Conclusions and Recommendations for Changes in Approach
Recommendations for Follow-on Research

Regional vegetation mapping
Wildlife habitat modeling
Monitoring environmental change
Regional conservation planning and reserve design

Progress in conservation assessment and planning has been severely and unnecessarily limited by hardware and software for mapping and spatial analysis. Specifically: 1) biogeographers and conservation biologists have not had adequate computing resources to analyze the large volumes of data involved in conservation assessments; 2) data management systems in general use are poorly designed for manipulation of heterogeneous biogeographic data; 3) there is practically no coupling among database management systems and analytical software used in biodiversity analyses; 4) it is difficult to visualize biogeographical data sets and model outputs with existing display tools; and 5) spatial modeling and decision support are constrained by inadequate hardware and cumbersome protocols for conducting sensitivity and error propagation analyses.

This research project was to design and enable a prototype regional computing facility for storage, analysis, and visualization of biodiversity data. More specifically, applications were to be developed to support regional gap analysis and siting of nature reserve systems. As originally conceived, the project was intended to develop an object-oriented database with a suite of operators to perform the most important, standard data transformations needed in conservation work and integrated through a common user interface. As the research progressed, however, the technology advanced so rapidly that it made more sense to adapt these technologies to our needs rather than to develop new, parallel solutions. For instance, the advent of the World Wide Web provided a universal environment for data cataloging. Commercial GIS software began to include tools for customizing user interfaces.

Also, the concept of a single user interface to support the gamut of conservation analysis tasks from image processing of satellite image data through location-allocation modeling to site new reserves became unwieldy. Each regional assessment would have unique data and processing requirements, making a prototype system impractical. Instead we focused our effort on advancing some of the individual research problems by improving the approaches, particularly those where processing vast quantities of data were involved. These tasks generally required both scientific and computational advances in the approaches currently used in conservation studies.

Regional vegetation mapping

Most tasks in regional conservation assessment and planning require comprehensive and consistent land-cover maps at a taxonomic level detailed enough to reflect variation in the native biota. For large ecoregions, this mapping typically involves massive data sets of satellite imagery, which are difficult to acquire under similar viewing conditions. We have addressed this issue by developing and testing a new method of compositing daily images into composites covering 10-14 day periods with improved viewing angle and elimination of most cloud cover (Stoms et al. 1997). This compositing strategy was used to compile AVHRR data for a growing season over the Intermountain Semi-Desert Ecoregion covering parts of nine western states. With further testing, this approach could improve the quality of composites being generated to support an international global land-cover mapping and global change research efforts.

A new image classification technique developed for this IBM project was used to classify this multi-temporal image data set into vegetation alliances, using the state GAP maps as training information for the maximum likelihood classifier (Stoms et al. in review). The major innovation of this map-guided classification technique is that it is iterative, assigned pixels to map classes only when there is strong agreement at the current iteration between spectral clusters and map information classes. This technique was applied to the AVHRR data set for the Intermountain Semi-Desert Ecoregion to compile a spatially and taxonomically consistent land-cover map, where the individual state maps had created abrupt discontinuities at boundaries between states. The nation's first multi-state regional gap analysis was conducted using this land-cover map as an expanded coarse-filter for assessing the status of biodiversity in the region (Stoms et al. in review). Map-guided classification is also being used to monitor changes in land-cover over time and could be useful in any large area mapping project as a means of integrating data from different sources.

Much useful information can be derived from the California GAP database with its rich set of landscape attributes. It achieves a view of vegetation over large, heterogeneous regions while containing considerable floristic information and spatial detail. This view was only possible by integrating many kinds of spatial data ranging from modern satellite imagery to air photos and archival vegetation maps. It is intermediate in detail between traditional regional biogeography and local ecological studies, and helps to bridge those very different perspectives. An important distinction between the map-based study in coastal sage scrub (Davis et al. 1994) and earlier studies is that the GAP database is spatially exhaustive across the range of the community type and therefore better suited to regional planning and policy analyses than strictly plot-based information. By relating the information to other spatial data we can readily answer queries such as: Which coastal scrub types occur on national forest lands? Which lands dominated by Salvia leucophylla are zoned as open space? Where are large areas of coastal scrub vegetation that are likely candidates for new reserves?

Wildlife habitat modeling

The orange-throated whiptail study (Hollander et al. 1994) illustrates how different distribution and environmental data at various scales can generate predictive distribution maps and hypotheses about the factors controlling them. None of these representations can be considered definitive, but each has its uses. The advantages of each map approach become more apparent when these different representations are considered together. Thus we envision a mapping environment where the researcher struggles no longer to produce a single map, but produces suites of them at will. Data integration is one component to this, but so are the flexibility and clarity of the underlying models, the multiple images thereby creating a better representation of the complex reality underlying diverse data sources.

From the point of view of wildlife habitat modeling in general, there are a number of results to be highlighted from the wild pig study. The first element to the wildlife habitat modeling is incorporation of human disturbance as a factor affecting relative abundance levels. This is done multiplicatively, with higher levels of hunting pressure or greater road densities corresponding to lower relative abundances. Also, additional influencing factors can be incorporated in the network model by adding nodes and links to the diagram. Finally, both the local and regional models are spatialized in that they integrate habitat factors across the landscape rather than focusing on a single point. This is a step towards building dynamic models of wildlife populations through space and time.

With respect to GIS methodology, the wild pig habitat modeling project has illustrated how expert review of the component layers in a GIS model can enhance the modeling process. Our methodology illustrates how interactive review of the components of the GIS model can aid in its development. This has been accomplished in a workshop where the components of GIS models were presented to wildlife experts. The portability of the presentation has been facilitated by technology such as the use of IBM laptop computers, overhead display devices, and current GIS visualization software. Nevertheless, the present technology did not allow these models to be altered interactively during the workshop. Another issue concerning expert review is whether to elicit feedback at a workshop, as done here, or through interviews with experts carried out individually. The latter offers the possibility of reaching the opinion of more people, whereas group review allows for more synergism and consensus from participants. Another component of expert review is the ability to embed expert knowledge into a formal system for GIS modeling. This can be carried out at several levels. The first is simply to translate the GIS model into a script written in the macro language of the GIS system. This allows the modeling process to be replicated using new datasets. Another level is to create a graphical representation of the formal model for ease in communication. This has been done here in the network diagrams for both the local and regional models. Moreover, this network diagram structure is closely related to influence diagram models from decision analysis. Algorithms have been developed to evaluate such diagrams, which means that these that diagrams can constitute a formal expert system.

Monitoring environmental change

Environmental monitoring and change detection is a very active area of geographic research. Many technical challenges (for example, scene-to-scene radiometric and geometric rectification, scale-dependence, classification strategies, accuracy assessment) and conceptual issues (e.g., criteria for defining and recognizing environmental change) remain. Although monitoring was not a central focus of our IBM-ERP research, we did make substantive progress in the areas of AVHRR pre-processing, compositing methods for obtaining cloud-free imagery (Stoms et al. 1997), and map-guided classification for detecting change using satellite remote sensing.

Regional conservation planning and reserve design

Major advances were made in the area of regional conservation planning. In particular, our linking of the emerging science of conservation biology with the traditions of operations research led to very fruitful research directions. The reserve selection problem in the past was usually solved with simplistic greedy-adding heuristics. Rare attempts to find optimal solutions were thwarted by the apparent combinatorial dilemma for any problem of the dimensions of real-world planning situations. By bringing the problem solving insights of linear programming, we were able to formulate it as a maximal covering location problem and to find optimal solutions to the basic reserve selection problem in reasonable computing time to be useful to conservation planners (Church et al. 1996). By reformulating the MCLP as a p-median problem, we also succeeded in integrating the model into a commercial GIS which most planners can use (Gerrard et al. in press).

While other enhancements to the basic covering model were being developed (Church et al. in review, Stoms et al. in review), we took a different approach to selecting biodiversity management areas with a model that could meet more sophisticated conservation objectives. The BMAS model can select a set of sites that maximizes dual objectives, both efficiency and suitability, while meeting areal representation targets. The model is useful for exploring the implications of planning objectives, assumptions, and policy choices. Not only was the model useful in a research environment for the Sierra Nevada Ecosystem Project (Davis et al. 1996), it was then successfully adapted to an actual planning project with The Nature Conservancy in the Columbia Plateau ecoregion in the Pacific Northwest.

Regional vegetation mapping
Wildlife habitat modeling
Monitoring environmental change
Regional conservation planning and reserve design

Regional vegetation mapping

Image compositing

The comparison of alternative compositing algorithms was limited to one 14-day period in a single study area and thus makes generalization difficult. California's geographic position at the extreme western limit of range from the ground receiving station in South Dakota precludes the availability of some eastward viewing opportunities. Located in the northern hemisphere means that California is imaged close to the principal plane of the sun, which accentuates anisotropic effects of both atmosphere and surface. Thus, the study is not fully representative of all regions of the country or the world. Consequently, the set of weights we selected for the MOC algorithm tested here may well not be appropriate for general use. Nevertheless, the multiple objective approach of weighting greenness, temperature, and viewing angle in compositing can be useful in discovering their relative importance over a range of circumstances. Some problems remain in the use of the selected weights in our study for specific land cover situations, such as snow cover and wildfires, which limits the utility of the algorithm for global applications. Additional study is needed to account for these unusual circumstances. The Land Cover Working Group of the International Geosphere-Biosphere Programme Data and Information System is developing global land cover datasets based on the traditional algorithm. We recommend that alternative methods such as the multiple objective compositing strategy be tested further in other regions and seasons.

Map-guided classification

The map-guided classification method provided an innovative means of compiling a regional land-cover map, but additional improvements may be possible. The proposed national schema for classifying land-cover is hierarchical, beginning with structural or physiognomic features at the highest levels. This suggests that a hierarchical approach to mapping might be appropriate as a two-stage classifier. Logic rules could be used for structural classification, similar to those proposed by Running et al. (1995). If the source map labels are consistent with the inference of logic rules applied to AVHRR data at the formation level, the source map alliance label would be assigned. If not, the map-guided classification or further ecological rules would be invoked. It should be noted that the classification logic in Running et al. (1995) uses AVHRR-derived variables to determine leaf type and phenology and permanence of aboveground live biomass, but does not distinguish life form or canopy density. Therefore that particular hierarchical approach would not directly lead to a classification schema compatible with the NVCS. Alternatively, a map-guided classification could be used as part of an evidential reasoning approach which would integrate spectral data and derived indices or biophysical metrics with the source map labels and environmental variables. Such a two-step approach may lead to higher accuracy but would require greater effort to develop the set of deductive rules.

Vegetation classification

Federal agencies and other resource groups are developing a national standard classification system. Our vegetation database approach used in the California GAP provided a flexible data source for developing more detailed classifications (Davis et al. 1994). A great deal more work needs to be done to mine this database for identifying new floristic alliances that have been neglected in the past.

Wildlife habitat modeling

The set of representations for the orange-throated whiptail suggests several additional lines of investigation. One is to further sample in poorly-studied areas which become obvious in the multiple representations. Further sampling would allow for the iterative refinement of the habitat models. To refer back to the hypercube, this three-tiered scheme illustrates the importance of casting the diversity of data into a suitable structure. Within this structure, as the whiptail example has shown, different datasets are played off of one another, looking for common patterns and making inferences. Effectively this process constitutes a sensitivity analysis on the predicted distribution for the species. Unfortunately, existing commercial GISs do not facilitate this sort of interactive work. The heterogeneity of these datasets - composed as they are of vector maps, images, tabular and statistical models, and so on - means much effort must be put into converting data from one form to another. GISs are weak in enabling spatial statistical analyses, and systems that assist in searching for geographical patterns are only at the research stage. But conversely, improving ability to visualize and integrate complex datasets is an active area of development and research in GIS. A promising approach for such integration has come from declarative or logic programming languages that have emerged from the artificial intelligence paradigm. These languages, if linked to a database, can be viewed as powerful extensions to the relational database model standard to many GISs, providing inference capabilities and more expressive representations of knowledge). Rule-based reasoning systems supported by these languages can be valuable assets to complex resource management tasks. Moreover, the rigor automation enforces on expressing models in logic makes evident their structure and assumptions. Many of the models used in conservation planning are logical ones (e.g. a plant species is predicted to be on a site if it is below 500 m elevation and is on clay soils and is on a flat or gentle slope). Indeed, such logical formulations of models may be more appropriate for conservation planning than the gradient models traditional to vegetation science.

Though the Fauna-List and Fauna-Map programs have shown their utility, they have a number of limitations that affect their ability to serve as a general-purpose modeling tool for WHR in a GIS. First of all, these programs work with a fairly specialized database. They use a database that is a simplified extract of WHR, the input files need to be specially formatted, and the output file is in its own ASCII format. These characteristics are fairly easy to work with, but the set-up is neither self-explanatory nor well-integrated with other outside databases. Moreover, the system cannot easily accommodate the full WHR database, which includes the habitat elements and all the vegetation structural elements. Likewise, the system only treats suitability as a binary characteristic, though ideally it should make use of the different suitability levels in the full database. Secondly, the processing flow in these programs is too one-directional which makes interactive manipulations of the databases difficult. For instance, it would be useful to directly change the habitat characteristics of a polygon and see the resultant effects on maps of habitat suitability and species distributions. Finally, this sort of interface designed with these Form menu tools is not easy to extend and integrate into other applications. If these two programs were to be redesigned, it would make sense to integrate them with ARCVIEW, thus taking advantage of its consistent graphical interface and its scripting language (AVENUE) that allows a user to make their own custom interface. There are a number of advantages to embedding this system within ARCVIEW. For instance, one could directly edit the tables describing the habitat information, recalculate the distribution of a species, and immediately display the resultant change. Another advantage is that it is easier to control the display of the distributions against the background of any GIS coverage. Most importantly, WHR would be integrated fairly seamlessly with the rest of a well-supported GIS visualization package.

Monitoring environmental change

We anticipate that much of the research that follows on from our IBM-ERP will be focused on regional environmental monitoring. With funding from the U.S. Environmental Protection Agency, we recently initiated research with several other investigators and institutions with the objective of using physically-based models of energy and water balance, coupled to remote sensing and GIS data, to identify a spatial hierarchy of ecological "response units" for environmental monitoring.

Regional conservation planning and reserve design

The BMAS model developed for the Sierra Nevada Ecosystem Project was a notable advancement in reserve selection models. Nevertheless, we identified a number of features that could make the BMAS model more realistic and effective. Some improvements would be in the model structure or formulation, while others involve refinements in input data. The basic issue is in the assessment of biotic vulnerability. Using simple percentage representation targets, e.g., 10% of the distribution of a plant community, is only an estimate of the real conservation needs. The real objective is to maintain viability of species and communities. Viability is based on a combination of size and spatial configuration of reserves, management of unreserved lands, and biotic responses to that management. The current version of the BMAS model does not consider the spatial pattern of the selected watersheds. Based on general principles of conservation biology one could argue that larger, better connected BMAs would tend to maintain biodiversity better than small, poorly connected systems. We want to explore analytical means of evaluating solutions more rigorously in terms of the viability of protected populations. Contiguity is difficult to incorporate as a suitability factor, however, because it is not a property than can be measured a priori for a watershed but is dynamic in that it changes as its neighbors are selected. The BMAS model provides solutions that are the most efficient solutions only in terms of requiring the least area. Thus the solutions can be considered planning benchmarks in terms of the area requirements for representative BMA systems. Any additional constraints such as spatial design will increase the area of the solution. Further, new methods must be developed for evaluating viability of different solutions.

The BMAS model also does not handle scheduling of reserve allocation over time, variations in land costs, and trade-offs with other resources. Given that implementation of a BMA system might need to be scheduled rather than instantaneous, a method is needed to prioritize sites for allocation, analogous to a budget constraint. All private lands are currently treated as being equally unsuitable for BMA selection and less suitable than public lands in recognition of the cost of land acquisition. However, private lands can vary widely in value. We have discussed ideas with another SNEP science team member how these land costs could be estimated for each planning unit and incorporated into the suitability data of the model. Data for the Sierra Nevada were obtained for public lands allocated to grazing and commercial timber harvest. These were used in defining management classes which in turn were used to determine vulnerability of biodiversity elements. They were also used to evaluate alternatives in regard to selection of resource management lands as BMAs. It should be possible to revise the suitability data (or add an additional objective) such that the model minimizes conflict with other resources.

IBM-ERP Project Home Page

Biogeography Lab Home Page