Center For Advanced Spatial Technologies (CAST)

THE ARKANSAS GAP ANALYSIS PROJECT

FINAL REPORT

LANDCOVER CLASSIFICATION AND MAPPING

TABLE OF CONTENTS

1. INTRODUCTION

2. LANDCOVER CLASSIFICATION AND MAPPING

2.1. Introduction

2.2. Landcover Classification

2.3. Methods

2.3.1. The Landcover Classification Scheme

2.3.2. Imagery Acquisition

2.3.3. Map Development

2.3.4. Map Editing

2.3.5. Aggregation

2.3.6. Final Map Editing

2.4. Results

2.5. Accuracy Assessment

2.5.1. Introduction

2.5.2. Reserved Assessment Data

2.5.3. Rapid Assessment Track (RAT)

2.5.4. Assessment Results

2.6. Limitations and Discussion

3. PREDICTED ANIMAL DISTRIBUTIONS AND SPECIES RICHNESS

4. LAND STEWARDSHIP

5. ANALYSIS BASED ON STEWARDSHIP AND MANAGEMENT STATUS

6. CONCLUSIONS AND MANAGEMENT IMPLICATIONS

7. DATA USE AND AVAILABILITY

8. LITERATURE CITED

9. GLOSSARY

10. GLOSSARY OF ACRONYMS

11. APPENDICES AND MAPS

2.1. Introduction

Mapping natural landcover requires a higher level of effort than the development of data for vertebrate species, agency ownership, or land management, yet it is no more important for gap analysis than any other data layer. Generally, the mapping of landcover is done by adopting or developing a landcover classification system, delineating areas of relative homogeneity (basic cartographic "objects"), then labeling those 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 reevaluation of results from another step. For example, the mapping and truthing process may reveal

needed corrections to the classification scheme. Finally, an assessment of the overall accuracy of the data is conducted. Where the database is appropriately maintained, the final assessment of accuracy will show where improvements should be made in the next update (Stoms 1994).

In its "coarse filter" approach to conservation biology (e.g., Jenkins 1985, Noss 1987), gap analysis relies on maps of dominant natural landcover 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. The Arkansas Gap Analysis Project (AR-GAP) mapped these patterns at a 1:100,000 scale using combinations of remotely sensed data (e.g., air photos, air videography, and various transformations of satellite imagery) along with field data and previous surveys. 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 (Natural Community) that are not resolved when mapping at the level of dominant canopy vegetation types (Natural Community 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 Natural Community Alliances.

2.2. Landcover Classification

Landcover 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 landcover classification system for GAP are: (a) an ability to distinguish areas of different actual dominant vegetation; (b) a utility for modeling vertebrate 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 mesoscale base map in addition to being one source of actual landcover 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 urban environments.

For gap analysis, the system that fits best is provisionally referred to as the Natural Landcover Classification System (NLC). In recent times, this system has also been referred to as the UNESCO/TNC system (Lins and Kleckner 1996) 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).

Tom Foti, Cheif of Research for Arkansas Natural Heritage Commission described our effort to develop a landcover classification system:

"At the inception of the AR-GAP, a national vegetation classification was not in existence. However, communication with TNC indicated that such a framework was being developed. Therefore a committee was created to develop a classification that would be based on the same concepts as the projected national classification and therefore create vegetation units that would be consistent with it. The committee cross-walked existing classifications of Arkansas vegetation and integrated them into a new classification that consisted of approximately 200 community types (Foti, et. al 1994). Since we did not know what level of distinction the national classification would use for defining alliances, nor did we know at what level these could be distinguished by remote sensing, we hierarchically clustered them into 90 and 57 more general units. These have now been cross-walked into the national classification (A.S. Weakley, K.D. Patterson, S. Landaal, and others, compilers. Working draft of March 1998. International classification of ecological communities: terrestrial vegetation of the southeastern United States. The Nature Conservancy, Southeast Regional Office, Southern Conservation Science Department and Natural Heritage Programs of the Southeastern States. Chapel Hill, NC 689 pp.). However, the names and alphanumeric codes used in the AR-GAP products are those originally developed for this project."

2.3. Methods

2.3.1. The Landcover Classification Scheme

According to the national GAP program standards handbook, each map theme must meet specific criteria. These guidelines incorporate both thematic and spatial map components.

National GAP elected to use the UNESCO alliance division (level 5) to categorize mapped natural vegetation (Jennings 1993). Within this framework, AR-GAP mapped a total of 37 landcover categories (32 level 5 classes based on Foti et al. 1994, and 5 other categories).

2.3.2. Imagery Acquisition

Landsat TM data (Table 2.1.) constitutes the base data layer for the landcover map. AR-GAP utilized automated mapping (computer processing) techniques to develop the landcover map. Classification utilized 10 "system corrected" Landsat TM scenes (Figure 2.1.). "System corrected" refers to the corrections performed at the ground receiving station based on previously known sensor (system) distortions such as the pitch, roll, and velocity of the satellite platform. Acquisition dates for those scenes range between 1990 and 1993 (Table 2.2.).

Figure 2.1. Landsat TM WRS (World Refernce System) Path/Row over Arkansas.
Table 2.1. Landsat TM sensor characteristics.

2.3.3. Map Development

Image processing was conducted on a scene by scene basis using EASI/PACE software from PCI Remote Sensing Corp., Arlington, VA.

Geometric Registration

To remove geometric distortions in TM imagery, each scene was geocoded to a UTM (NAD 27) coordinate system based on ground control points (GCP) collected from 1:100,000 scale U.S. Geological Survey (USGS) Digital Line Graph (DLG) roads. Vector to image GCP collection, first order transformation, and nearest neighbor resampling of uncorrected imagery was performed in PCI's GCPWorks. Within this software environment, the operator viewed registered vectors overlaid

Table 2.2. Landsat TM image acquisition dates.

upon imagery. Image identifiable points were selected and matched to vectors until a semi-regular grid of GCP's covered the entire scene. Those GCP's were then used to project the uncorrected imagery into a UTM coordinate system. Each GCP was ordered by the residual error it contributed to the polynomial fit. Points with high error were discarded before registration. Image fit was considered acceptable if the RMS error was < 30 m or one pixel wide (RMS = 1). Transformation was previewed prior to resampling. Each scene was registered using first order, nearest neighbor resampling. Nearest neighbor resampling was selected for several reasons. Computer processing time was shorter compared to other interpolation methods. Further, nearest neighbor interpolation better maintained original gray level values. Also, locating clearly defined image to vector GCP's with low error across an entire TM scene was a difficult task. As a result, first order transformation provided sufficient accuracy for 1:100,000 scale statewide mapping applications and reduced potential introduction of unwanted geometric distortions in areas with no GCP's to provide precise control.

Tasseled cap transformation

There are numerous methods available for enhancing spectral information content of Landsat TM data. In fact, many enhancements are specifically designed to feature vegetation. After geometric correction, a tasseled cap transform (Crist and Cicone 1984) vegetation index was calculated for each scene. The tasseled cap index related six TM bands (1-5 and 7) to measures of vegetation (greenness), soil (brightness), and the interrelationship of soil and canopy moisture (wetness) (Lillesand and Kiefer 1994). The index fit a linear transformation (1) to six TM bands using a set of empirically derived coefficients (Crist and Cicone 1984). Information present in the 6 original bands into was compressed into 3 tasseled cap transform bands: greenness, brightness, and wetness. A large amount of image variability is expressed within those three bands. Tasseled cap bands could also be directly related to physical scene characteristics (Crist and Cicone 1984). The transformation's emphasis on presence and condition of vegetation helped add interpretive value to the composite imagery. Since AR-GAP landcover mapping was primarily concerned with natural landcover, the tasseled cap transform yielded relevant information for input to the classification of natural landcover. The procedure also reduced the number of bands subjected to the classification

process which economized computer processing time.

Eq. 1

Brightness = 0.3037(TM1) + 0.2793(TM2) + 0.4743(TM3) + 0.5585(TM4) + 0.5082(TM5) + 0.1863(TM7)

Greenness = (-0.2848(TM1)) + (-0.2435(TM2)) + (-0.5436(TM3)) + 0.7243(TM4) + 0.0840(TM5) + (-0.1800(TM7))

Wetness = 0.1509(TM1) + 0.1973(TM2) + 0.3279(TM3) + 0.3406(TM4) + (-0.7112(TM5)) + (-0.4572(TM7))

Image Segmentation

Prior to classification, each image was stratified by Natural Resource Conservation Service's (NRCS) Arkansas STATSGO (State Soil Geographic Data Base) soil boundary polygons (Figure 2.2.). It was hypothesized that 1:250,000 scale STATSGO units would help isolate spectrally homogeneous regions within scenes prior to classification (Stewart and Lillesand 1995). It was hoped that more spectrally homogeneous units would improve classification by limiting diverse regional environmental variables (topography and land use) over the entire Landsat TM scene. Other research has determined that image segmentation procedures helped increase classification accuracies (Bauer et al. 1994). STATSGO polygons also standardized the spectral segmentation process. Manual delineation of spectrally homogeneous areas would have added time, cost, and inconsistency to the process (Franklin et. al. 1986). Instead, digitally registered STATSGO data were ready for input to this automated map delineation procedure.

Unsupervised Classification

Unsupervised clustering was then performed on the STATSGO unit using the ISODATA algorithm. In this approach, a number of desired clusters was selected. The image was sampled to determine cluster means based upon a user specified range. Then, each pixel (image value) was assigned to a cluster based upon its statistical distance from cluster mean. The output image contained pixel groupings of spectrally homogenous information. An image realistic tasseled-cap pseudo-color table was constructed for spectral clusters to assist the labeling (interpretation) stage of AR-GAP landcover mapping. At this point, clustered data were transferred from PCI image processing software to GRASS (Geographic Resource Analysis Support System) GIS. Labeling was accomplished in GRASS GIS. The image interpreter associated landcover classification labels to natural spectral groupings that were identified in unsupervised clustering. A final merging of each classified scene produced 2,778 different spectral classes across the state. That composite image (Figure 2.3.) reflects the "Spectral Diversity" of the state of Arkansas.

Information Extraction and Training Data

Information extraction is the process that correlates and aggregates spectral classes to information classes. The information classes, for purposes of gap analysis, were categorized by natural landcover. Spectral classes were compared to existing digital ground information and interpreted in context of environmental and photo identifiable variables such as slope, aspect, elevation, shape, texture, and color. Initial landcover labelling was guided using the following existing digital ground information including the USDA Forest Service Southern Forest Inventory and Analysis (SOFIA) plots and Continuous Inventory Stand Condition (CISC) data and US Army Corps of Engineers Computerized Environmental Resources Data System (CERDS).

  • · SOFIA (SOuthern Forest Inventory and Analysis) Database

The SOFIA database covered the state in a network of 3200 1 acre (2.5 ha) plots located at the intersection of a 3 mile (4.8 km) grid system (Figure 2.4.). The database is updated on a five year cycle. The latest update occurred in 1988. Landcover attributes for SOFIA plots were unloaded into

Figure 2.2. Arkansas STATSGO soil units. Figure 2.3. Arkansas "Spectral Diversity" Image.

INFORMIX database and then transferred to GRASS4.1 as a GIS reference data layer.

· USDA Forest Service Digital Stand Data

Digital stand data were available for two USDA Forest Service Ranger Districts: Buffalo Ranger District (Ozark-St. Francis NF) and Jessieville Ranger District (Ouachita NF) (Figure 2.5.). Stand data were formatted as ARC/INFO GIS data sets in Arkansas State Plane projection. Stand coverages were converted to UTM Zone 15 projection and transferred to GRASS4.1 with landcover attributes. A set of 1982 false-color infrared photography (1:15,840) also complemented Buffalo Ranger District stand data. More USDA Forest Service ranger districts in both Ozark-St. Francis NF and Ouachita NF became available in digital form in 1995. Those and other new data were later incorporated to refine the classification. Those refinements are described in the next section: Map Editing.

Figure 2.4. SOFIA plot distribution.

· CERDS (Computerized Environmental Resource Data System)

CERDS data consisted of hardcopy landcover and aquatic habitat maps for Mississippi River leveed floodplain areas. Maps depict 29 landcover categories compiled from 1982 false-color infrared photography (1:24,000).

Labeling was conducted independently for each classified STATSGO polygon; however, neighboring labeled polygons were considered for within scene edge mapping of STATSGO units. Non-forest cover spectral categories were defined. Then, ground truth data were compared to remaining spectral categories. When all spectral categories were affixed with information labels, spectral classes were reclassed to landcover classes. Each scene was reconstructed by patching together all labeled STATSGO polygons.

State-wide Composite

AR-GAP's composite landcover map involved producing a single legend of consistent UNESCO based map classification categories from 10 individually classified TM scenes. The number and type of vegetation categories for each interpreter's final classification differed due to the combination of changing vegetation structure across diverse physiographic regions, the extent and quality of available ground-reference data, and interpreter bias. Resultant differences, therefore, required cross-walking individual interpreter classification into one master classification list. For each interpreter, a list of total vegetation categories was compiled. Then, common categories were matched. Unique categories were maintained in the master list while some closely related categories were collapsed into others. Once the crosswalk was completed, each scene was reclassed by the crosswalk scheme. After reclassing, scenes were mosaiked with GRASS (r.patch). An arbitrary color table was assigned to the patched landcover map with a legend reflecting the UNESCO based natural vegetation classification system for Arkansas Gap Analysis (Foti et al.1994). Then, the map was reviewed for immediately apparent thematic errors. Errors attributed to categorical differences across scene boundaries and mislabeled general landcover categories were corrected by further reclassing.

2.3.4. Map Editing

Map Revision

Two additional data sets were acquired after initial classification. Eight digital USDA Forest Service ranger districts and Timber Manager Inventory (TMI) plot data from the Arkansas Game and Fish Commission were added to the ground truth database. Digital CISC stand data were now available over 10 ranger districts (Figure 2.5., Table 2.3.). Those data were used to improve the assignment of spectral classes to information classes. CISC data were compared to the classes in the spectral diversity image. Area weighting algorithms were employed (Edwards et al. 1995) to refine information class assignments. TMI data were correlated to spectral data on the

Figure 2.5. USDA Forest Service Ranger Districts.

basis of coincidence of the two data sets. Where inconsistencies in classification were discovered

through the use of those additional data, information classes were updated.

Categorical Expansion Of The Landcover Data

Gap analysis has traditionally ignored nonnatural landcover categories such as "Agriculture: Crops", "Agriculture: Pasture", "Urban: Residential", and "Urban: Commercial" preferring to group these classes simply as "Agriculture and Urban." It was obvious that differentiating these classes would improve the utility of the landcover product both for gap analysis and general usage. Urban categories were mapped using a combined geographic information system and spectral approach. US Census Bureau TIGER (Topologically Integrated Geographic Encoding and Reference) incorporated boundaries were used to "mask" or confine spectral analysis. Within these areas, commercial and residential categories were mapped on the basis of their unique spectral character. Because there are often mixed spectral signatures in residential areas leading to difficulty in differentiating trees surrounding houses and trees without houses, residential zones were further constricted by proximity to TIGER roads. In other words, to receive an "Urban: Residential" designation an area must fall within a TIGER incorporated boundary, be near a road, and possess a candidate "Urban: Residential" spectral signature.

Separating "Agriculture: Pasture" from "Agriculture: Crops" was accomplished on the basis of general STATSGO landform regions. Because 1:250,000 scale STATSGO boundaries are generally intended for regional usage, their spatial precision fluctuates. As a result, these categories represent a less accurate model of their spatial distribution which can be useful in the terrestrial vertebrate mapping process but may be misleading for other applications.

TIGER water was also added (GRASS: r.patch) to AR-GAP's landcover map to improve its spatial discrimination across the state. Initially, many medium sized streams and ponds were not fully resolved by the Landsat 30 meter pixel resolution. TIGER water helped provide a more realistic depiction of water in the state. There may however be some unwanted consequences that are seasonal in nature. For example, water levels in reservoirs are dynamic and therefore their spatial extents may not coincide with other data sources that were interpreted at different times. Similarly, seasonal flooding and management practices in bottomland forests may cause TIGER to overstate water area at some locations.

2.3.5. Aggregation

To meet the national spatial scale requirement, AR-GAP produced a final 100 ha minimum mapping

Table 2.3. Digital CISC data - Map index no. refers to figure 2.5..

unit (MMU) landcover product with allowable 40 ha inclusions for special features (e.g., lakes, palustrine landcover) from the base 30 m resolution landcover data. Aggregated 100 ha AR-GAP landcover will provide the basis for broad ecoregional terrestrial vertebrate modeling efforts described later in chapter 3. Arriving at the 100 ha aggregated dataset was challenging (Stoms 1994).

Many researchers have addressed practical reasons for generalizing mapped data. McMaster and Shea (1992) presented a response to "why generalize" through their framework of philosophical objectives which included elements related to computational efficiency, application needs, and theoretical concerns. Lagrange and Ruas (1994) explained that generalization is basically conducted through two scenarios according to need: model and cartographic generalization. Model generalization redefines (mapped) data reality into a new (scale dependent) perspective and cartographic generalization alters data for eventual cartographic representation. Li and Su (1995) insisted that model generalization is simply an interpretation of purpose. In their estimation, scale predominates generalization. Accordingly, model and cartographic generalization are translated to their simplest form: "After transformation in the scale dimension, spatial reality in a form of digital data is generalized (simplified)" ( Li and Su 1995, 46). AR-GAP implemented those concepts by aggregating the 30 m landcover data in the "scale dimension" to 2 ha, 10 ha, 40 ha, and 100 ha MMUs and converting the aggregated 100 ha raster data to a vector data structure for subsequent line generalization (Figure 2.6.).

Many gap analysis projects are challenged by aggregation of TM derived digital landcover data (with 30 m resolution) to the 40 to 100 hectare minimum mapping unit (MMU) landcover product. Research into rule based automated abstraction methods is still ongoing. In fact, further emphasis on approaches that minimize time spent on algorithm modifications or adjustments is needed (Richardson 1994).

AR-GAP tested and implemented an aggregation method developed by the Montana GAP. This approach assumes that the data undergoing aggregation were constructed from multispectral remote sensing techniques. First, the method derived a binary similarity cell matrix based upon multispectral and classified image inputs. To overcome memory limitations, the state was divided into seven subsections. Interfaces were written to the Montana program to derive similarity matrices for those sections of Arkansas. The Montana program utilized four items to lower memory requirements: number of columns, number of cells, number of categories, and number of output polygons from each aggregation pass. The interfaces reclassed only those categories which were present in the section (then restored the original category numbers at the end of the process), constructed GRASS GIS supporting files, and did other miscellaneous tasks. With these interfaces and 100 megabytes of available RAM, six of the seven sections were processed in one day. Testing was necessary to ensure that parameters would not exceed memory requirements. At each larger aggregation unit the program was slower than the previous level, which would be expected. Aggregation levels were 2, 10, 40, and 100 ha (Figure 2.6.). On some of the wider (more columns) sections, additional aggregations at the 60 and 80 ha level were required to further reduce the number of polygons so that the available memory was not exceeded. To maximize the potential user base within the state the full range of MMUs were maintained toward compilation of the final 100 ha gap analysis landcover database.

While processing speed is an important issue, logical and explainable results are the desired output of any aggregation process. Output from the algorithm appeared visually acceptable. However, evaluating aggregated results to assess potential information loss poses another difficult task.

Remember that any clump of cells will be subsumed and its identity changed if it is not large enough

Figure 2.6. Aggregation process stages 2, 10, 40, 100 ha (left to right).

to remain at the current aggregation level. Clearly, the mechanics of data aggregation are complex and dependent upon GIS mapping strategies, assumptions, classification structures, and the character of the phenomena underlying the abstraction.

To move the 100 ha product through the entire "model generalization" process, a second step was needed to remove small water and urban pixel groupings that were held out of the initial aggregation. Water and urban categories were originally held out of the aggregation to prohibit the competitive aggregation of those categories and the natural and agronomic landcover classes. The secondary aggregation procedure was accomplished with ARC/INFO polygon eliminate. The program merged polygons smaller than 40 ha with the adjacent polygon with which they shared the longest common border. Based solely on the geometric relationship of neighboring polygons, this aggregation approach in some instances lead to conversion of water to land. As a result, the desired outcome may not always be achieved through this type of automated processing for a given user application. Thus, users need to be aware of potential impacts of scale change.

If the goal is to simplify thematic information, we must be aware of implications of scale change. What information is lost or gained? Is the information thematic, spatial, or both? While the motivations for generalization are becoming better defined, it seems that the implications of generalization on modeling and decision making are unclear and require further investigation especially in the context of spatially and thematically hierarchical data structures in an operational environment (Richardson 1994).

2.3.6. Final Map Editing

Cartographic Adaptation Of The Landcover Data

AR-GAP benefited from early lessons in cartographic communication and miscommunication. The process of portraying the landcover database in hardcopy map form was iterative and provided valuable learning experiences toward production of the final map. Burt (1995: 154) discussed the concept that "clarity of understanding can be a function of experience" suggesting that, "the more ways an image is experienced the clearer its meaning becomes." For the Arkansas project, experience

and experimentation in design truly played a functional and valuable role in improving cartographic realization of the landcover map.

The initial goal of our earliest effort was to plainly display the state-wide landcover product as notification to AR-GAP cooperators that project work was being accomplished. That illustration was flawed not because we did not have the data or we did not accomplish our original aim; rather, the failing was in the representation of the data. It should be recognized that a map's ability to carry and convey information can be impaired by primitive color selection, generic language usage in the legend, vague or inadequate notes, and insufficient investment in production time and appropriate cartographic technology. At the statewide scale, portraying community level (level 5) information proved too ambitious. Assigning discernible and logical color hue for 37 landcover categories was troublesome. Instead of communicating with clarity, the map overwhelmed and confused its audience. The resulting experience suggested the need to add a staff cartographer to the project and to carefully consider the many important issues related to cartographic design and presentation of results.

The second iteration of the landcover map totally avoided the difficult issue of linking the mapped image to the legend. Taking into account our original goal to simply illustrate the data, this map included the mapped categories in the legend, however, colors were not associated with those categories. Essentially, the map remained a purely image-based depiction of Arkansas' landcover. Still, that solution neglected the linkage between mapped categories and image. More time was required to think through this problem.

Explicit color hue association with mapped categories required a reduction in the number of categories. The UNESCO/TNC hierarchy provided logical order for that task. While the classification structure was accommodating, the hierarchy itself posed another communication problem. If lower levels (e.g. formation - level 4) in the hierarchy were mapped to reduce complexity in category to color hue assignments, how could the legend indicate that the database contained further detail and hierarchical structure? A visualization tool (legend) was needed to organize and conceptualize this landcover classification structure.

A similar problem was encountered in the process of linking landcover (habitat) to gap analysis terrestrial vertebrate mapping models. To build these relations, an organizational tree was formulated that more clearly expressed the hierarchical structure and relationships among mapped landcover classes. As map producers, the tree supported our understanding of the landcover hierarchy. Obviously, map readers might also benefit from an expanded legend that conveyed the multiple hierarchies of the UNESCO/TNC based Arkansas landcover classification system. Thus, the tree structure was incorporated into the cartographic design of the map legend (Appendix 11.4.).

This legend applied innovation from other aspects of the gap analysis project to successfully communicate the structural composition of the classification system and to facilitate comprehensible landcover category to color hue allocation at the formation level. Clearly, the project benefited from the dynamic interaction and creativity derived from a multidisciplinary team with diverse backgrounds in geography, anthropology, forestry, biology, and cartography assembled together at the same location. Keates (1993) has lamented that the cartographic design process is often limited not by technological factors but rather by conceptual barriers like the extensive color choices available to modern cartographers that potentially obscure other map design alternatives outside the color

domain. Similarly, Kuchler (1967: 143) presented potential ambiguities of vegetation type description in map legends and noted that "it is easy enough for the mind to get into a groove, but it is difficult to get from there into another one." The AR-GAP team's strongly cooperative working relationship united many innovative ideas that facilitated cartographic production of the final landcover map. Our experience would seem to lend credibility to Ferris' (1993, 123) assertion that "maps must not be designed in a vacuum."

Color is a strong determinant in the character of the AR-GAP landcover map. While reduction of colors (and categories) was brought about by the need to generalize for graphical clarity, selection of colors was equally critical in maintaining map clarity and consistency. Immediately recognizable are the solid red and yellow hues symbolizing "Urban: Commercial/Industrial" and "Urban: Residential" respectively. Conceived to draw attention to those relatively small but important reference features, this significant contrast in color hue is consistent with the genuinely distinct character of urban features and the surrounding phenomena. Colors that implied additional ecological meaning in landcover categories were also considered. Palustrine map units, for example, were shown in hues of blue in an effort to impart the physical reality of a flooding regime (or moisture) to those mapped categories. This technique conforms to the traditional Gaussenian approach of attributing ecological significance to color (Kuchler 1967), which has been suggested for small scale (1:200,000 to 1:1,000,000) maps (Gaussen 1953). Coincidentally, the color scheme in the AR-GAP landcover map closely follows the general color design for the area covering Arkansas in Kuchler's (1985) vegetation map which has a color scheme modeled according to Gaussenian color principles.

AR-GAP produced its final gap analysis landcover map of Arkansas during spring of 1996. It was illustrated on a single map sheet using the 2 ha landcover map product at 1:600,000 scale. Margin statistics were calculated based upon the ungeneralized 30 meter product listing statistics for each of the 37 categories although the mapped image presented a total of 15 natural landcover and 5 other categories. A beige background was used to improve the map's figure to ground relationship and to refine the clarity of accompanying text. Upon completion, the map was mailed to state gap analysis cooperators for comments. One criticism of the map was that the legend used binomial Latin scientific names for the alliance level map units rather than common names. This confirmed our assumption that the map would be read by a diverse audience without specialist knowledge of the specific map content. Although these units were not explicitly portrayed on the map, they are part of the digital database. A second map was constructed that translated the legend from scientific names to common names (Moore 1960). In addition, a translation table was mailed to all map recipients. Because maps can have authoritative stature, it is important to formally present the data as accurately as possible. Doing so, however, requires significant investment in project time and effort.

Line Smoothing and Tiling

Once the 30 meter data were aggregated to the 100 ha MMU, the next challenge was to vectorize the raster-based aggregation output to meet the national digital submission standard. The visual character of the resultant polygons were defined by the regular grid of the input raster data that defines the raster. Landcover data for Benton Co., Arkansas were selected for testing line smoothing algorithms both in ARC/INFO and in Intergraph Map Finisher. Different algorithms were initially compared for computational efficiency. Map Finisher moving averages method had superior processing speed. That algorithm also appeared to have greater capacity for maintaining the general form of the original polygons. AR-GAP proceeded to test different input parameters required for operation of the moving averages algorithm to generate the most aesthetically pleasing output. It

was determined that equal weighting parameters produced the most effective results.

National Gap standards required the final landcover product to be in an ARC/INFO vector file format. All previous product development had occurred in the raster domain. Raster to vector conversion was required. While the output minimum mapping unit remained 100 ha, the base resolution was 30 meters. The resulting vectorized 100 ha landcover map maintained the raster legacy in its very irregular grid shaped pattern. The 30 m resolution created the appearance of a very detailed and precise boundary shape. In reality, those boundaries (100 ha MMU) are not as detailed as the raster might indicate. AR-GAP considered the representation of these boundaries inconsistent with the scale at which the data had been aggregated. A more generalized depiction of the data was required to insure scale consistencies at the required 100 ha MMU level in a vector format. To achieve the desired generalized product, AR-GAP smoothed the derived linework that resulted from the vectorization of the raster landcover map. MGE Map Finisher was chosen for the line smoothing. It provided superior results when compared to the original 30 meter raster-based vector linework. AR-GAP employed a weighted moving averages algorithm with equal weights (20) on the first, second, and third vertices. The resulting output vectors achieved a generalized shape while maintaining most of the original character of the polygon.

Line smoothing implementation revealed many unforeseen pitfalls. First, moving the data from GRASS to MGE posed an initial hurdle. Because of its complexity and size (data volume), the entire state consisting of over 19,000 landcover polygons could not be transferred as a single file. Efforts were made to split the state in half along the Arkansas River, a natural feature that essentially bisects the state. Still, the data volume continued to overwhelm the data translation software. Finally, the landcover map was tiled into 35 30' x 60' USGS 1:100,000 scale quadrangles (Figure 2.7.). Unfortunately, the data volume (number of vertices) prohibited direct tiling of the vector file in ARC/INFO (SPLIT). Instead, the data were first tiled in the raster (GRASS) domain. These raster tiles were vectorized (r.poly) and then output to MGE ASCII Loader format through v.out.mgal.sh

Figure 2.7. USGS 30 X 60 Minute Series 1:100,000 Scale Map Tiles.

(developed by Malcolm Williamson - CAST). The tiles were individually migrated into MGE via ASCII Loader. All migrated tiles maintained proper topology through the format conversion process (GRASS to MGE).

Line smoothing introduced topological problems where centroids near polygon boundaries were located in the adjacent polygon after smoothing. To retain proper polygon attribution, those centroids needed to be relocated to their appropriate polygon. The lines were first smoothed, then a list file containing the centroid locations was created. The list file allowed the operator to review centroid placement by cycling through the queued list file. Duplicate centroid locations were readjusted upon visual inspection when they were found to jeopardize polygon attribution. After the relocation process, topology was built for the tile. Other topological errors, however, still existed in areas characterized by narrow polygon features with few vertices defining that narrow peninsular shape. In these situations, line smoothing would collapse the feature on itself subsequently creating a second polygon without a centroid These errors were unique since the point at which the line crossover occurred lacked any spatial connectivity. As a result, they were corrected by applying line cleaning (intersection processor) tools to flag false intersections. Since intersection processor is a memory intensive activity, some of the most complex tiles needed to transferred to a CLIX (UNIX) environment to complete processing.

After line smoothing, all 30' x 60' tiles were edgematched across their 1:100,000 scale quadrangle boundaries to surrounding quadrangles. This was a manual process that proceeded in Microstation. To guide edgematching and minimize potentially confusing situations, every polygon centroid was labeled according to its landcover code. Original vectorized polygons including all details of the original raster boundaries were also displayed as attached reference files to verify line position.

Finally, the tiles were exported to ARC/INFO ungenerate format. Each tile was generated into an ARC/INFO coverage. Topology was built. Thirty five 30' x 60' tiles were patched (MAPJOIN) together into a single 100 ha landcover coverage. Some 200 label errors attributed to the MAPJOIN process were manually corrected in ARCEDIT. Approximately 20 edgematching error polygons were removed from the coverage using ELIMINATE with a area threshold of 40,000 square meters yielding a total of 19,748 polygons.

2.4. Results

AR-GAP landcover classification included 36 landcover classes (Table 2.4.) with 31 natural landcover classes mapped into the hierarchical UNESCO classification (Foti et al. 1994). Agricultural (pasture and crops) landcover ranged over 44% of the state. No other category occupied more than 10% of the land area of Arkansas. Dominant natural landcover classes (greater than 5% of total state area) were T.1.A.9.b.I (Pinus echinata; 7.13%), T.1.A.9.b.II (Pinus taeda; 5.42%), T.1.B.2.b.II (Quercus spp. - Pinus echinata - Carya spp.; 8.08%), T.1.B.3.a.II (Quercus alba - mixed hardwoods; 9.03%), and P.1.B.3.c.VII (Quercus phellos; 5.41%). Sparse landcover classes (less than 1,000 ha) included T.1.B.3.a.I (Fagus grandifolia; 361 ha), T.2.A.2.b.I (Juniperus virginiana - Quercus spp.; 757 ha), T.2.B.3.a.II (Juniperus ashei - Quercus spp.; 944 ha), T.4.B.3.a.II (Mixed shrub species; 741 ha), P.5.A.4.a.I (Tall grass; 214 ha), and P.5.A.4.b.III (Arundinaria gigantea; 785 ha). Minor representation of those classes may be connected in part to combined circumstances

of sparse source data for classification of those categories and differential effects of aggregation of initial 30 meter classification to a 100 ha MMU (with allowable inclusions of 40 ha for Palustrine

classes). Changes in area due to aggregation are discussed in the following section.

2.4.1. Comparison to 30 meter (pre-aggregation) classification

Agriculture increased in area by 4% compared to the 30 meter (pre-aggregation) classification (Appendix 11.3.) causing an increase of over 5,800 km2. Interestingly, only nine classes gained in area with the aggregation to a 100 ha MMU (with allowable inclusions of 40 ha for Palustrine classes). Of those nine, only two classes other than Agriculture, T.1.B.3.a.IV (Quercus falcata - Quercus spp.; +1.15%) and P.1.B.3.c.VII (Quercus phellos; +1.6%), gained more than one percent. Conversely, only two of the 27 classes that lost area, T.1.A.9.b.II (Pinus taeda; -1.32%) and T.1.B.3.a.II (Quercus alba; -3.05%), lost more than one percent. One category, T.2.B.4.a.I (Quercus

Table 2.4. Total area (ha), percent, and number of polygons for AR-GAP landcover classes.

spp. - Carya texana), disappeared from the map. While very few classes experienced more than a one percent gain or loss, even smaller changes attributed to the aggregation procedure appeared to have a significant effect on the distribution of AR-GAP landcover classes. Twenty two classes have less than one percent of the total state area and 18 of those lost over half their original area. In fact, no class with greater than one percent of the area of the state lost more than half of its area. Percent change in area suggested that aggregation to 100 ha MMU caused extreme effects on some classes. T.1.B.3.a.I (Fagus grandifolia; 361 ha), for example, decreased in area from 58 km2 to 3.6 km2. Illustrating the scale-dependent nature of the landcover database, the isolated geography (coves, north facing bluff lines, etc.) and resultant small patch size of Fagus grandifolia provided demonstrative evidence of selective discrimination in the decision to aggregate at a specific threshold. More work is needed to fully evaluate the outcome of the 100 ha aggregation and explore whether or not an allowable 40 ha inclusion MMU for Palustrine features was beneficial.

2.5. Accuracy Assessment

2.5.1. Introduction

Gap analysis combines state-wide Landsat TM derived natural landcover data with predicted terrestrial vertebrate distributions to model biological diversity at ecoregional scales (Scott et al. 1993). Under this scenario, it is important to evaluate accuracy of thematic information, upon which models and potential management decisions may be based.

Error matrices are one commonly used method for assessment of landcover maps produced by analysis of remotely-sensed digital multispectral satellite data (Congalton 1988, Lark 1995, Verbyla and Hammond 1995). The error matrix is a two-way contingency table that compares classified pixels to a set of independently collected reference data (Janssen and van der Wel 1994). While error matrices provide one approach to map validation, it should be recognized that there are potential problems with this methodology. Lark (1995), for example, indicated that map accuracy may be described by a variety of error rates which change in accordance with specific mapped categories and accuracy requirements of different map users. The error matrix samples and reports those particular error rates (probabilities) as determined by map classification system, of sample design methodology, and spatial data complexity (Congalton 1988). Those map accuracy factors form an intricate interface among remote sensing, cartography, and scale. Puech (1994) demonstrated that classification of pixels often reveals local landscape anomalies rather than discrete homogenous map units because the physical components described by the remote sensor are fixed by its spatial and spectral resolution characteristics. As a result, Puech (1994) suggested that the error matrix may not be a sufficient measure by which a classification should be evaluated. Still, some comparison criterion of mapped landcover to ground reality is needed and error matrices provide a logical point to begin that exercise (Congalton 1991).

AR-GAP developed and combined two assessment strategies. First, an opportunistic approach relied on reserving a portion of the original training data for accuracy assessment (USDA Forest Service CISC and AGFC TMI). Second, a cooperative field based data collection effort (Rapid Assessment Track (RAT)) was implemented to gather additional information for the assessment.

2.5.2. Reserved Assessment Data

TMI database contains nearly 15,000 georeferenced points attributed up to three dominant canopy tree species. Most of the state's physiographic regions are represented by the network of AGFC Wildlife Management Areas (WMA). Each WMA is covered with a regular grid of TMI inventory plots at roughly 100 m intervals. Plots were converted from site to raster data on an aligned 30 m resolution to match the AR-GAP landcover map. Opportunistic assessment of the first-cut landcover map with AGFC TMI data provided initial area indications of high agreement, low agreement, and areas lacking enough information for competent assessment.

2.5.3. Rapid Assessment Track (RAT)

Much of the following discussion of this accuracy assessment technique was previously published as Dzur et al. 1996. Collecting unbiased independent reference samples across a 13,798,200 ha study area such as Arkansas becomes problematic. Because of their size, state-wide natural landcover maps offer many difficult logistical issues for systematic thematic accuracy assessments. Among these issues are limited access to private lands, time, and cost constraints. These broad issues characterize the problem of implementing an accuracy assessment on a state-wide scale. While those issues are readily apparent and relatively easy to isolate, the challenge remained to devise available options that addressed those broad constraints. AR-GAP required an implementation plan - a "Rapid Assessment Track" (RAT) - for gap analysis landcover map accuracy assessment that minimized direct data collection time, costs, and land access limitations.

Perhaps the most influential factor on the magnitude of logistical obstacles lies in the character of the sample design used for accuracy assessment. Janssen and van der Wel (1994) suggested that poor accessibility and limited budgets may result in a cluster-based sampling approach. While clustered sampling might reduce the cost to travel between sample units, field crews still needed to visit and record sample site information. Field crews for accuracy assessment were solicited from AR-GAP cooperators gathered at the third annual state-wide gap analysis meeting in February of 1995 in Little Rock (Dzur et al. 1996). Here, Arkansas Forestry Commission (AFC) demonstrated a willingness to participate in the RAT initiative.

As the state agency responsible for the protection and development of Arkansas' forest resources, AFC possessed key qualities needed to implement RAT. AFC provided a well-developed organizational structure. Furthermore, AFC provided a knowledgeable and well-equipped (vehicles, etc.) personnel base trained in land navigation and natural landcover recognition. This state-wide personnel network facilitated simultaneous sample data collection across the entire state, thus minimizing collection time.

The following section presents development of cooperative implementation strategies for systematic accuracy assessment of a natural landcover map of Arkansas. Research objectives described include: automated sample design development and automated field material production directed toward realization of a simple, systematic, and cost effective state-wide accuracy assessment approach that relied upon an established partnership between CAST and AFC.

RAT operational viability resulted from a carefully devised methodology related to sample design and field material composition.

Sample Site Selection

A critical RAT objective was to employ random sampling at clustered locations. This objective framed another compelling reason to involve AFC in the sampling effort. Commanding state-wide presence and authority, AFC field crews gained ready access to lands otherwise inaccessible due to the geography of land ownership. AFC participation helped enable a random site selection design.

Sample Site Processing

Sample point selection methods were adopted and modified from Utah gap analysis (Edwards et al. 1995). AR-GAP spectral classification procedures utilized National Resource Conservation Service (NRCS) State Soil Geographic Data Base (STATSGO) for stratification of spectral space. Accordingly, RAT sampling was stratified by major landforms of Arkansas (Figure 2.8.). U.S. Geological Survey (USGS) 7.5 minute quadrangles served as the first filtering strata for the clustering of random samples. A total of 100 USGS 7.5 minute quadrangles (Figure 2.9.) were randomly selected in proportion to the relative area of Arkansas' 10 major landforms. For example, since the Arkansas River Valley and Ridge landform region comprised 10% of the state, this region was randomly assigned 10 quadrangles.

Once identified, quadrangles were further subdivided into two road-based strata within which a total of 20 1-ha sample points per quadrangle were distributed. Each quadrangle was buffered by 360 m to ensure that roads on surrounding quadrangles did not influence random site selection. Then, the 1:100,000 scale USGS Digital Line Graph (DLG) road network was buffered by 300 m. Within this buffer zone, 10 sites were randomly selected. Next, 10 additional sites were systematically located adjacent to one another based upon a single randomly placed starting point selected outside the roads buffer. These 10 off-road sites were positioned so as not to touch the roads buffer and to lie on a 1 km transect oriented north for easy compass location by the ground crew. A GRASS GIS UNIX script automated random site selection processing for the 100 quadrangles.

Figure 2.8. Major Arkansas Landforms. Figure 2.9. Sample Quadrangle Locations.

Field Material

While AFC provided the labor force required to visit those sample sites, CAST needed to ensure a quality data collection effort. Accurately locating sample points in the field was the central concern and most difficult expectation. Since it was impractical to equip AFC with Global Positioning System (GPS) equipment, an alternative was to provide AFC crews with clearly delineated and detailed field map sheets. Designing the most effective field map for AFC personnel required

intimate understanding of AFC land navigation training guidelines. AFC outlined these guidelines and recommended a ranked priority of routine AFC land reference features for inclusion on field maps. Since the chain (20 m) is the fundamental unit of field measurement used by AFC personnel, AFC requested a chain map scale. In keeping with AFC needs, maps were printed with a mile barscale subdivided by tenths (80 chains = 1 mile). According to AFC, Public Land Survey System (PLSS) line work was an absolute requirement. Once PLSS data became available from USGS, CAST manufactured a field data collection package complete with detailed field maps, map classification keys, and disposable cameras.

Field Map Production

Two sets of 8.5 x 11 field maps (one color and one black and white) were automatically generated for each quadrangle using the Interactive Mapper ( http://www.cast.uark.edu/local/mapper/ ) GRASS script modified to print field map specific features. Distinct color, weight, and style symbology were used to maximize contrast between map features. Separate background raster maps were rendered for color and black and white maps. A Landsat TM unsupervised classification with tasseled-cap (Crist and Cicone 1984) pseudo color table (RGB equals wetness, greenness, brightness) was used as the background raster for color maps. Figure 2.10., for example, depicts the landscapes of the St. Francis National Forest (upper right) on Crowley's Ridge and the Mississippi alluvial valley. Topologically Integrated Geographic Encoding and Referencing System (TIGER) water was used as the background raster layer for black and white maps (Figure 2.11.).

Numerous 1:100,000 scale reference layers were then printed over the raster background. These reference layers appeared on both map sets and included: TIGER roads (with varying line weights and styles for distinct road classes), TIGER railroads, DLG hydrology, DLG contours, and DLG PLSS data (with labelled sections). USGS Geographic Names Information System (GNIS) populated place names and their site locations were printed and oriented such that the names would always be contained entirely within the map area. A white dot surrounded by a black ring marked the 20 sites obtained from the site selection procedure. Sites were given an alphabetical (A-K) letter designation from north to south where K always indicated the 1 km transect. Margin information contained barscale, state locator map, north arrow, quadrangle name and unique number identifier, regional coverage (Latitude/Longitude), source data, and CAST contact information.

A total of 300 (200 color, 100 black and white) maps were delivered to AFC. One set of color maps was maintained at AFC headquarters in Little Rock, AR. The remaining maps were bundled by AFC district for distribution to district foresters and then to actual field crews. To avoid potential bias in

data collection, no explanation or meaning was attributed to the pattern of color portrayed by the Landsat TM background layer. The black-and-white map doubled as field data sheet. On the reverse of this map were 20 data entry rows beginning with sample site A. Data entry columns included: date, landcover code, photo #, source, and comments. While on-site visits were preferred, it was recognized that some sites may have been previously inventoried. Therefore, the source column listed three data survey alternatives: on-site visit (V), inventory records (I), or personal communication (P). At the top of the field data sheet, name and phone number of the field observer were recorded. Upon RAT completion, black-and-white map/field data sheets were returned to CAST.

Classification Key Development

A UNESCO/TNC based vegetation classification scheme was adopted for AR-GAP landcover

mapping (Foti et al. 1994). The UNESCO/TNC classification system is characterized by its

Figure 2.10. Color AFC Field Map.

Figure 2.11. Black and White AFC Field Map.

hierarchical structure (Jennings 1993). To help field crews match ground data to this classification

system a coded semi-dichotomous key was assembled from the classification system (Appendix 11.3.). The key was designed to lead the user through a possible 35 main entries with multiple contrasting subentries defining particular landcover criteria to identify the correct landcover map class. Flowing linearly in accordance with the hierarchical structure of the UNESCO based classification, the key first delineates class membership at a system level (e.g., Terrestrial, Developed Cover, Palustrine, Riverine, etc.). The two-page (double-sided) key reads from left to right starting with landcover code, followed by a code description, the equivalent gap analysis classification notation (e.g., T.1.A.9.b.I - Pinus echinata), and a "Go To" column with a number designating the next step down the hierarchy toward a level five (cover type) classification.

To locate shortleaf pine (Pinus echinata) dominated forest with the key, for example, AFC field personnel start with code 1a. As 1a - "primarily water (bottomlands)" does not apply, proceeding to 1b - "primarily land" satisfies the present conditions and points to 4. The first entry (4a), "areas dominated by trees with a total canopy cover of 61% or more, tree crowns usually interlocking", fits the definition of forest leading the search to 5. Again the first entry (5a), "mainly evergreen forest (greater than 75% evergreen trees)", corresponds to ground evidence and leads to 6. As 6a, "dominated by eastern redcedar (Juniperus virginiana) , does not agree, the next alternative, (6b), "dominated by pines", leads to 7 where 7a, "Shortleaf pine (Pinus echinata) dominant", matches the site characteristics. Then, 7a was recorded on the field data sheet. Instructions to field crews were to record the code about which they had the most confidence.

Site Verification and Pictorial Key Concept

When landcover codes were listed on the field data sheet, a fundamental question remained. What did they look like? Since data were collected by AFC field crews rather than CAST staff, site characteristics were essentially unknown. To supplement the recorded codes, 25 of the 100 sample quadrangles were arbitrarily allocated a disposable Kodak Fun Camera. At least 2 quadrangles in each major landform unit were assigned cameras. For each sample location, field crews took one picture (Figure 2.12.) that best captured the landcover code and recorded the corresponding exposure number in the photo # column. To track camera and quadrangle, a CAST address label printed with quadrangle name and number was attached to each camera. Using RAT as a foundation, these photos might also provide the initial basis for a Web enabled multiscale pictorial key of the AR-GAP classification that would support written descriptions with pictorial examples of landcover map categories using ground based photos, aerial videography frames, and satellite imagery.

Figure 2.12. Example AFC Field Photo.

RAT Collection Results

A total of 90 field data sheets (quadrangles) were returned. From those quadrangles, AFC collected a total of 1777 field sample points. Two quadrangles lacked the off-road points because they were not printed on the field maps. Only 3 sites were not collected because the landowner would not allow the

AFC crew to enter the property. Of the total sample sites, 1601 were actual on-site visits, 156 were based on inventory records, and 20 were based on personal communication.

A total of 23 cameras were returned. From those cameras, 412 photos were developed. Although every roll was not completely developed, most photos appeared very informative. Even underexposed photos provided informational clues related to canopy closure. Expected benefit relative to a total cost of $330 for camera purchase and film developing justified the experimental camera concept.

Conclusions

RAT's primary accomplishment was implementing a cooperative state-wide systematic data collection effort. As noted in the methods section, two major elements guided successful RAT implementation. First, the sample design achieved immediate reductions in data collection time through stratification on quadrangles and roads. Second, clear and concise field materials supported AFC field observations. Field materials were generated to assist sample site location, description, and documentation. Furthermore, field materials formed the vital link between sample design and field data collection. As potential users of the Arkansas natural landcover map, AFC recognized legitimate value in RAT objectives and contributed a proportional commitment to the successful completion of RAT. Although AFC's participation in RAT minimized direct costs to AR-GAP, AFC costs were not minimal. In fact, Robert McFarland, AFC Assistant State Forester, estimated manpower costs alone at one half man year (over $10,000). Still, RAT demonstrated that partnerships strengthened by well-developed planning can overcome major impediments to state-wide accuracy assessment of land access, collection time, and cost.

2.5.4. Assessment Results

Janssen and van der Wel (1994) have suggested that landcover information generated from "per-pixel" classification of Landsat TM be viewed as "point-sampled" data. Accuracy assessment of the landcover map was conducted using AGFC Timber Manager Inventory (TMI) digital database, AFC field data collection, and a 5 percent sample of the 20% reserved USDA Forest Service CISC stand data. Sampling the stand data only served as a necessary data reduction technique to facilitate the conversion of those sites into a manageable reference data set. To boost sample sizes for most categories, all available ground reference data were combined for the presented accuracy assessment. Those data were compared to the 100 ha AR-GAP landcover map using PCI's Accuracy Assessment Tools.

Accuracy was tested on level 5 (alliance) map categories. Output of the pixel to pixel comparisons was arranged in a standard error matrix (Appendix 11.6.). Aggregated error matrices were also produced for each level in the AR-GAP classification hierarchy (Appendix 11.18.). To visualize those results, a set of binary maps (Figures 2.13. through 2.18.) was created for each level. The maps illustrate the area of the state that met an overall map accuracy criterion of 75 percent. All individual map category accuracy highlighted in this section are described as the percentage of pixels in each category of the AR-GAP landcover map that are actually that category on the ground (as determined by the reference data). This metric is known as user's accuracy or map reliability (Congalton 1991). Overall accuracy is computed by calculating the sum of the main diagonal entries in the error matrix divided by the total number of reference pixels (Senseman et al. 1995). Other measures of accuracy may be evaluated by consulting the error matrices (Appendix 11.7. - 11.8.). Agriculture (crops) and urban areas were unaffected by the classification hierarchy and were mapped respectively at 75% and

67% accuracy. Since agricultural landcover constitutes a large portion of Arkansas' land area, most of the eastern third of the state (known as the Delta) remains in all binary accuracy maps.

At level 1, 77% of the state was classified at 75% or greater accuracy (Figure 2.13.). At this level, terrestrial forest was 98% accurate and palustrine forest was determined 78% accurate. Overall accuracy at level 1 was reported at 92%.

For level 2, overall accuracy dropped to 69%. At this level, 50% of the state met the 75% or greater accuracy threshold (Figure 2.14.). Mainly evergreen forest was 76% accurate while bottomland forest maintained 78% accuracy. Mainly deciduous or mixed forest was 67% accurate due to the difficulty of classifying woodland versus forested areas.

At level 3, woodland and forest are separated in the classification structure. As a result, 64% of the state was displayed (Figure 2.15.) at 75% or better accuracy. Temperate lowland and submontane broad-leaved forest was 85% accurate and temperate evergreen needle-leaved forest was 76% accurate. Again, bottomland forest remained unchanged at 78% and overall accuracy was estimated at 49%.

Very little change in accuracy occurred between level 3 and level 4 (formation). Here, 62% of the state was expressed as being at least 75% accurate (Figure 2.16.). With some minor separation in the bottomland forest classification at level 4, it was discovered that Cold-deciduous alluvial forest was 76% accurate. At the formation level (level 4), overall accuracy remained identical to level 3 at 49%.

Level 5 overall accuracy fell to 36% (Figure 2.17.). With increasing detail in classification, interpretation of classification error became more strictly defined and specific. At this most detailed level in the classification hierarchy, 41% of the state was classified at 75% or greater accuracy. T.1.A.9.b.I, Pinus echinata, shortleaf pine was 75% accurate and T.1.B.3.a.II, Quercus alba - mixed hardwoods; white oak - mixed hardwoods was 79% accurate. Other notable level five categories include P.1.B.3.c.I, Quercus lyrata, overcup oak mapped 55% accurate and P.1.B.3.c.III, Quercus falcata var. Pagodifolia, cherrybark oak mapped with 45% accuracy.

2.6. Limitations and Discussion

The coarse nature of these data (1:100000 scale and 100 ha MMU) limit their usage to large area or regional analysis. The AR-GAP landcover map was not designed for analyses at scales finer than 1:100000. Some landcover classes presented problems for mapping from Landsat TM. Bottomland hardwoods have similar spectral signatures, occur in similar environments and are often in mixed stands. We used AGFC TMI data to help refine bottomland classification. Landcover is a dynamic feature that can change rapidly across the landscape which adds to the difficulty of making generalizations and extensions of sparse data over an entire state. In the highlands (Ozark -Ouachita) forest regeneration areas were difficult to distinguish using automated classification techniques and limited ground truth input. Most forest regeneration areas were classified with Agriculture: Pasture.

Gap analysis uses landcover distribution to define wildlife habitat for predictive terrestrial vertebrate modeling. Gap analysis landcover mapping standards required large area coverage on a clearly defined time/financial budget. Budgetary elements combined with scale and an objective to produce the most detailed Arkansas landcover map ever promoted an operational atmosphere within a research environment. AR-GAP investigated methods for large area vegetation cover mapping. Concentrated efforts were made to incorporate unique research methodologies in the AR-GAP

project. Applied techniques included tasseled-cap transformation of TM data, image segmentation by

Figure 2.13. Level 1 Accuracy > 75%. Figure 2.14. Level 2 Accuracy > 75%.

Figure 2.15. Level 3 Accuracy > 75%. Figure 2.16. Level 4 Accuracy > 75%.

Figure 2.17. Level 5 Accuracy > 75%.

a nationally recognized and digitally available database, unsupervised classification of transformed and segmented imagery, and spectral to information class assignment based upon existing ground-reference data. As a result, the derived landcover map is a highly data driven product both with