Janet Franklin and John Stephenson
We are using image processing of satellite imagery, and ecological gradient models implemented as cartographic models in a GIS, to produce regional (totaling about 2 million hectares) fine-scale (2-ha minimum mapping unit) digital vegetation maps for the US Forest Service in Southern California. Attributes mapped include vegetation community type, and forest cover and structure (canopy size class). Automated image segmentation is used to delineate mapping units corresponding to vegetation stands from satellite imagery. The databases will be used for a number of natural resources and land management planning applications. We illustrate the usefulness and limitations of these data for evaluating the effect of forest fragmentation on the viability of the California Spotted owl (Strix occidentalis occidentalis), a Forest Service management indicator species. A habitat suitability model for the Spotted Owl was developed and applied to the vegetation database, and evaluated for its ability to capture known owl nesting sites in the San Bernardino Mountains. The resulting habitat quality map will be used as input to a spatially explicit simulation model of owl population dynamics, which includes habitat-specific demographic parameters.
The management of ecosystems requires inventory and monitoring of large areas of natural landscapes at fine scales. Increasingly, modeling is being used as a research and management tool to examine spatio-temporal processes such as land use conversion, natural disturbance, resource harvesting, and species dynamics (hence, this conference). A lot has been written about trade-offs among spatial, temporal, and categorical resolution in geographic databases, often with reference to existing and planned satellite sensors, issues of data storage and processing, issues of scale and accuracy, and so forth. A number of continental- and global-scale land cover, vegetation, and biodiversity databases are being developed or planned, for example continental and global land cover maps (Loveland et al. 1991, Townshend et al. 1991), MODIS land cover products (Running et al. 1994), global vegetation databases used in modeling the general circulation of the atmosphere (Henderson-Sellers 1994), Landsat Pathfinder datasets (Maiden 1994, James and Kalluri 1994), and the Gap Analysis Program biodiversity datasets (Scott et al. 1993). However, these data may lack the spatial detail and categorical attributes required for some ecosystem management applications.
In this paper we will discuss a vegetation database developed for a large area (2 million ha) at a relatively fine spatial scale (2 ha MMU), with moderate taxonomic detail (vegetation types mapped at the series level) but including information on forest cover and canopy structure. Mapping is based on image processing of Landsat satellite data, and predictive mapping based on terrain and other variables in a GIS. The database is expensive to produce, but we will demonstrate that data of at least this categorical and spatial resolution are required over large regions (several million ha) for ecosystem management. Ecosystem processes operate across large spatio-temporal scales, therefore modeling is essential to addressing landscape management questions (Sample 1994, J.F. Franklin 1993, 1995). The example we will use, from our ongoing research, is that of a habitat suitability model developed for the California Spotted Owl and applied to the vegetation database. The resulting habitat quality map will be used as input to a spatially explicit simulation model of owl population dynamics, which includes habitat-specific demographic parameters.
Certain salient aspects of the mapping methods will be reviewed briefly, as they relate to the integration of remote sensing with GIS, and issues of data structures. Then, the data requirements of a spatially explicit wildlife population simulation model will be outlined, to demonstrate the necessity for stand-specific estimates of vegetation structure and composition. Finally, the preliminary results from our attempt to parameterize and implement a habitat suitability model for the California Spotted Owl will be presented. The habitat model is evaluated based on its ability to predict known owl nesting site locations in the San Bernardino Mountains. Errors can result from lack of precision and accuracy in the vegetation database. However, a database lacking the attributes or spatial resolution of ours would not be useful at all for this type of modeling.
As growing numbers of researchers and resource managers rely on digital geographic data, and look to remotely sensed imagery as a source of data for their GIS, issues of geographic data models and remote sensing scene models (Goodchild 1992, 1994, Strahler et al. 1986), as well as image processing algorithm development, are central to research on the integration of remote sensing and GIS. Our mapping methods integrate advanced image processing algorithms (segmentation, canopy modeling, mixture modeling), applied to Landsat Thematic Mapper (TM) imagery, with simple cartographic modeling using GIS, to produce fine-scale digital vegetation maps for large areas. These methods have been described in detail elsewhere (Woodcock et al. 1994a, Franklin and Woodcock, submitted, and references therein, and see Strahler 1981, and Franklin et al. 1986 for background).
Briefly, our methods result in a digital vegetation map with the following attributes:
Importantly, each attribute listed above is interactively edited by an image analyst based on air photo interpretation, usually resulting in the correction of 5- 20% of the labels for each attribute.
Central to our mapping methods, attributes are derived from fine-grained raster data (e.g., 30-m TM and DEM pixels), but the fundamental mapping unit is the "stand" composed of contiguous pixels, derived automatically using image segmentation, and with a minimum size specified (Woodcock and Harward 1992). White and Running (1994, 690) noted that errors arise at all steps of topographic and satellite data processing, and "these errors can be reduced, assuming they are random, by aggregating the land units which have similar environmental characteristics." Averaging our estimates of categorical and continuous attributes over areas larger than the single grid cell tends to reduce errors in the estimates of those spatially autocorrelated geographic phenomena.
The accuracy assessment for the digital vegetation map referred to in this paper has not been completed. However, the accuracy of other forest maps based on the same methods have been evaluated using an innovative approach based on fuzzy set theory (Gopal and Woodcock 1994), resulting in overall accuracy for the life form attribute of about 95%, and for the vegetation series and cover labels of about 60-80% with the greatest errors occurring among similar classes (Woodcock et al., 1994b). The database was costly to produce, very roughly $0.4/ha, not including the initial purchase of satellite imagery, digital elevation models, digital line graphs, and other digital coverages (streams and water bodies, roads, land ownership and administrative boundaries). Photointerpretation-based map editing and field data collection represent the greatest costs.
Spatially explicit population models (SEPMs) are simulation models that provide a means to assess how habitat patterns influence population dynamics and viability (Dunning et al. 1995). Species' populations are frequently distributed in disjunct patches of suitable habitat embedded in a large landscape matrix. Spatial and temporal variation in habitat patches (e.g. size, structure, vegetation type) and underlying landscape patterns (e.g. distance between patches, dispersal corridors) can profoundly influence population viability (Soule et al. 1988). Habitat patterns influence populations because demographic rates (i.e. survivorship, fecundity, and dispersal success) are affected by variations in habitat quality and configuration (Pulliam 1988). High quality habitats generally have positive demographic rates and contribute to population growth (i.e. "source" habitats). Conversely, low quality habitats can have negative rates and promote population decline (i.e. "sinks") (Pulliam 1988).
SEPMs address the influence of habitat patterns by incorporating digital maps of the landscape (that can be changed over time to simulate dynamic processes) and habitat-specific demographic parameters into a temporal population model. This makes it possible to investigate how changes in landscape configuration, brought on by human land use practices and natural disturbance and successional processes, will affect wildlife populations over time. It has been noted, however, that problems associated with parameter estimation and model complexity are magnified for spatially explicit models, which may give the "illusion of exactitude in the absence of hard information" (Doak and Mills 1994). This is particularly true for the estimation of habitat-specific survivorship rates and dispersal behavior, since there are very few species for which we have empirical data on these parameters. Nevertheless, a SEPM designed to describe the spatial dynamics of wildlife populations in fragmented habitats with a minimum of life history data could at least be used to separate species for whom habitat pattern alone is a good predictor of carrying capacity from those whose population dynamics require a more detailed understanding of mobility and demographics (Schumaker, unpublished manuscript). To date most SEPMs have been developed for relatively well-studied endangered species (Pulliam et al. 1992, McKelvey et al. 1993).
One objective of our ongoing research is to evaluate the effect of present and projected habitat fragmentation on the Spotted Owl in the southern California mountains using a SEPM developed specifically for this species (Schumaker 1995). We focus on the spotted owl because: a) the southern California population is small, its habitat is fragmented, and a panel of experts has expressed concern about its long-term viability (Verner at al. 1992); b) it is considered an "umbrella species" for other forest-dependent species and a management indicator species by the Forest Service (Verner et al. 1992); c) an ongoing demographic study in the San Bernardino Mountains provides the habitat-specific information about demography, dispersal behavior, and habitat selection required to parameterize a SEPM (LaHaye et al. 1992); d) prior analyses of spotted owl dynamics in the southern California mountains have not explicitly considered landscape habitat patterns (Noon and McKelvey 1992, LaHaye et al. 1994); and e) the digital vegetation database for the study area, described in the last section, has a set of attributes mapped at a spatial scale (2 ha MMU) that is appropriate for modeling spotted owl habitat suitability.
Detailed studies of spotted owl habitat in the Pacific Northwest (Thomas et al. 1993), the Sierra Nevada and southern California (Verner et al. 1992) have shown vegetation type (forest composition), forest canopy cover, and forest structure (distribution of tree size) to be correlated with habitat suitability. Prior to the mapping effort described above, there was no digital map of southern California forests that estimated each of these attributes. It has been suggested in the literature that the grain of a map of habitat attributes should be no more than 1% of the target species' average home range size (reviewed in Schultz and Joyce 1992). Our vegetation type labels nominally meet this requirement; the minimum polygon size is 2 ha, the median is roughly 4 ha and the mean is approximately 15 ha (Franklin and Woodcock, submitted); the territory size for Spotted Owl in the study area is highly variable, roughly 200-1000 ha depending on elevation (LaHaye et al. 1992). The only other completed vegetation database that covers the study area has a MMU of 100 ha (Davis et al. 1995) and an average polygon size of 600-1000 ha (Franklin and Woodcock, submitted). No other existing regional database has information on conifer and hardwood forest cover for all forest types.
The SEPM we are using partitions the landscape into a uniform array of hexagonal cells, where the hex cell size approximates the average territory size for a pair of owls. The model evaluates the quality of habitat within hex cells based on user-defined scores for each habitat class, and identifies cells which are suitable breeding territories based on a user-defined threshold score. The goal is to select a hex cell size, habitat classes, class scores, and a territory threshold score that will produce a map that closely approximates the true number and spatial configuration of potential owl territories in the landscape. To identify appropriate values for these parameters, we are fortunate to have one southern California mountain range (the San Bernardino Mountains) where the spotted owl population has been intensively studied and a full inventory of owl territories has been completed (LaHaye et al. 1992). What we present here is preliminary work using known owl nest locations from the San Bernardino Mountains to develop reliable habitat suitability parameters for the SEPM.
As previously described, the three mapped vegetation attributes are vegetation type, tree crown size, and percent canopy cover. Going into this analysis, there were several clear trends in owl/habitat relationships that guided our thinking: 1) empirical data shows that spotted owls nest and roost predominately in dense, mature forest stands with large trees (see summary in Verner et al. 1992); and 2) owl territories in low-elevation forest types (bigcone Douglas fir and live oak) are in distinctly smaller patches of forested habitat than territories found in high elevation forest types (mixed conifer and Jeffrey pine). The latter relationship can be seen clearly in Figure 1, where the small low- elevation forest patches in the lower left portion of the map contrast sharply with the continuous forests found at higher elevations. The sizable variability in habitat quantity per territory presents a problem for parameterizing a model that requires a single hex cell size. We addressed this problem by adopting a hex cell size (345 ha) that is skewed towards the lower end of the estimated territory size range (200-900 ha) to increase the relative importance of small forest patches, and assigning higher scores to the low- elevation forest types to compensate for their smaller quantities.
Examinations of the mapped vegetation immediately around owl nest sites indicated that the nest sites were not as highly correlated with high crown size and canopy classes as would be expected from field measurements taken at the nests. Based on the field data, most nests should fall in the 80-100% canopy and 24-40+ crown diameter classes and essentially all of them should be in or above the 50-79% canopy class. What we found for the entire San Bernardino Mountain study area, using the mapped categories, was 48% (147 of 305) of the nest sites in the 80-100%/24-40+ class and 82% (250 of 305) in or above the 50-79% canopy class. Also, 10% (31 of 305) of the nest sites were identified as being within vegetation types that spotted owls do not nest in (chaparral and pinyon pine).
There are several possible explanations for the observed discrepancies between the field data and the map. One is that some stands may simply be mislabeled; there are undoubtedly some errors in the map and a classification accuracy of 82% for nest sites by canopy class is reasonable to expect. Also, preliminary accuracy assessments suggest that the three mapped vegetation attributes (i.e. type, size, and cover) may vary significantly in their accuracy level, with tree size being the least accurate of the three. However, it is also likely that scale differences between the field data and the map is a factor. Field data on forest structure around nest sites was collected using tenth-acre plots, while the map's minimum mapping unit is approximately 5 acres. It is probable that in some areas the forest structure immediately around the nest site is not representative of the larger mapped unit. This is almost certainly a factor in the appearance of nest sites in non-forest types; some low-elevation nests are located in extremely narrow riparian forest stands surrounded by chaparral.
These map scale and accuracy issues were considered when identifying the habitat classes to use in the model. Although we initially started with a good idea of the habitats that are most important to spotted owls at the micro-scale, it was necessary to determine which mapped, stand-scale vegetation classes best correlated with known nest sites. We did this by quantifying the amounts of each vegetation class found within the 345 ha hexagon cells that encompassed known nest sites (Table 1). In this way, we were able to identify which mapped attributes were good predictors of owl nests and which were not. Based on this analysis, we decided to drop tree crown size from the model and focus on vegetation type and canopy cover. Although the highest crown size class correlated well with nests in some areas, it did not correlate well with known nests in oak woodlands. With over 60% of this vegetation type concentrated in a single, mid-range size class, this attribute was not effective at discriminating good habitat in live oak woodland.
Possible scores that can be assigned to habitat classes range from 0 to 99. Bigcone Douglas fir with 50-100% canopy cover (the most important low-elevation forest type) received the maximum score of 99. Other vegetation type/canopy cover classes were scored proportionally lower based on their importance as owl habitat, their relative abundance in a territory, and the vegetation types they are associated with. A key factor in scoring habitat classes is the threshold score that a hex cell must exceed in order to be identified as a territory. We selected a threshold score of 27 based on trial and error. This threshold value achieved adequate representation of the low-elevation territories embedded in chaparral. At this threshold and with chaparral habitat scored at 8, a hex cell with 21% bigcone Douglas fir (72.5 ha) surrounded by chaparral would be designated as a territory.
The most important high-elevation forest type, mixed conifer with 80-100% canopy cover, was scored based on its mean quantity (areal extent) in known owl territories relative to mean quantities found in bigcone Douglas-fir territories. On average, it is 56% more abundant than bigcone Douglas-fir, thus its score was calculated to be 56% of 99 (55) (see Table 1). Forest types in the lower canopy cover classes were scaled down based on their considerably lower value as owl habitat. However, an exception to this was in mixed conifer habitats embedded in Jeffrey pine forest. Our analysis found that owl territories in these more arid, rainshadow forests were strongly associated with patches of mixed conifer habitat that have canopy cover of 50-79%. To account for this, we gave higher weighting to this habitat class.
Figure 2 provides a visual representation of how the model-predicted territories correspond with known owl nest sites and Table 2 provides a numeric comparison by both vegetation type and geographic region. The model predicted 131 territories in a mountain range where there are 135 known territories. More importantly, although there is significant error in locational accuracy, the model performed well in estimating total numbers both by vegetation type and by geographic region. Thus, both the size and the spatial distribution of the population are realistically simulated. Clearly, the model isn't perfect. Model-predicted territories are more tightly clustered than actual ones and they tend to overvalue oak woodland habitats in some areas because of difficulties in distinguishing mature stands of trees from dense brushy ones. However, for spatially explicit population modeling, this territory map captures far more spatial and categorical detail about the owl population's distribution than have previous studies (Noon & McKelvey 1992, LaHaye et al. 1994). The real test will be how the model performs when applied to the other southern California mountain ranges.
We have implemented a suit of mapping methods, relying on digital image processing and the integration of remote sensing with GIS, to estimate vegetation attributes (composition and structure) at a fine grain (2-20 ha stands) for a large region (2 m ha). This paper demonstrates that the resulting database contains a necessary set of attributes at a resolution sufficient for modeling the distribution of suitable habitat for the California Spotted Owl, a forest-dependent umbrella species whose numbers are declining. The habitat-specific, spatially explicit population modeling that is currently in progress will hopefully provide results that indicate the effect of habitat fragmentation on population viability, and will identify habitat that is critical, on the basis of its location, for the viability of the metapopulation.
The lack of perfect correspondence between modeled and actual territory locations, based on the habitat suitability model, could reflect inaccuracy or imprecision in the mapped attributes as discussed above. We are currently conducting an analysis of the sensitivity of the habitat suitability model to error in the attribute labels (Ellen Hines, Masters thesis in progress). It might be possible, using photointerpretation- or field-based editing, to improve the accuracy of the variables that are key to a particular application; our mapping methods lend themselves to this kind of updating. Another possibility is that habitat variables important to spotted owls are not mapped as attributes, and may be difficult to capture in a large-area map. Examples would be local and subtle variations in canopy structure or microclimate, affecting prey populations. This has been discussed in the literature on GIS-based habitat suitability modeling. The forest structure attribute that is most difficult to estimate from remote sensing, or from field data for that matter, is tree crown size. We are exploring the use of spatial variance in high-resolution SPOT panchromatic imagery (10 m pixels) to better estimate forest canopy structure (David Shaari, Masters thesis in progress).
This southern California database has also been used for other environmental modeling applications. Jennifer Swenson (1995) simulated future urban development in the region surrounding the Santa Monica Mountains National Recreation Area (administered by the National Parks Service), and examined its potential effect on the connectivity of the remnant natural vegetation. The results predicted a disproportionate impact of potential future urbanization on certain sensitive vegetation types (oak woodland and coastal sage scrub) primarily due to their proximity to current urban land use. Modeling also indicated that patterns of future urban development are likely to increase the fragmentation of the remaining wildlands, due to the patterns of land ownership in the Santa Monica Mountains.
We hope, in the future, to use predictive vegetation modeling (Franklin 1995) to add detail on vegetation composition to the database. We have tested this approach, with some limited success, in the Cleveland National Forest in San Diego County. Chaparral series were predicted from satellite and terrain data using classification tree modeling (McCullough 1994), and riparian vegetation types were also predictively mapped from terrain variables using a classification tree (Gray 1994). While the accuracy of predictive mapping was low relative to thematic map accuracy standards, the predicted patterns made sense. We feel there may have been problems with the field data used for calibration and validation but collected for other purposes, including ambiguous labeling of vegetation type and poor registration with the digital database. However, being able to predictively map the distribution of dominant plant species at this scale with reasonable confidence would allow us to model the spatial dynamics of the vegetated landscape based on characteristics of the disturbance regime ,and the life history traits of those plant species (see, for example, Mladenoff in press). This, in turn, could be tied back in to a SEPM to examine in a realistic way how a landscape altered by environmental change or land management practices might affect a management-indicator species.
We would like to thank the Greg Nichols, Nathan Schumaker, and Joseph Shandley for their contributions to this research.
Davis, F.W., Stine, P.A., Stoms, D.M., Borchert, M.I. and Hollander, A.D. (1995) Gap analysis of the actual vegetation of California 1. The southwestern region. Madrono 42: 40-78.
Doak, D.F. and Mills, L.S. (1994) A useful role for theory in conservation. Ecology 75: 615-626.
Dunning, J.B., Jr., Stewart, D.J., Danielson, B.J., Noon, B.R., Root, T.L., Lamberson, R.H., and Stevens, E.E. (1995) Spatially explicit population models: current forms and future uses. Ecological Applications 5: 3-11.
Franklin, J. (1995) Predictive vegetation mapping: geographic modelling of biospatial patterns in relation to environmental gradients. Progress in Physical Geography 19:494-519.
Franklin, J., Logan, T., Woodcock, C. E., and Strahler, A. H. (1986) Coniferous forest classification and inventory using Landsat and digital terrain data, IEEE Transactions on Geoscience and Remote Sensing GE-24, 139.
Franklin, J. and Woodcock, C.E., (submitted) Multiscale vegetation data for the mountains of Southern California: spatial and categorical resolution, in, Scaling of Remotely Sensed Data for GIS, D. A. Quattrochi and M. F. Goodchild, eds., Lewis Publishers.
Franklin, J. F. (1993) Preserving biodiversity: species, ecosystems, or landscapes?, Ecological Applications 3: 202-205.
Franklin, J. F. (1995) Why Link Species Conservation, Environmental Protection, and Resource Management?, in Linking Species and Ecosystems, Jones, C. G. and Lawton, J. H., Eds., New York: Chapman & Hall. pp. 326-335.
Goodchild, M. F. (1992) Geographical data modeling, Computers & Geosciences 18: 401-408.
Goodchild, M. F. (1994) Integrating GIS and remote sensing for vegetation analysis and modeling: methodological issues. Journal of Vegetation Science 5: 615-626.
Gopal, S. and Woodcock, C. E. (1994) Theory and methods for accuracy assessment of thematic maps using fuzzy sets. Photogrammetric Engineering and Remote Sensing 60: 181-188.
Gray, C.A. (1995) Predicting the location of riparian vegetation using Landsat TM and digital terrain data in the Cleveland National Forest. Masters Thesis, Department of Geography, San Diego State University, San Diego CA.
Henderson-Sellers, A. (1994) Global terrestrial vegetation 'prediction': the use and abuse of climate and application models. Progress in Physical Geography 18: 209-246.
James, M., and Kalluri, S.N.V. (1994) The Pathfinder AVHRR data set: an improved coarse resolution data set for terrestrial monitoring. International Journal of Remote Sensing 15: 3347-3363.
LaHaye, W.S., Gutierrez, R.J., and Call, D.R. (1992) Demography of an insular population of spotted owls (Strix occidentalis occidentalis). In: McCullough, D., Barrett, R.H., eds. Wildlife 2001: Populations. New York: Elsevier Press; pp. 803-814.
LaHaye, W.S., Gutierrez, R.J., and Akcakaya, H.R. (1994) Spotted owl metapopulation dynamics in southern California. Journal of Animal Ecology 63: 775-785.
Loveland, T., Merchant, J., Ohlen, D., and Brown, J. (1991) Development of a land-cover characteristics database for the coterminous United States. Photogrammetric Engineering and Remote Sensing 57: 1453-1463.
Maiden, M. (1994) The Pathfinder project. International Journal of Remote Sensing 15: 3333-3345.
McCullough, P.E. (1995) Predictive vegetation mapping using classification tree analysis. Masters Thesis, Department of Geography, San Diego State University, San Diego CA.
McKelvey, K., Noon, B.R., and Lamberson, R.H. (1993) Conservation planning for species occupying fragmented landscapes: the case of the northern spotted owl. In: Kareiva, P.M., Kingsolver, J.G., and Huey, R.B., eds. Biotic Interactions and Global Change. Sunderland, MA: Sinauer Press. pp. 424- 450.
Mladenoff, D.J., Host, G.E., Boeder, J., and Crow, T.R. (in press) LANDIS: A spatial model of forest landscape disturbance, succession, and management. In, Goodchild, M., Steyaert, L.L., Parks, B.O., Crane, M.P., Johnston, C.A., Maidment, D.R. and Glendenning, S., editors, GIS and environmental modeling: Progress and research issues, Ft. Collins, CO: GIS World, Inc.
Noon, B.R. and McKelvey, K.S. (1992) Stability properties of the spotted owl metapopulation in southern California. In: Verner, J., McKelvey, K.S., Noon, B.R., Gutierrez, R.J., Gould, G.I., Beck, T.W., tech. coordinators. The California Spotted Owl: a technical assessment of its current status. General Technical Report PSW-GTR-133. Pacific Southwest Research Station, Forest Service. U.S. Department of Agriculture. pp. 187-206.
Pulliam, H.R. (1988) Sources, sinks, and population regulation. American Naturalist 132: 652-661.
Pulliam, H.R., Dunning, J.B. Jr., and Liu, J. (1992) Population dynamics in complex landscapes: a case study. Ecological Applications 2: 165- 177.
Running. S.W., and 13 other authors (1994) Terrestrial remote sensing science and algorithms planned for EOS/MODIS. International Journal of Remote Sensing 15:3587-3620.
Sample, V. A., Ed., (1994) Remote Sensing and GIS in Ecosystem Management. Washington, DC: Island Press.
Schultz, T.T. and Joyce, L.A. (1992) A spatial application of a marten habitat model. Wildlife Society Bulletin 20:74-83.
Schumaker, N.H. (1995) Habitat connectivity and spotted owl population dynamics. Ph.D. Dissertation. University of Washington: College of Forest Science.
Schumaker, N.H. (Unpubl. manuscript) A simulation approach to quantifying landscape quality.
Scott, J.M. and 11 other authors (1993) Gap analysis: a geographical approach to protection of biological diversity. Wildlife Monographs 123:1-41.
Soule, M.E., Bolger, D.T., Alberts, A.C., Sauvajot, R.S., Wright, J., Sorice, M., and Hill, S. (1988) Reconstructed dynamics of rapid extinctions of chaparral-requiring birds in urban habitat islands. Conservation Biology 2: 75-92.
Strahler, A. H., (1981) Stratification of natural vegetation for forest and rangeland inventory using Landsat digital imagery and collateral data. International Journal of Remote Sensing 2: 15-41.
Strahler, A. H., Woodcock, C. E., and Smith, J. A. (1986) On the nature of models in remote sensing. Remote Sensing of Environment 20: 121-139.
Swenson, J. (1995) Habitat fragmentation in the Santa Monica Mountains: current status and future predictions. Masters Thesis, Department of Geography, San Diego State University, San Diego CA.
Thomas, J.W., Forsman, E.D., Lint, J.B., Meslow, E.C., Noon, B.R., and Verner, J. (1990) A conservation strategy for the Northern Spotted Owl. Interagency Scientific Committee to Address the Conservation of the Northern Spotted Owl (USDA: Forest Service, USDI: Bureau of Land Management, Fish and Wildlife Service, and National Park Service). 1990-791-171/20026. United States Government Printing Office, Washington, D.C. USA.
Townshend, J.R.G., Justice, C., Li, W., Guerney, C., and McManus, J. (1991) Global land cover classification by remote sensing: present capabilities and future possibilities. Remote Sensing of Environment 35: 243-255.
Verner, J., McKelvey, K.S., Noon, B.R., Gutierrez, R.J., Gould, G.I., Beck, T.W., Technical Coordinators. (1992) The California spotted owl: a technical assessment of its current status. General Technical Report PSW-GTR-133. Pacific Southwest Research Station, Forest Service. U.S. Department of Agriculture.
White, J. D. and Running, S. W. (1994) Testing scale dependent assumptions in regional ecosystem simulations. Journal of Vegetation Science 5: 687- 702.
Woodcock, C. D., Collins, J., Gopal, S., Jakabhazy, V. D., Li, X., Macomber, S., Ryherd, S., Harward, V. J., Levitan, J., Wu, Y., and Warbington, R. (1994a) Mapping forest vegetation using Landsat TM imagery and a canopy reflectance model. Remote Sensing of Environment 50: 240-254.
Woodcock, C. W., Gopal, S., Macomber, S., and Jakabhazy, V. (1994b) Accuracy Assessment of the Vegetation Map of the Plumas National Forest, Boston: Boston University Center for Remote Sensing, 18 pp.
Woodcock, C. E. and Harward, V. J. (1992) Nested-hierarchical scene models an image segmentation. International Journal of Remote Sensing 13: 3167-3187.