Table of Contents

CHAPTER 2

Land Cover Classification and Mapping

Of all the branches of botany there is none whose elucidation demands so much
preparatory study, or so extensive an acquaintance with plants and their affinities,
as that of their geographic distribution.
- Sir Joseph Dalton Hooker

2.1 Background

Vegetation patterns are an integrated reflection of the physical, chemical, and biotic factors that shape the environment of a given land area (Whittaker 1965). As such, gap analysis relies on maps of dominant land cover types as the most fundamental spatial component for the analysis of terrestrial environments (Scott et al. 1993). The mapped extent and distribution of existing land cover is used in gap analysis to evaluate the management status of natural land cover types in Wyoming, to provide a spatial database for modeling wildlife habitat and vertebrate distributions across Wyoming, and to establish a single temporal data set of current land cover patterns in Wyoming for future reference (Stoms 1994). Because gap analysis was conceived to provide conservation assessment of large areas, Landsat Thematic Mapper (TM) data were chosen as the basis for mapping land cover. TM data provide sufficient spectral and spatial resolution for land cover discrimination and are available for the entire United States, providing a consistent base for the National GAP (Scott and Jennings 1994).

Although each state conducting gap analysis uses methods appropriate to mapping land cover in their region, land cover mapping standards have evolved to insure that the products of state gap projects are compatible and allow their integration into regional and national products (Jennings 1993). National standards for land cover mapping required the use of TM satellite imagery less than 3 years old at the initiation of the project, classification of land cover types and wetlands consistent with a national template (Jennings 1993, Cowardin et al. 1992), specific cartographic criteria (i.e., MMU of 100 ha for land and 40 ha for wetlands, map products at a scale of 1:100,000) and land cover mapping into adjacent states to facilitate regional edge-matching of land cover maps. A review of existing land cover maps in Wyoming showed that neither state-wide maps (Wyoming Department of Agriculture 1987), nor maps of large portions of the state (Despain 1990, United States Forest Service Resource Inventory System [USFS RIS] data) provided both the spatial resolution and the land cover classes necessary to satisfy the GAP standards. As a result, a new land cover map for Wyoming that met national standards was developed based on the protocols described below.

2.2 Methods

2.2.1 Rationale For Visual Interpretation vs. Digital Classification.

Two general approaches have been used to develop land cover maps from digital TM imagery for GAP: digital classification and visual interpretation. Digital classification assigns image pixels to cover classes based on statistical differences in spectral characteristics. Classes are defined either before classification (supervised) or after (unsupervised) and pixels are assigned to the classes using any of a suite of statistical techniques (Richards 1993). The resulting classes can be refined using other sources of information, such as elevation data, existing maps, or field reconnaissance. Digital classification requires considerable computational resources both for preparation of images prior to classification and for the digital classification. Each TM scene must be classified either individually or all scenes must be corrected to eliminate differences caused by atmospheric characteristics unrelated to the target land cover before classification. The resulting per-pixel classification must be aggregated to the standard MMU of 100 ha, a non-trivial task because individual pixels must be merged with adjacent pixels by applying aggregation rules that can vary across the landscape (Stoms 1994). The primary advantages of digital classification are that classes are statistically consistent and the classification results are repeatable.

The second approach, visual interpretation of the satellite imagery, uses a human interpreter to define areas of homogeneous land cover. Difficulties with the visual interpretation method arise from subjective interpretation by different analysts and from human errors, some of which are difficult to document. On the other hand, visual interpretation requires fewer computer resources than computer classification, both in data storage and central processing unit time, and aggregation is not necessary because units are drafted to fit the MMU. In effect, aggregation is accomplished during mapping using rules that make sense in the landscape context. Individual TM scenes are not atmospherically corrected, and edge-matching between scenes is accomplished by extending the map from one scene to the next as it is created. Perhaps most importantly, the ability of the human analyst to integrate texture and context with spectral information allows discrimination of cover types which might not be discernible based on spectral characteristics alone (Estes et al. 1983). For these reasons, and based on the success of mapping efforts by the CA-GAP (Davis et al. 1995), Wyoming chose to adopt the visual interpretation approach.

2.2.2 Classification System

Development of the land cover classification for the WY-GAP project was constrained by several practical considerations. First, the land cover map had to be compatible with the habitat types used to map vertebrate distributions. Second, the cover types had to be discernible on Landsat TM imagery. Third, types had to be consistent with national standards and the classifications of surrounding states (Jennings 1993).

The Wyoming land cover classification was developed in 1991 based on a vegetation classification by Jones (1992) and was consistent with the UNESCO classification scheme for vegetation (Driscoll et al. 1984). Later, Jennings (1993) outlined the UNESCO system as a template for GAP classifications. The UNESCO system organizes vegetation communities into a hierarchical structure with classes based on gross physiognomy at the coarsest level (also referred to as level 1), and community types based on dominant species composition at the finest level (level 6). GAP required land cover at the cover type or alliance (level 5) whenever possible, but practical constraints sometimes forced the mapping of combinations of several cover type units. The classification system developed for WY-GAP (Appendix 2.1) was crosswalked to the Wyoming Game and Fish Department habitat classification at the outset of the project to ensure that our types were compatible with existing vertebrate habitat associations. Detailed descriptions and range maps for the Wyoming cover types are provided in a separate land cover atlas (Merrill et al. 1996a) and an example of the atlas is presented in Appendix 2.5.

Because of their disproportionate importance in an arid state like Wyoming, wetlands were considered at several levels in WY-GAP. We use the term "wetlands" to refer to areas defined by Cowardin et al. (1992) as both wetlands and as deep water habitat. These areas include bogs, swamps, marshes, ponds, lakes and riparian areas (vegetation associated with streams and rivers) and any other environments where standing or moving water is present or where saturation by water is the key factor controlling the ecology of the area. Wetlands are included as types in the classification (e.g. open water, forested riparian, grass-dominated wetlands) and are mapped as primary or secondary types within polygons when they are larger than the wetland MMU (40 ha). We also used a wetland attribute to describe wetland inclusions within polygons using the classification of Cowardin et al. (1992), even when the inclusion was small in extent. Finally, a riparian/aquatic model was developed for the purpose of improving the predicted distributions of species with riparian/aquatic associations (see section 3.2.3).

2.2.3 Imagery Acquisition and Processing

All image processing for WY-GAP was performed using the Map and Image Processing Systems (MIPS) (MicroImages Inc., Lincoln, Nebraska) and is described in detail by Thurston (1993). Twenty-three Landsat TM scenes were used to create the bulk of the Wyoming land cover map (Fig. 2.1, Appendix 2.2). All imagery contained < 10% cloud cover and was acquired from mid-June to late August between the years 1984 and 1993; scenes older than 1988 were updated with new TM data prior to the release of the map in 1995. Cloudy areas, though minimal in Wyoming, were handled either by using alternative cloud-free TM data, or, in a few cases, by extrapolating polygon boundaries across small clouds. Eight of the 23 TM scenes were terrain-corrected (Appendix 2.2). A small area in southeastern Wyoming was digitized from a combination of Satellite Pour l'Observation de la Terre (SPOT) imagery and the 1987 Wyoming Land Inventory (WLI) map (Wyoming Department of Agriculture 1987) because TM data for that area were not available (Fig 2.1).

Images were georeferenced by establishing a relationship between an image coordinate system (line, column) and a map coordinate system (e.g. Universal Transverse Mercater [UTM], Lambert). We identified control points on the image that could also be located on 1:24,000 scale USGS topographic sheets. Approximately 18 control points were distributed across each image as 9 pairs of 2 points each. This strategy was a compromise between the CA-GAP approach, which used 8 to 12 control points, and the UT-GAP approach, which used 9 clusters of 5 points.

  Index      TM       Index      TM       Index      TM       Index      TM     

 Number   Path/Row   Number   Path/Row   Number   Path/Row   Number   Path/Row  

    1      38/29        8       38/30      15       36/31      22       36/32   

    2       38/29       9       37/30      16       35/31      23       35/32   

    3       37/29      10       36/30      17       35/31      24       34/32   

    4       36/29      11       35/30      18       34/31      25       33/32   

    5       35/29      12       34/30      19       WLI*                        

    6       34/29      13       38/31      20      SPOT**                       

    7       38/30      14       37/31      21       37/32                       

Figure 2.1. Landsat TM scenes used to develop the WY-GAP land cover map. Numbers on the map refer to the path and row of the TM satellite imagery in the table.
(graphic figure not available in html)

Points with root mean square (RMS) errors greater than one pixel (30 m) for rows and two pixels (60 m) for columns were inactivated. Column errors were slightly larger than row errors due to the interaction of terrain with the geometry of the TM sensor (Thurston 1993). Data were not warped to fit the control points because tests showed that although warping could force residual errors of the control points to zero, areas between control points showed little improvement (Thurston 1993). The TM data were resampled from full resolution to a 100-m degraded pixel size to reduce data storage and processing time. Davis et al. (1995) found that for mapping to a relatively large MMU (100 ha) over large areas at the 1:100,000 scale, little information was lost by degrading the original data to 100-m pixels. We used an affine (linear) transformation model and nearest neighbor resampling for most of the TM data. Four scenes processed in the latter part of the project were resampled using a 3rd-order polynomial transformation (Appendix 2.2). A normalized contrast enhancement was applied to each of the three spectral bands used for interpretation. Contrast enhanced TM spectral bands 3, 4 and 5, representing red and near-infrared portions of the spectrum, were used to create false color composite images for photointerpretation.

2.2.4 Land Cover Mapping

On-screen digitizing

Vector polygons enclosing "landscape units" (Davis et al. 1995) were drafted manually, on-screen, using the enhanced TM composite images as guides. These units consisted of either a single homogeneous land cover type or mixtures of several land cover types which together occupied an area equal to or greater than the 100 ha MMU. Polygons were generally drawn over the imagery displayed such that a 100-m TM pixel covered about 1 mm on the screen. This simulated an approximate scale of 1:100,000, but image magnification was increased or decreased to more accurately delineate features when necessary. Although paper maps for WY-GAP are produced at a scale of 1:100,000, the concept of scale for digital data has no meaning, since the data may be viewed on the computer screen at any scale. As digitizing progressed from one TM scene to the next, lines were extended into the new scene to create a seamless final product.

Riparian and wetland areas are spectrally distinct regions on the satellite imagery. These areas were mapped on the land cover map as separate polygons when they were both larger than the wetland MMU (40 ha) and wider than 2 pixels in the imagery. Smaller or narrower riparian/wetland areas were subsumed by surrounding polygons and noted as polygon attributes. Riparian areas were also modeled in more detail as a separate GIS layer because of their disproportionate importance as vertebrate habitat (see section 3.2.3).

Disturbance (e.g. logging, fire) in some parts of Wyoming affects areas larger than the 100 ha MMU. Disturbed land cover types were included in the classification system as clearcut conifer and burned conifer (Appendix 2.1). These types were mapped from the satellite imagery using the same procedures as for other, non-disturbed types, because they comprised a significant part of the Wyoming landscape and because existing vegetation (rather than potential) was used to predict animal habitat. Less clearly defined seral vegetation (e.g. old growth forest) was not mapped because it is difficult to distinguish using satellite imagery without extensive ground truthing.

Polygon topology was built after the initial digitizing using Arc/Info and problems such as dangling nodes, unclosed polygons, and polygons smaller than the MMU were identified, and corrected or eliminated. The positions of polygon boundaries were examined, and corrected if necessary, during polygon attributing and after field review. In most cases this involved deleting polygon boundaries that did not correspond to features in the imagery and re-drafting them. In a few cases, map notes by field reviewers were used to re-draft boundaries.

Polygon attributing

Attributes assigned to each polygon describe primary and secondary cover types, the relative area of each polygon occupied by these types as well as other important features occurring in the polygon (Table 2.1). Because predictions of vertebrate species distributions were based on primary and secondary land cover types in each polygon, these attributes were completed for all polygons in Wyoming. Other data fields provided important information (i.e., disturbance, forest crown closure) about the composition of the polygons and were filled when information was available.

Literature, existing maps (Appendix 2.3), and field reconnaissance were used to assign land cover attributes to polygons. Published papers, theses, and federal and state reports were useful for local areas. Small-scale maps of the entire state (e.g., WLI) and larger scale maps of particular areas of the state (e.g., USFS RIS data) were used when they were available. In addition to existing documentation, we conducted field reconnaissance along nearly 16,000 km of road transects throughout the state, and recorded land cover on USGS 1:100,000 scale topographic maps for photointerpretation of the satellite imagery. Sources of information for attributing polygons, and whether the polygon attributes were checked on the ground, are documented in tables linked to each polygon (see Metadata section in Chapter 7).

Edge-matching to other states

Polygon boundaries were extended at least to closure and often to > 10 km into surrounding states (Montana, South Dakota, Nebraska, Colorado, Utah, and Idaho) to facilitate regional edge matching. Edge matching from Colorado to Wyoming was performed by CO-GAP personnel. At the completion of the WY-GAP land cover map, corresponding maps were not available for Montana, South Dakota, Nebraska or Idaho. Edge matching between the western states will be accomplished by consensus between these states, orchestrated by the National GAP.

Area calculations

In this chapter, we present two area calculations for land cover types in Wyoming. The area of land cover polygons in Table 2.2 is the sum of the area of all polygons for each primary and secondary type. The proportional area of land cover was derived by multiplying the area of a polygon by the percent of the polygon occupied by the primary and secondary land cover types (Table 2.2). The proportional area gives a closer approximation to the area of each of the land cover types in the state than either the primary area or secondary area alone. While these proportional areas are useful for approximating the actual area of cover types in the state, they cannot be used to determine the location. This is because the database only records a percentage of variation of the primary and secondary cover types, but the variation is not mapped. Therefore, all area statistics presented in this report (with the exception of this chapter) are based on the area of land cover polygons, not the proportional area of land cover.

2.3 Results

The WY-GAP land cover classification includes 41 primary and secondary cover types (Table 2.2, Map 2.1). Not all these types are consistent with the cover type level (level 5) of the UNESCO classification, the template provided for the land cover classification (Jennings 1993), since practical constraints forced mapping of some combinations of cover type units. For example, herbaceous tundra and shrub-dominated tundra types were combined into a single alpine tundra class since the two types were indistinguishable on TM imagery. Other examples where combinations occurred are listed in the separate volume of appendices (Appendix 2.5, Merrill et al. 1996a), along with definitions of the 41 cover type classifications presented here.

Two cover types, Wyoming big sagebrush (30.8%) and mixed-grass (20.2%), occupied about half of the land area of the WY-GAP land cover map, based on the proportional area of land cover (Table 2.2). Lodgepole pine (6.1%) and Ponderosa pine (2.7%) comprised the greatest amount of forested area. Irrigated agriculture occupied 4.2% of the land area of Wyoming. The rarest land cover types in the state were basin big sagebrush, bur oak, and bitterbrush (Table 2.2). Mesic shrub, bur oak and basin big sagebrush occurred more often as a secondary type than a primary type. These types were rare in Wyoming, did not usually occur in patches larger than 100 ha, were difficult to distinguish from other types using satellite imagery, or were not mapped due to a combination of these reasons. Rare types were often found in the ecotones between the more common cover types or in unique micro-habitats, such as places where topography and wind interacted to enhance snow accumulation.

2.4 Accuracy Assessment

As of this writing, no formal state-wide validation of the Wyoming land cover map has been undertaken. Additional funding has been provided to validate the map using aerial videography, initiated in fall of 1996 and be completed by the end of 1998. Aerial videography is currently being used to provide an error estimate of thematic accuracy in the land cover map. It may also provide useful training data for a next generation mapping effort.

Prior to this validation, two informal efforts were conducted as pilot studies for full validation. During the summer and fall of 1993, WY-GAP personnel conducted a statistically designed assessment of 4, small subsections of the land cover map which included both montane and basin land cover (Ball et al. 1994). A priori accuracy estimates for each cover type were used to determine the number of field samples necessary to estimate map accuracy within 10 % of the true value, 95% of the time. The a priori estimates were "best guesses" by the original interpreter. Accuracy of primary and secondary attribute data for the test polygons was determined in the field by surveying a 450-m transect through the approximate center of each polygon. The proportion of each land cover type encountered along the transect was recorded and eventually compared to the primary and secondary cover designations from the land cover map by analyzing an error matrix with rows representing cover from the land cover map and columns representing cover from field observation (Story and Congalton 1986).

This preliminary accuracy assessment was not successful for two reasons. First, it was difficult or impossible to access a large number of the randomly chosen polygons due to private ownership and poor roads. Second, even when polygons were accessible, their large size made it impossible to sample intensively enough from the ground to assess the overall, relative proportions of primary and secondary types in the polygon. Thus, differences found in land cover designations of polygons between ground sampling and photointerpretation of satellite imagery were more a function of the scale of perspective than a true test of the accuracy of polygon classification (Ball et al. 1994). To gain a true measure of polygon composition on the ground would require many long transects located randomly throughout the polygon. This pilot study provided a basis for estimating the costs of more intensive validation efforts.

During the summer of 1994, personnel from the Bureau of Land Management (BLM), U.S. Forest Service, National Park Service, Soil Conservation Service (SCS), U.S. Fish and Wildlife Service (USFWS), Wyoming Game and Fish Department, and TNC performed informal spot checks of primary and secondary attributes by visiting polygons during the course of their normal activities. In some cases, there were multiple reviews of the same map area. In total, 133 copies of 1:100K quadrangle maps were distributed and 51 were returned, covering 38 of the 56 (68%) quadrangles in Wyoming. These 38 maps were either partially or completely checked by field personnel during the course of their normal activities (Appendix 2.4). The field personnel either noted the correct cover type on the mylar or indicated that the original designation was correct. Additional notes on polygon content were also made on separate data sheets which we provided. Of the 14,690 polygons, 1809 (12.3 %) were checked. Reviewers reported that based on their field reconnaissance, 1439 polygons (79.6 %) were labeled correctly for primary land cover. Mislabeled polygons were corrected before the release of the map. The most common errors reported were confusion between agricultural areas and riparian zones (these types are frequently intermingled in Wyoming and were mapped as single polygons) and confusion between juniper woodland and xeric shrub communities (both occur in similar spectral situations). These problems are discussed in more detail in the following section.

2.5 Limitations and Discussion

Visual interpretation of satellite imagery required subjective decisions during the drafting of polygon boundaries, and during interpretation of cover within each polygon. Several steps were taken to mitigate this subjectivity. A priori rules were used (e.g., zooming guidelines, riparian corridor minimum widths; see Methods) to increase consistency among digitizing personnel. In addition, boundaries were checked and, if necessary, adjusted, several times between the first draft and the final product. Polygon attributes were assigned by a single interpreter for all but a few polygons in the southeastern corner of Wyoming. The identity of the interpreter for each polygon is included in associated metadata tables (see Chapter 7).

The coarse scale (1:100,000) and large MMU (100 ha) of the land cover map restricts it to use for large area management and for regional analyses. The Wyoming land cover map was not designed for use in analyses and management at finer scales. Cognizance of the issues and limitations imposed by map scale for spatial analyses is critical, and is the responsibility of the map user. Areas calculated using spatial data, such as the Wyoming land cover map, are very sensitive to map scale and resolution (Davis et al. 1995). Areas occupied by the Wyoming cover types reported here are not comparable to areas calculated from map products at other scales, because finer-scale maps depict boundaries with more detail than is possible in the WY-GAP map, which in turn affects area calculations. Also, fine-scale maps may have a smaller MMU, and therefore may include smaller units in calculation of area.

Some of the cover types mapped for WY-GAP occupied huge areas and spanned environmental gradients. Because of this, there can be large variation in the appearance of some of these types across the landscape. Canopy coverage, physiognomic habit and subdominant species can vary within a single land cover type and this variation could not be mapped within the constraints of WY-GAP. The most important example of this is Wyoming big sagebrush, the single most common cover type in Wyoming. Very often it occurred in rolling terrain over which it varied in its coverage and composition by orders of magnitude. This land cover type should be understood and interpreted as a complex gradient-mosaic, of which Wyoming big sagebrush is the dominant species over most of the area (Reiners et al. 1989, Burke et al. 1989).

Some land cover types presented particular problems for mapping from TM imagery. These problems were overcome using additional data when available. Conifers in Wyoming (e.g., lodgepole pine, Englemann spruce, and subalpine fir) have similar spectral signatures, occur in similar environments and are often in adjacent or mixed stands. We used USFS RIS data, field reconnaissance and, in a few cases, digital elevation data to help identify boundaries between conifer types. Shrublands and grasslands in Wyoming form a complex matrix that is patchy in some places and homogeneously mixed in others, and spectral separation was difficult. To separate these types, we relied primarily on field data and site context. Areas in the eastern part of Wyoming are more likely to be grass-dominated, while the western two-thirds of the state are primarily shrub-dominated. Juniper woodlands and xeric shrub communities both occur on shallow soils and rock outcrops in Wyoming and their spectral signatures are dominated by the substrate rather than by vegetation. Efforts to correct this problem were based on field review.

Because irrigated agriculture and riparian areas are often intermingled and difficult to separate spectrally, and because of their disproportionate importance to vertebrates, especially in the arid Wyoming climate, we made additional efforts to model riparian area in more detail as a separate layer (see section 3.2.3).

2.6 Summary and Conclusions

Wyoming big sagebrush (30.8%) and mixed grass (20.2%) occupied about half the land area of Wyoming. Lodgepole pine (6.1%) and Ponderosa pine (2.7%) comprised the greatest amount of forested area. Rare types more often occur as secondary types than a primary types. Formal assessments of the land cover map will be completed in 1998. Informal assessment of the map indicated a thematic accuracy of 79.6%. Despite several caveats we discuss about the WY-GAP land cover map, it is a useful representation of Wyoming land cover that represents a "snapshot" of the actual land cover of the state in time. Although land cover in Wyoming changes very slowly for the most part, it is by nature dynamic (i.e., the 1988 Yellowstone fires) and change with time will not be reflected in the current version of the map. We hope that this map will be updated and maintained over time, but users should be aware that some changes in land cover may have already occurred since the completion of the Wyoming land cover map.