Chapter 2 Land Cover Classification and Mapping

2.1 Introduction

Mapping natural land cover requires a higher level of effort than the development of data for animal species, agency ownership, or land management, yet it is no more important for gap analysis than any other data layer. Generally, the mapping of land cover is done by adopting or developing a land cover classification system, delineating areas of relative homogeneity (basic cartographic “objects”), then labeling these areas using categories defined by the classification system. More detailed attributes of the individual areas are added as more information becomes available, and a process of validating both polygon pattern and labels is applied for editing and revising the map. This is done in an iterative fashion, with the results from one step causing re-evaluation of results from another step. Finally, an assessment of the overall accuracy of the data is conducted. The final assessment of accuracy will show where improvements should be made in the next update (Stoms et al. 1994).

In its “coarse filter” approach to conservation biology (e.g., Jenkins 1985, Noss 1987), gap analysis relies on maps of dominant natural land cover types as the most fundamental spatial component of the analysis (Scott et al. 1993) for terrestrial environments. For the purposes of GAP, most of the land surface of interest (natural) can be characterized by its dominant vegetation.

Vegetation patterns are an integrated reflection of the physical and chemical factors that shape the environment of a given land area (Whittaker 1965). They also are determinants for overall biological diversity patterns (Franklin 1993, Levin 1981, Noss 1990), and they can be used as a currency for habitat types in conservation evaluations (Specht 1975, Austin 1991). As such, dominant vegetation types need to be recognized over their entire ranges of distribution (Bourgeron et al. 1994) for beta-scale analysis (sensu Whittaker 1960, 1977). These patterns cannot be acceptably mapped from any single source of remotely sensed imagery, therefore, ancillary data, previous maps, and field surveys are used. The central concept is that the physiognomic and floristic characteristics of vegetation (and, in the absence of vegetation, other physical structures) across the land surface can be used to define biologically meaningful biogeographic patterns. There may be considerable variation in the floristics of subcanopy vegetation layers (community association) that are not resolved when mapping at the level of dominant canopy vegetation types (alliance), and there is a need to address this part of the diversity of nature. As information accumulates from field studies on patterns of variation in understory layers, it can be attributed to the mapped units of alliances.

2.2 Land Cover Classification

Land cover classifications must rely on specified attributes, such as the structural features of plants, their floristic composition, or environmental conditions, to consistently differentiate categories (Kuchler and Zonneveld 1988). The criteria for a land cover classification system for GAP are: (a) an ability to distinguish areas of different actual dominant vegetation; (b) a utility for modeling animal species habitats; (c) a suitability for use within and among biogeographic regions; (d) an applicability to Landsat Thematic Mapper (TM) imagery for both rendering a base map and from which to extract basic patterns (GAP relies on a wide array of information sources, TM offers a convenient meso-scale base map in addition to being one source of actual land cover information); (e) a framework that can interface with classification systems used by other organizations and nations to the greatest extent possible, and (f) a capability to fit, both categorically and spatially, with classifications of other themes such as agricultural and built environments.

For GAP, the system that fits best is referred to as the National Vegetation Classification System (NVCS) (FDGC, 1997). The origin of this system was referred to as the UNESCO/TNC system (Lins and Kleckner in press) because it is based on the structural characteristics of vegetation derived by Mueller-Dombois and Ellenberg (1974), adopted by the United Nations Educational, Scientific, and Cultural Organization (UNESCO 1973) and later modified for application to the United States by Driscoll et al. (1983, 1984). The Nature Conservancy and the Natural Heritage Network (Grossman et al. 1994) have been improving upon this system in recent years with partial funding supplied by GAP. The basic assumptions and definitions for this system have been described by Jennings (1993).

As noted in the introductory, Pennsylvania’s contemporary landscapes are largely a legacy of historic human disturbance including marginal agriculture, strip mining, and extensive deforestation, often followed by fire. Whereas the initial GAP view relegated disturbed lands to minor interest, they are integral to Pennsylvania’s habitat and landscape integrity issues that are crucial to conservation. Physiography has substantially determined degradation under human influence and, therefore, whether land-use pressure has intensified or been alleviated in favor of return to more naturalistic conditions. Whether as a consequence of restoration or progressive fragmentation, structure and composition of vegetation often varies in a complex manner at relatively fine scales with a limited suite of species in different mixes comprising the overstory. Floristic distinctions could not be made consistently in this vegetative complex on the basis of Landsat satellite data that varied widely with respect to season of collection. Since the primary goal of the Pennsylvania GAP Project has been to lay a foundation of landscape-level conservation perspectives upon which to build in future work, it has been viewed as important that data sources not be over-extended in any manner that would be cause for future investigators to essentially discard these first efforts and start over completely.
Instead of vegetation alliances, therefore, the gap analysis effort for Pennsylvania has focused on land cover in terms of physiognomy, disturbance, and landscape integrity.

2.3 Mapping Standards

Land cover mapping for Pennsylvania has been conducted at two scales. Eight types of physiognomy have been combined with 3 intensities of disturbance at a resolution of 2 hectares or better giving a 24-class mapping of generalized land cover to serve both habitat modeling and a broad range of other uses. A binary mapping of landscape matrix as naturalistic (forest and water) versus humanistic (herbaceous and barren) has been conducted at 100-hectare resolution for landscape ecological analysis. The 8 types of physiognomy are: water, evergreen forest, mixed forest, deciduous forest, (woody) transitional, perennial herbaceous, annual herbaceous, and barren/hard-surface/rubble/gravel. The 3 intensities of disturbance are rural, suburban, and urban. Comprehensive mapping of more detailed floristics and/or vegetation structure has not been undertaken for this initial conservation analysis, although it was investigated and found to be infeasible using the remote sensing imagery available.

2.4 Methods

Generalized land cover and disturbance were mapped in several modes from Landsat Thematic Mapper (TM) digital image data collected during a period from 1991 through 1994. The Landsat image data were obtained from USGS EROS Data Center through the Multi-Resolution Land Characterization (MRLC) consortium. Image data were compressed through a hyperclustering protocol configured at Penn State Univ. for display and classification using commercial software. The compressed images have been made available to the public and have received considerable use in Pennsylvania as backdrops for GIS applications. An initial binary classification into naturalistic and humanisitic types of landscape matrix at 100-ha resolution was done by interactively digitizing with a mouse on a computer display. An interpretive classification of disturbance was done similarly, but with no specific minimum resolution. By reference to digital orthophoto quarter-quads (DOQQs), clusters were interpretively assigned to 8 general physiognomic land cover classes. Combining land cover and disturbance yielded 24 map classes for habitat modeling and analysis. Aerial videography was acquired along transects and used for validation. Details in these regards are given in the ensuing paragraphs.

2.4.1 The Land Cover Classification Scheme:

The Pennsylvania 100-ha binary landscape matrix layer has codes as follows:

10 = naturalistic (forest, water);
20 = humanistic (transitional, perennial herbaceous, annual herbaceous, barren).

The Pennsylvania 2-ha land-cover/disturbance layer has a two-digit coding scheme, for which the first digit is coded as:

1 = Rural (wild land or agriculture);
2 = Suburban (primarily low-density residential);
3 = Urban (primarily high-density residential and/or commercial/industrial); with the second digit being a code for physiognomy as follows:

1 = Open water or wetlands with standing water;
2 = Evergreen forest (not more than 30% of tree canopy cover deciduous);
3 = Mixed forest (deciduous and evergreen both > 30% of tree canopy cover);
4 = Deciduous forest (not more than 30% of tree canopy cover evergreen);
5 = Woody transitional (5%< cover of woody plant foliage<40%), also shrubland
or forest regeneration;
6 = Perennial herbaceous (grasslands, pasture, forage, old fields <5% shrubs);
7 = Annual herbaceous (row crops, grain crops, exposed mineral soil);
8 = Barren, hard-surface, rubble, gravel.

These latter classes form a natural ordination not only for physiognomy, but also for near-infrared spectral brightness. Spectral confusion is more likely for classes that are adjacent in the ordination than for classes that are further apart. Additional levels of classification were considered in cooperation with other northeastern states, but could not be implemented consistently using available image data.

2.4.2 Imagery Used:

The primary source of remotely sensed image data used in land-cover/disturbance mapping was from the Landsat Thematic Mapper (TM) sensor in paths 14-18 and rows 31 & 32, with coverage as shown in Figure 2.1. The data for these images were acquired from USGS EROS Data Center through GAP participation in the MRLC (Multi-Resolution Land Characterization) consortium. Each frame consisted of six bands, not including the thermal infrared. The image dates obtained for these path/row positions are listed in Table 2.1. Delivery of the image tapes was considerably delayed, which became a major cause of protraction for the Pennsylvania GAP Project. Several of the image dates were also considerably less than ideal for land classification, being acquired in phenological circumstances when trees had only partial foliage or were devoid of foliage. Clouds in portions of several images also required substantial remedial effort.

Figure 2.1. Landsat TM coverage by path/row position.

Table 2.1. Dates of Landsat TM imagery.

Digital orthophoto quarter-quads (DOQQs) derived from 1:40,000 scale black and white aerial photographs were used as a supplement to the Landsat data. The DOQQs are made publicly available by the Pennsylvania DCNR Topographic and Geologic Survey via the PASDA website http://www.pasda.psu.edu for downloading at no cost.

For purposes of validation, dual-resolution aerial videography was collected along transects using a light plane and camcorder equipment maintained by national GAP to support state projects. General location of aerial videography transects is shown in Figure 2.2.

Figure 2.2. Aerial videography transects for use in validation.

2.4.3 Land Cover Map Development:

The 100-ha binary landscape matrix mapping was the first land-cover product developed by the Pennsylvania GAP Project. This was accomplished by displaying a 3-band color-infrared composite of Landsat TM on a computer workstation, and interactively performing interpretive digitizing with a mouse. The digitizing work was then cleaned, edited, and assembled into polygons. When the interpretive mapping was topologically consistent, it was then generalized to 100-ha resolution by dissolving the boundaries of smaller polygons.

The second major image analysis undertaking was inspired by the Spectrum? initiatives of the Khoros Group at Los Alamos National Laboratory. The hypercluster concept was appealing, but not its implementation. An alternative scenario for hyperclustering was configured by customizing ERDAS? image analysis facilities. This was approached as a form of image data compression that produces a hybrid image-map layer to be displayed and analyzed using ESRI™ GIS software facilities. The available Landsat TM images were compressed in this manner, and the compressions made publicly available on both CD-ROM and the PASDA website (http://pasda.psu.edu). Since the customized method of hyperclustering with ERDAS was still somewhat awkward and restrictive, a suite of independent software facilities carrying the acronym PHASES (Myers 1999) was developed in generic C language and made publicly available on the Worldwide Web. This latter software development work was conducted under an NSF/EPA project. Studies on extensions of hyperclustering concepts for a variety of image analysis purposes are continuing under different sources of support.

Hypercluster compressions of the 10 scenes for spring dates provided the primary basis for unsupervised classification of physiognomy. Each scene had 255 clusters, making a total of more than 2,500 clusters to be labeled. Cluster labeling was accomplished by interactive image interpretation using ArcView? by ESRI™. Two computers were used simultaneously for the interpretation, with one displaying the clustered image-map, and the other displaying a panel of DOQQs. A sample of DOQQs was used in the manner of training sets, with additional samples being used to check consistency of interpretation across the image.

Clouds were an obstacle to image analysis, as they usually are in Pennsylvania. Most of the clouds could be eliminated in mosaicking the classified scenes by carefully choosing which scene took precedence in the area of overlap. Some clouds, however, were too centrally located to allow this expedient. Fortunately, dual dates were available for these instances. Accordingly, the alternate date was also classified and the map from the clouded image was patched from the second date.

Unsupervised classification was conducted at Landsat TM 30-meter pixel resolution. The classification of physiognomy was then generalized to a 2-ha level of resolution in two steps using the ANXPHASE facility of the PHASES image analysis system (Myers 1999). The first step performed generalization from pixel level to 1-ha level. The second step went from the 1-ha level to the 2-ha level. ANXPHASE is computationally intensive, but its strategy is parallel in several respects to the way a human interpreter would generalize. As it works, it always looks for the next smallest patch in the vicinity and blends it with the neighboring type having the greatest border.

The mapping of urbanized disturbance was developed in a manner analogous to the mapping of landscape matrix, but with a few modifications in technique. The first modification was to speed up and simplify the process by displaying the compressed image instead of a 3-band composite, which allowed for faster reloading of graphics when panning and zooming. The second was to overlay the compressed image with a digital file of roads in order to lend emphasis to urbanized areas. Digitizing was then done interactively via mouse. There was no specific minimum mapping unit for digitizing urbanized areas.

In combining the raster theme of physiognomy with the vector theme of urbanized areas,
The first operation was to rasterize the urbanized layer. The urbanized raster was then reclassified to a second digit coding; after which the two layers could be combined by directly adding their codes.

2.4.4 Special Feature Mapping:

Wetlands data were extracted from a land-use/land-cover classification performed by MRLC. Although the MRLC mapping was deemed to have insufficient landscape fidelity for habitat modeling in most respects, MRLC analysts had used National Wetlands Inventory (NWI) as an ancillary data source. Along with open water, two wetland types were transferred from NWI by MRLC: palustrine herbaceous wetlands and palustrine woody (shrub and forested) wetlands. For use in habitat modeling, each of these two classes was isolated from the MRLC digital map as a separate layer. Digitized NWI quads are only available for 2/3 of Pennsylvania, and they are not merged into a single layer; thus, we elected not to use them in partial manner. In addition, NWI coverage usually detects only about half of Pennsylvania’s wetlands. We used variable width buffers along steams, rivers, lakes, and wetlands to capture a considerable portion of aquatic habitats.

2.5 Results

The landscape ecological insight that arises from mapping at both broad and fine scales is evident from the contrast between the broad-scale landscape matrix map in Figure 2.3 and the fine-scale landscape matrix map in Figure 2.4. The broad-scale landscape matrix depiction comes directly from mapping naturalistic versus humanistic cover at 100-ha resolution; whereas, the fine-scale landscape version comes from combining water with forest classes and herbaceous with barren classes in the mapping of physiognomy at 2-ha resolution. In the broad-scale mapping, the Pittsburgh Low Plateau in southwestern Pennsylvania appears to be nearly as deforested as the Great Valley and Piedmont Lowland in southeastern Pennsylvania. In the fine-scale mapping, the Great Valley and Piedmont Lowland continue to appear as essentially devoid of forest cover; whereas, the Pittsburgh Low Plateau shows a partial but considerably fragmented forest cover. The Pittsburgh Low Plateau may also lose much of its remaining naturalistic cover if fragmentation continues. According to the broad-scale mapping, 69% of Pennsylvania has a naturalistic landscape matrix with 65% of the state being in one expansive unit that encompasses much of the northern third of the Commonwealth and extends through the mountains to the southern border. In this sense, landscapes are relatively intact over much of Pennsylvania.

The separable mapping of urbanization as shown in Figure 2.5 is also revealing from a landscape perspective. Pennsylvania is predominantly rural, with 1.5% of its area being intensively urbanized and another 4.1% being suburban. Much of the urbanization is due to a few large metropolitan areas such as Philadelphia, Pittsburgh, Harrisburg, Erie, and Wilkes-Barre/Scranton. Suburban sprawl is, however, a contemporary issue of concern that has been emphasized by the Governor’s 21st Century Environment Commission (Seif and Glotfelty 1998).


Figure 2.3. Broad-scale mapping of naturalistic versus humanistic landscape matrix.

Figure 2.4. Fine-scale mapping of naturalistic versus humanistic landscape matrix.

Figure 2.5. Urbanized and suburban areas of Pennsylvania.

Table 2.2 provides a percentage breakdown of the 11,618,719 hectares mapped in Pennsylvania according to land cover and urbanization. Figure 2.6 is a color plate showing the major components of the land cover and urbanization mapping.

Table 2.2. Percentage breakdown of Pennsylvania by land cover and urbanization.

2.6 Accuracy Assessment

2.6.1 Introduction:

GAP land cover maps are primarily compiled to answer the fundamental question in gap analysis: what is the current distribution and management status of the nation’s major natural land cover types and wildlife habitats? Besides giving a measure of overall reliability of the land cover map for Gap Analysis, the assessment also identifies which general classes or which regions of the map do not meet the accuracy objectives for the Gap Analysis Program. Thus, the assessment identifies where additional effort will be required when the map is updated. We report the results of the accuracy assessment, believing that the map is the best map currently available for the project area.

The purpose of accuracy assessment is to allow a potential user to determine the map’s “fitness for use” for their application. It is impossible for the original cartographer to anticipate all future applications of a land cover map, so the assessment should provide enough information for the user to evaluate fitness for their unique purpose. This can be described as the degree to which the data quality characteristics collectively suit an intended application. The information reported includes details on the database’s spatial, thematic, and temporal characteristics and their accuracy.

Assessment data are valuable for purposes beyond their immediate application to estimating accuracy of a land cover map. The reference data is, therefore, made available to other agencies and organizations for use in their own land cover characterization and map accuracy assessments (see Data Availability for access information). The data set will also serve as an important training data source for later updates.

Even though we have reached an endpoint in the mapping process where products are made available to others, the gap analysis process should be considered dynamic. We envision that maps will be refined and updated on a regular schedule. The assessment data will be used to refine GAP maps iteratively by identifying where the land cover map is inaccurate and where more effort is required to bring the maps up to accuracy standards. In addition, the field sampling may identify new classes that were not identified at all during the initial mapping process.

2.6.2 Methods:

Aerial videography transects were flown as shown in Figure 2.2 with a light plane using a high resolution 8mm camcorder system loaned to the Pennsylvania GAP Project by the national GAP coordinators. The system featured dual wide-angle and zoom video cameras with linkage to a Global Position System (GPS) unit though a notebook computer. There were numerous logistical difficulties in obtaining aerial videography. Cloud conditions and otherwise inclement weather were also persistent problems. As a consequence of these difficulties, the available videography came from different years and different seasons. Some represented full foliage phenology, some were from fall foliage transition, and others were from leafless periods.

The strategy for obtaining reference data from the video was to randomly select frames from the respective flight lines on the basis of time code. The selected frames were then located on the basis of time code and the center of the zoom image was classified photointerpretively and recorded. The time code was translated to spatial coordinates on the basis of the GPS data. A circle having 40-meter radius was then located on the map, and pixels having their center in the circle were examined. If the circle captured a pixel having the reference class, then the map was considered to be correct. Otherwise, it was recorded as confusion for the map class of the pixel nearest the center of the circle. The accuracy assessment was performed on the individual pixel map before generalization to a 2-ha level.

The physiognomic land cover categories form a general ordination with respect to infrared brightness on Landsat TM imagery. Confusion of categories adjacent in the ordination was expected to occur more frequently, and was not further investigated. If there was confusion of classes farther apart in the ordination, then the video interpretation was rechecked to make sure there had not been a problem in this regard.

An original sample of 498 frames was drawn randomly by flight line and interpreted, which was found to be primarily deciduous forest. Predominance of this class in the map made it unlikely that other classes would be reasonably represented through additional unrestricted random sampling. Therefore, a supplemental random sample was taken after excluding areas where flight lines passed through deciduous and transitional map classes. Although precluding formal calculation of standard errors for accuracy, it was decided as most reasonable to pool data from the unrestricted and supplemental samples.

Table 2.3 shows the results of pooled accuracy assessment, with video reference categories as columns and map land-cover codes as rows. The last column shows percentage user’s accuracy. The last row shows percentage producer’s accuracy. The lower corner shows overall accuracy as percent of correct classification.

Table 2.3. Accuracy assessment for physiognomic land cover categories. R=reference; M=map. Water (watr), evergreen forest (evrgrn), mixed forest (mixd), deciduous forest (decid), transitional (trans), perennial herbaceous (pherb), annual herbaceous (aherb), barren (bare). Producer’s accuracy (%pac);
user's accuracy (%uac).

2.7 Limitations and Discussion

In spite of having broad physiognomic land-cover classes and more than 2,500 clusters, the accuracy target of 80% was achieved only for the deciduous forest class. Water in the landscape was also classified with 85% producer’s accuracy, but there was commission error recorded which could arise in different ways. Cloud shadows were not consistently distinguishable from water in the clustered image data, and some cases of very deep terrain shading could also appear in water clusters. The majority of the confusion for water, however, was with deciduous forest that is usually spectrally distinct from water. Thus, it seems likely that this apparent confusion arises from forested wetlands that are more evident in the spectral infrared than on video taken with visible wavelengths. Likewise, seasonally wet areas may well have dried in the fall video timeframe. It must also be kept in mind that the videography was taken as much as 6 years later than the satellite imagery.

It is evident that evergreen forest is not consistently distinguished from mixed forest, so the evergreen and mixed forest categories would be better combined as “evergreen component” forests. This would give approximately 70% producer accuracy, but still less than 60% user accuracy. Thus, it appears that separate sets of leaf-on and leaf-off imagery are needed for accurate recognition of forests having an evergreen component.
Otherwise, deciduous forests on heavily shaded aspects tend to be clustered with those having an evergreen component.

It is likewise evident that annual herbaceous vegetation is not consistently separated from perennial herbaceous vegetation. Pooling these two classes would give close to 80% producer accuracy, and 70% user accuracy. A major contributor to remaining confusion lies in the fact that northern Pennsylvania forests in several of the spring Landsat scenes had only partially foliated. In these cases, herbaceous vegetation in the understory was contributing strongly to spectral signatures.

Since large paved areas, quarries, etc. were quite evident in the clustered images, it can be concluded that low accuracy in this class is due largely to mixed-pixel contamination in the clustering.

The clustered images have a good deal more fidelity to landscape pattern than the error matrix might suggest. When display scales between 1:50,000 and 1:100,000 are used for the clustered images, they portray the landscape pattern well enough that they have been popular in Pennsylvania as backdrops for GIS work. Major roads and paved urban surfaces are evident, but more minor roads are lost to mixed-pixel effects. Overlaying GIS layers of roads and streams serves to verify consistency of the landscape pattern in these clustered images.

This verification exercise does, however, underscore the importance of phenology for acquisition and processing of remotely sensed image data in Pennsylvania. The best situation is to have a combination of two phonological conditions. The primary set of imagery would be late spring or early summer when trees and shrubs have a full complement of leaves, but agricultural crops have not yet developed complete coverage of the soil surface. The secondary set would be from late fall after the leaves have been shed, but before snowfal