions from the calibrated model in the French Creek watershed were more dispersed, therefore, a larger percentage (44%) of the stream segments were under protection. The gaps in the French Creek watershed (56%) were again indicative of somewhat effective conservation of high fish diversity stream segments, but efforts could be made to further protect areas of unusually high fish diversity.

The automated and calibrated models provided predictions of fish diversity which were similar in location and of good accuracy when used on larger areas such as the Allegheny River watershed. This was evidenced by the significant correlation in both models between predicted and observed fish diversity in the Allegheny River watershed at the 1:100,000-scale and uniformly lower correlations in all tests on the smaller French Creek watershed at the 1:24,000-scale. Overall, fish diversity predictions from the calibrated model were less selective since only three categories of prediction were used, as opposed to the 18 used in the automated model, however, accuracy levels were almost equivalent.

Observed macroinvertebrate diversity values were well correlated with predictions from the automated model for the Allegheny River and French Creek watersheds. As in the fish diversity models, macroinvertebrate predictions for high diversity habitat types were likely to overpredict to account for presence of both tolerant and intolerant families. Thus we would expect the predictions to be well correlated with but not significantly equal to observed macroinvertebrate diversity as evidenced in French Creek. Again, macroinvertebrate predictions at the 1:100,000-scale were uniformly closer to observed diversity than those at the 1:24,000-scale. The percentage of gaps in macroinvertebrate diversity protection in the three watersheds (58% Allegheny River, 56% Fall Creek, and 73% French Creek) indicated that the level of protection currently in place in these watersheds was somewhat effective at conserving macroinvertebrate diversity but more directed efforts would be helpful.

Predictions of fish species occurrences using the calibrated model were quite poor except for rainbow darter. The low levels of accuracy were not surprising given the opportunistic tendencies for fish species in their use of habitat. Further, four of the seven fish species tested (brook trout, brown trout, smallmouth bass, and muskellunge) were stocked and, therefore, may be found in many habitats not normally supporting such fish species. This complicates the discriminant analysis process of fitting linear equations to the data and may alter the success of the accuracy analysis results. The time of year, amount of water, availability of food, presence of predators, and collection method all contribute to the recorded presence or absence of a single species and are difficult to factor into the calibration model predictions or accuracy analysis of the results. New methods should be reviewed for enhancement of prediction at the species level.

Effective conservation of biodiversity in aquatic communities requires the identification and protection of key landscapes and communities (Angermeier and Winston 1999). To do so, ecologists and landscape planners need to design protocols to assess community diversity and develop strategies for conservation. Geographic information systems can assist in modeling large areas and identifying regions where conservation efforts may be focused. It is clear that the modeling procedures presented here have considerable potential as coarse but feasible methods in predicting the diversity of fish and macroinvertebrate communities on the watershed scale.