Vertebrate Modeling








Avian Modeling Pilot Project with USFWS






Project Description

The Biodiversity and Spatial Information Center (BaSIC) at NC State University has partnered with the U.S. Fish and Wildlife Service (USFWS) to explore the potential of enhancing GAP vertebrate models in the southeastern U.S. (Southeast Gap Analysis Project) to address specific needs of the avian conservation community. The partnership benefits from high quality datasets and spatial modeling expertise offered by BaSIC, as well as from the network of USFWS biologists available to coordinate expert review and link the modeling work directly to management and monitoring activities. The primary goal of this research is to develop a scientifically rigorous framework for iteratively improving predictions of species distribution patterns. In addition, BaSIC is working with state, regional and national partners to improve techniques for mapping habitat suitability (Figure 1). After these methods are developed for a few species within a pilot study region, they will be applied to the 9 state SE-GAP region and may be expanded to include more species and/or spatially explicit population dynamics. The pilot project focuses on six bird species occurring along a shrub-forest gradient of landscape characterization within the North Carolina portion of the Appalachian Mountains Bird Conservation Region. A team of local experts selected these species in part because they were listed as priority species for monitoring and/or conservation efforts by Partners in Flight (Rich et al. 2004). Efforts will focus on 1) determining which SE-GAP species distribution model inputs have the greatest influence in predicting species' occurrences areas; 2) developing an extensive database of spatially explicit species-habitat-population relationships, 3) incorporating knowledge of habitat suitability, which is critical in setting conservation goals and objectives, into traditional presence/absence predictions; 4) assimilating existing bird survey records from multiple sources into a spatially explicit database: 5) validating North Carolina GAP and SE-GAP species distribution models; and 6) extrapolating and validating bird species occurrences using occurrence data and state-of-the-art techniques.

Figure 1. Contributions by BaSIC and the US Fish and Wildlife Service to regional habitat suitability models used in setting regional conservation goals.


Return to Top

Background

In September 2004, a meeting of regional biologists and modelers was hosted by the USFWS in Atlanta, GA to review a variety modeling approaches and the datasets available for modeling in the Southeast. The objectives of the meeting were to:

  1. Inform partners about current regional modeling efforts by the USFWS and BaSIC.
  2. Get feedback on preliminary efforts to match functional ecological systems (Comer et. al. 2003) with mappable land cover categories.
  3. Identify priority bird species.
  4. Review existing avian models for those species.
  5. Provide background on ancillary data available for use in modeling.
  6. Identify additional ancillary data that would be helpful for modeling efforts.
  7. Receive feedback from partners on additional methods/data that could be used to improve modeling.

At the meeting, presentations by the Mississippi Alluvial Valley Joint Venture office (MAV-JV) and the Upper Midwest Environmental Sciences Center (UMESC), as well as BaSIC described a range of modeling approaches that could be taken. Some of the feedback obtained at the meeting resulted in modifications and enhancements to the ancillary datasets developed for use in GAP models. USFWS biologists also declared interest in expanding the use of science-based modeling for conservation planning efforts. In particular, there was a desire to identify areas on the landscape that are either currently suitable for priority bird species or could be managed to improve the suitability. Scaling factors were requested for some model parameters so areas could be ranked with respect to habitat quality. The audience was also adamant about the need to quantify the sensitivity of prediction accuracies to model variables and thresholds. These recommendations were compiled and evaluated for their utility to increase the accuracy and information content of occurrence predictions. A pilot project was then established to develop and implement these enhancements to current SE-GAP modeling efforts.

A meeting in Asheville was used to help refine the pilot project objectives and list of focal species. Meeting participants included regional researchers, managers and bird species experts. To reduce computer processing times, it was decided to limit the pilot project study region to the portion of the Appalachian Mountains Bird Conservation Region falling within North Carolina. The focal bird species list was also reduced to six to simplify the project. Focal species' habitat associations span a wide range of forest age and structure. Reasons for species selection included priority listing by Partners in Flight and because most changes in landscape composition and structure in the revised study region area are associated with forest loss or conversion due to timber harvesting and residential development. The final list of focal species for the pilot project includes:

Study Species


Return to Top

Objectives

  1. Develop a map of avicentric land cover classes to define landscape composition in a way that may be more relevant to birds than existing land cover classes
  2. Develop a database of spatially explicit species-habitat associations and population responses described in the literature.
  3. Conduct sensitivity analysis on input map layers used to make SE-GAP predictions of species' occurrences.
  4. Create habitat suitability models based on expert rankings of species- habitat associations.
  5. Locate and integrate various bird inventories for knowledge-based model validation.
  6. Extrapolate and validate bird occurrences using existing inventories and state-of-the-art techniques.
Return to Top

Literature Review Database

Literature reviews of species-habitat relationships are conducted everyday for the purpose of scientific research and management. These efforts are generally inefficient as they are rarely documented and often repeated. Furthermore, information gleaned from the reviews is commonly summarized as text descriptions that cannot be queried or summarized in different ways for other purposes. BaSIC is currently developing a new method of documenting literature reviews by multiple persons that creates synergy from their activities through the use of a relational database and the Internet. BaSIC created database forms to record details of literature searches and research results. Details captured in the forms are spatially and temporally explicit and divided into modular units. In this way, each record of the database can be queried for information describing the study’s date and location, method of data collection, species studied, land cover types and landscape relationships (e.g., patch size, distance to water), as well as qualitative descriptions of habitat suitability and quantitative demographic parameters (e.g., density, mating success, daily nest survival, nest predation rates) under those conditions. Additional queries allow these data to be standardized and summarized in different ways. Efforts are currently underway to make the forms and queries available on-line so that the conservation community may work together to limit redundant efforts and build a robust repository of research results descriptions. Examples of database forms (Figures 2 & 3) and possible quesries (Figure 4) are provided below.

Figure 2. The database switchboard allows for selection of several forms to enter or view data. The search description form is used to document literature searches. Future searches will no longer need to include previously searched dates and/or databases. Articles identified in the searches are documented and linked to a separate bibliographic database.


Figure 3. Data entry form for describing spatially explicit habitat associations, ranking suitability estimates and quantifying population parameters. A single journal article often contains information that can be coded as multiple inputs (records) using this form. Values for a single input may contain data pulled from tables, figures, text results and study area descriptions.


Figure 4. Data can be transformed to standard units (e.g., individuals per ha) and queried in diverse ways.
Return to Top

Avicentric Classes

Landscape analyses of species-habitat associations should be conducted using maps representing functional differences in how focal species perceive and respond to landscape composition and structure. For this reason, USFWS has generated a list of avicentric land cover classes to spatially partition the southeast region. These classes were mapped for the study region using a combination of available data layers. In particular, National Land Cover Data (NLCD) were intersected with 16 landform classes. The resulting groups were aggregated and/or reclassified into avicentric classes (Figure 5).

Figure 5. 3-D projection of avicentric land cover classes in the Asheville region of NC.



Avicentric land cover classes.
Community Name Level 1 Level 2 Level 3 Code
Eastern Grasslands I     10000
    Tamaulipan Prairie I A   10100
    Tall Grass I B   10200
    Meadows / Florida and Georgia Prairies I C   10300
    Agricultural and Cropland I D   10400
    Pasture I E   10500
    Rank Annuals I F   10600
Freshwater Wetland Communities II     20000
    Non-Forested II A   20100
        Freshwater Emergent Marsh II A 1 20101
        Bogs / Fens / Ephemeral Wetlands II A 2 20102
        Mudflats / Sandbars II A 3 20103
    Forested II B   20200
        Bottomland Hardwood II B 1 20201
        Cypress-Tupelo II B 2 20202
        Atlantic White Cedar II B 3 20203
        Pocosin / Carolina Bays II B 4 20204
    Riparian II C   20300
    Open Fresh Water II D   20400
Coastal Communitites IV     40000
    Maritime Shrub-Scrub IV A   40100
    Maritime Forest IV B   40200
    Chenier / Oak Motte IV C   40300
    Estuarine Emergent Marsh IV D   40400
    Beaches and Dunes IV E   40500
    Tidal Mudflats IV F   40600
    Coastal Prairie IV G   40700
    Mangroves IV H   40800
Pine Communities V     50000
    Pine Flatwoods V A   50100
    Pine Savanna V B   50200
    Xeric Pine Scrub V C   50300
    Pine Plantations V D   50400
    Other Pine Forest V E   50500
        Other Pine Forest - Natural V E   50501
    Pine Sandhills V F   50600
Upland Hardwood / Pine-Hardwood Communities VI     60000
    Spruce-Fir VI A   60100
    Northern Hardwood VI B   60200
    Mixed Mesophytic (Cove Hardwood) VI C   60300
    Hemlock / White Pine / Hardwood VI D   60400
    High Elevation Oak / Oak-Pine VI E   60500
    Mixed Hardwoods VI F   60600
        Dry Mixed Hardwoods VI F 1 60601
        Mesic Mixed Hardwoods VI F 2 60602
    Pine Hardwoods VI G   60700
    Oak-Cedar VI H   60800
    Oak Savanna VI I   60900
    Hardwood Plantation VI J   61000
    Tropical Hardwoods VI K   61100
Pelagic VII     70000
    Continental Shelf VII A   70100
    Deep Open Water VII B   70200
    Gulf Stream VII C   70300
Cities / Towns / Suburbs VIII     80000
    Residential VIII A   80100
    Commercial Urban VIII B   80200
    Airfields / Golf Courses / Cemeteries VIII C   80300
Additional classes IX     90000
    Quarry / Strip Mines / Gravel Pits IX A   90100

Return to Top

Sensitivity Analysis

Each input data layer associated with species' occurrences will be assessed for its influence on modeled occupancy areas. For example, if a species known to occur within 100m of water, this occupancy rule will be tested for its independent contribution to the total area classified as species presence by the model. Independent contributions of each model parameter will be summarized for all species to identify parameters commonly making large contributions to model predictions.

The influence of map errors on model predictions will also be assessed. Map errors under consideration include both spatial errors and compositional errors. For example, we will reclassify random pixels as land cover classes in proportion to the cell frequency of correctly and incorrectly classified pixels in the confusion matrix of the land cover map to induce intrinsic spatial error into land cover descriptions. Compositional errors will be induced in several ways. For example, a low-pass filter may be applied to the land cover map to investigate the effects of land cover spatial precision on model output.

Return to Top

Habitat Suitability Models

Spatial models based on habitat affinities derived from the literature review database and expert review will be created for each of the priority bird species. Data layers of traditional Boolean GAP models depicting presence/absence will be ranked into suitability levels. Categorical variables (avicentric classes and their categorical modifiers) will be ranked directly. Continuous variables will first be coded as having one of many possible response curves (Figure 6). Thresholds of suitability along the response curve will then be set. The ranked data layers will then be summarized to describe spatial gradients of habitat suitability for focal species across the study region. Besides suitability, simple qualitative descriptions of variable importance and use will also be recorded when possible. Below are more detailed descriptions:

1. Importance: how important (either good or bad) is the condition (land cover category, or other data layer) to the species during the spring and summer? 2. Suitability: rank the suitability of the parameter for the species life history during the spring and summer. When ranking, consider that the species may not necessarily change behaviors under different suitability conditions. 3. Use: A simple, generic list of behaviors regarding how the specis will behave and use the condition.

Figure 6. Response types available for ranking habitat quality in the suitability models' continuous variables.


Return to Top

Locate and Integrate Data

In order to validate SE-GAP presence/absence maps and create a repository of data for the development of more sophisticated data-driven models, occurrence data are being collected and organized for the priority species throughout the three state study region. Data sources so far include NC Wildlife Resources Commission, Dr. Ted Simons, state breeding bird atlases, the USGS Breeding Bird Survey, local Natural Heritage Programs' element occurrences, R8Bird data collected by the U.S. Forest Service, data from the Monitoring Avian Productivity and Survivorship program, museum records and any other digital data sets that may become available. These diverse data will be imported into a relational database to derive spatially and statistically appropriate datasets for model creation and validation.

Return to Top

Data-driven models

Several powerful techniques have recently become available for extrapolating wildlife distribution patterns using relationships between locations where species have been observed and mapped environmental conditions at those locations. Environmental data available for modeling include data sets already being complied by SE-GAP (disturbance, canopy closure, imperviousness, distance to water, elevation, slope, etc.). Other regional data sets to be acquired include rainfall, temperature and frost-free days among others. Important predictors for individual species can be identified by several methods including principal components analysis, hierarchical partitioning (Mac Nally 2002), classification and regression tree analysis (CART; De'Ath and Fabricius 2000), as well as expert opinion.

Using these new modeling techniques, locations that have not been surveyed will be labeled for species presence or absence based on their "similarity" to surveyed locations where species were or were not detected. However, the techniques differ in how "similarity" is defined as well as their predictive accuracies and interpretability of the mechanisms behind the predicted patterns. For these reasons, SE-GAP will investigate the appropriateness of these techniques under different conservation objectives. Depending on data availability, some species will be selected for model development and evaluation. Only species with adequate data quality, quantity, and dispersion throughout the study area will be modeled. However, the "gap" in data availability for other species will be noted as a future research priority. Several inductive modeling techniques may be explored for each species including; "biophysical envelopes" such as DOMAIN (Carpenter et al. 1993) and PHASE1 (Laurent et al. 2005), multiple logistic regression, CART, and maximum entropy (Phillips et al. 2004) models. Each of these techniques has advantages and disadvantages which could make them suited for a particular species or data set. In addition, multiple models may be used for each species providing a more robust evaluation of habitat suitability including the possible use of deductive rules such as land cover use.

Return to Top

Habitat Model Variables

Return to Top

Literature Cited:

Carpenter, G., A.N. Gillson, and J. Witmer. 1993. DOMAIN: a flexible modeling procedure for mapping potential distributions of plants and animals. Biodiversity and Conservation 2:667-680.

Comer, P. D., Faber-Langendoen, R. Evans, S. Gawler, C. Josse, G. Kittel, S. Menard, M. Pyne, M. Reid, K. Schulz, K. Snow, and J. Teague. 2003. Ecological Systems of the United States: A Working Classification of U.S. Terrestrial Systems. NatureServe, Arlington, VA.

De'Ath, G. and K. E. Fabricius (2000). Classification and regression trees: a powerful yet simple technique for ecological data analysis. Ecology 81(11): 3178-3192.

Laurent, E.J., H. Shi, D. Gatziolis, J.B. LeBouton, M. Walters and J. Liu. (2005) Using the spectral and spatial precision of satellite imagery for analyses of wildlife distribution patterns. Remote Sensing of Environment 97:249-262.

Mac Nally, R. C. 2002. Multiple regression and inference in ecology and conservation biology: further comments on identifying inportant predictor variables. Biodiversity and Conservation 11: 1397-1401.

Phillips, S.J., J. Dudik, and R. E. Schapire. 2004. A maximum entropy approach to species distribution modeling. In Brodley, Carla E., editor. Machine Learning, Proceedings of the Twenty-first International Conference (ICML 2004), Banff, Canada, July 4-8, 2004. ACM Press, New York, NY.

Rich, T. D., C. J. Beardmore, H. Berlanga, P. J. Blancher, M. S. W. Bradstreet, G. S. Butcher, D. W. Demarest, E. H. Dunn, W. C. Hunter, E. E. Iñigo-Elias, J. A. Kennedy, A. M. Martell, A. O. Panjabi, D. N. Pashley, K. V. Rosenberg, C. M. Rustay, J. S. Wendt, T. C. Will. 2004. Partners in Flight North American Landbird Conservation Plan. Cornell Lab of Ornithology. Ithaca, NY.


Return to Top

For More Information:

  Contact: Dr. Ed Laurent   Contact: Steve Williams  
  Science Coordinator for the Bird Conservation Institute   Vertebrate Mapping Coordinator  
  NC Field Office, American Bird Conservancy   220 David Clark Labs  
      Biodiversity and Spatial Information Center  
    Department of Biology, NCSU  
    Raleigh, NC 27695-7617  
    919-513-7413  


Last updated: Oct. 4, 2006