5

Center For Advanced Spatial Technologies (CAST)THE ARKANSAS GAP ANALYSIS PROJECT

FINAL REPORT

LANDCOVER CLASSIFICATION AND MAPPING

TABLE OF CONTENTS1. INTRODUCTION2. LANDCOVER CLASSIFICATION AND MAPPING2.1. Introduction2.2. Landcover Classification2.3. Methods2.3.1. The Landcover Classification Scheme2.3.2. Imagery Acquisition2.3.3. Map Development2.3.4. Map Editing2.3.5. Aggregation2.3.6. Final Map Editing2.4. Results2.5. Accuracy Assessment2.5.1. Introduction2.5.2. Reserved Assessment Data2.5.3. Rapid Assessment Track (RAT)2.5.4. Assessment Results2.6. Limitations and Discussion3. PREDICTED ANIMAL DISTRIBUTIONS AND SPECIES RICHNESS4. LAND STEWARDSHIP5. ANALYSIS BASED ON STEWARDSHIP AND MANAGEMENT STATUS6. CONCLUSIONS AND MANAGEMENT IMPLICATIONS7. DATA USE AND AVAILABILITY8. LITERATURE CITED9. GLOSSARY10. GLOSSARY OF ACRONYMS11. APPENDICES AND MAPS

2.1. Introduction

Mapping natural landcover requires a higher level of effort than the development of data forvertebrate species, agency ownership, or land management, yet it is no more important for gapanalysis than any other data layer. Generally, the mapping of landcover is done by adopting ordeveloping a landcover classification system, delineating areas of relative homogeneity (basiccartographic "objects"), then labeling those areas using categories defined by the classificationsystem. More detailed attributes of the individual areas are added as more information becomesavailable, and a process of validating both polygon pattern and labels is applied for editing andrevising the map. This is done in an iterative fashion, with the results from one step causingreevaluation of results from another step. For example, the mapping and truthing process may reveal

6

needed corrections to the classification scheme. Finally, an assessment of the overall accuracy of thedata is conducted. Where the database is appropriately maintained, the final assessment of accuracywill 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 analysisrelies on maps of dominant natural landcover types as the most fundamental spatial component ofthe analysis (Scott et. al. 1993) for terrestrial environments. For the purposes of GAP, most of theland 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 theenvironment of a given land area (Whittaker 1965). They also are determinants for overall biologicaldiversity patterns (Franklin 1993, Levin 1981, Noss 1990), and they can be used as a currency forhabitat types in conservation evaluations (Specht 1975, Austin 1991) . As such, dominant vegetationtypes 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 anysingle source of remotely sensed imagery. The Arkansas Gap Analysis Project (AR-GAP) mappedthese patterns at a 1:100,000 scale using combinations of remotely sensed data (e.g., air photos, airvideography, and various transformations of satellite imagery) along with field data and previoussurveys. 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 definebiologically meaningful biogeographic patterns. There may be considerable variation in the floristicsof subcanopy vegetation layers (Natural Community) that are not resolved when mapping at the levelof dominant canopy vegetation types (Natural Community Alliance), and there is a need to addressthis part of the diversity of nature. As information accumulates from field studies on patterns ofvariation in understory layers, it can be attributed to the mapped units of Natural CommunityAlliances.

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) anability to distinguish areas of different actual dominant vegetation; (b) a utility for modelingvertebrate species habitats; (c) a suitability for use within and among biogeographic regions; (d) anapplicability to Landsat Thematic Mapper (TM) imagery for both rendering a base map and fromwhich to extract basic patterns (GAP relies on a wide array of information sources, TM offers aconvenient 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 tothe greatest extent possible; and (f) a capability to fit, both categorically and spatially, withclassifications of other themes such as agricultural and urban environments.

For gap analysis, the system that fits best is provisionally referred to as the Natural LandcoverClassification 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 ofvegetation derived by Mueller-Dombois and Ellenberg (1974), adopted by the United NationsEducational, Scientific, and Cultural Organization (UNESCO 1973) and later modified forapplication to the United States by Driscoll et al. (1983, 1984). The Nature Conservancy and the

7

Natural Heritage Network (Grossman et al. 1994) have been improving upon this system in recentyears with partial funding supplied by GAP. The basic assumptions and definitions for this systemhave been described by Jennings (1993).

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

�At the inception of the AR-GAP, a national vegetation classification was not inexistence. However, communication with TNC indicated that such a framework wasbeing developed. Therefore a committee was created to develop a classification thatwould be based on the same concepts as the projected national classification andtherefore create vegetation units that would be consistent with it. The committeecross-walked existing classifications of Arkansas vegetation and integrated them intoa new classification that consisted of approximately 200 community types (Foti, et. al1994). Since we did not know what level of distinction the national classificationwould use for defining alliances, nor did we know at what level these could bedistinguished by remote sensing, we hierarchically clustered them into 90 and 57more general units. These have now been cross-walked into the nationalclassification (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 NaturalHeritage Programs of the Southeastern States. Chapel Hill, NC 689 pp.). However,the names and alphanumeric codes used in the AR-GAP products are those originallydeveloped 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 specificcriteria. These guidelines incorporate both thematic and spatial map components.

National GAP elected to use the UNESCO alliance division (level 5) to categorize mapped naturalvegetation (Jennings 1993). Within this framework, AR-GAP mapped a total of 37 landcovercategories (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 utilizedautomated mapping (computer processing) techniques to develop the landcover map. Classificationutilized 10 "system corrected" Landsat TM scenes (Figure 2.1.). "System corrected" refers to thecorrections 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 thosescenes range between 1990 and 1993 (Table 2.2.).

8

Table 2.1. Landsat TM sensor characteristics.

DNAB DNAB DNAB DNAB DNAB HTGNELEVAW HTGNELEVAW HTGNELEVAW HTGNELEVAW HTGNELEVAW EGNARLARTCEPS EGNARLARTCEPS EGNARLARTCEPS EGNARLARTCEPS EGNARLARTCEPS )sretem(NOITULOSERDNUORG )sretem(NOITULOSERDNUORG )sretem(NOITULOSERDNUORG )sretem(NOITULOSERDNUORG )sretem(NOITULOSERDNUORG

1 25.0-54.0 )eulb5.0-4.0(neerg-eulB 03

2 06.0-25.0 )neerg6.0-5.0(neerG 03

3 96.0-36.0 )der7.0-6.0(deR 03

4 09.0-67.0 )derarfni-dim&raen0.5-7.0(RIraeN 03

5 57.1-55.1 RI-diM 03

6 05.21-04.01 RIlamrehT 021

7 53.2-80.2 RI-diM 03

2.3.3. Map Development

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

Geometric RegistrationTo 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, firstorder transformation, and nearest neighbor resampling of uncorrected imagery was performed inPCI's GCPWorks. Within this software environment, the operator viewed registered vectors overlaid

RO

W 3

5

PATH 26 PATH 25PATH 24 PATH 23

RO

W 3

5

RO

W 3

6

RO

W 3

6

RO

W 3

7

RO

W 3

7

PATH 23PATH 24PATH 25

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

9

Table 2.2. Landsat TM image acquisition dates.

wor-htapSRW wor-htapSRW wor-htapSRW wor-htapSRW wor-htapSRW ETADYREGAMI ETADYREGAMI ETADYREGAMI ETADYREGAMI ETADYREGAMI

53-62 39-72-40

53-52 29-42-90

53-42 29-30-01

53-32 29-21-01

63-52 09-50-01

63-42 29-91-01

63-32 29-21-01

73-52 09-50-01

73-42 29-91-01

73-32 29-21-01

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

Tasseled cap transformationThere are numerous methods available for enhancing spectral information content of Landsat TMdata. In fact, many enhancements are specifically designed to feature vegetation. After geometriccorrection, a tasseled cap transform (Crist and Cicone 1984) vegetation index was calculated foreach 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 ofempirically derived coefficients (Crist and Cicone 1984). Information present in the 6 original bandsinto was compressed into 3 tasseled cap transform bands: greenness, brightness, and wetness. Alarge amount of image variability is expressed within those three bands. Tasseled cap bands couldalso be directly related to physical scene characteristics (Crist and Cicone 1984). Thetransformation�s emphasis on presence and condition of vegetation helped add interpretive value tothe composite imagery. Since AR-GAP landcover mapping was primarily concerned with naturallandcover, the tasseled cap transform yielded relevant information for input to the classification ofnatural landcover. The procedure also reduced the number of bands subjected to the classification

10

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 SegmentationPrior to classification, each image was stratified by Natural Resource Conservation Service's(NRCS) Arkansas STATSGO (State Soil Geographic Data Base) soil boundary polygons (Figure2.2.). It was hypothesized that 1:250,000 scale STATSGO units would help isolate spectrallyhomogeneous regions within scenes prior to classification (Stewart and Lillesand 1995). It washoped that more spectrally homogeneous units would improve classification by limiting diverseregional environmental variables (topography and land use) over the entire Landsat TM scene. Otherresearch has determined that image segmentation procedures helped increase classificationaccuracies (Bauer et al. 1994). STATSGO polygons also standardized the spectral segmentationprocess. Manual delineation of spectrally homogeneous areas would have added time, cost, andinconsistency to the process (Franklin et. al. 1986). Instead, digitally registered STATSGO data wereready for input to this automated map delineation procedure.

Unsupervised ClassificationUnsupervised 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 determinecluster means based upon a user specified range. Then, each pixel (image value) was assigned to acluster based upon its statistical distance from cluster mean. The output image contained pixelgroupings of spectrally homogenous information. An image realistic tasseled-cap pseudo-color tablewas constructed for spectral clusters to assist the labeling (interpretation) stage of AR-GAPlandcover mapping. At this point, clustered data were transferred from PCI image processingsoftware to GRASS (Geographic Resource Analysis Support System) GIS. Labeling wasaccomplished in GRASS GIS. The image interpreter associated landcover classification labels tonatural spectral groupings that were identified in unsupervised clustering. A final merging of eachclassified 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 DataInformation extraction is the process that correlates and aggregates spectral classes to informationclasses. 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 ofenvironmental and photo identifiable variables such as slope, aspect, elevation, shape, texture, andcolor. Initial landcover labelling was guided using the following existing digital ground informationincluding the USDA Forest Service Southern Forest Inventory and Analysis (SOFIA) plots andContinuous Inventory Stand Condition (CISC) data and US Army Corps of Engineers ComputerizedEnvironmental Resources Data System (CERDS).

· SOFIA (SOuthern Forest Inventory and Analysis) DatabaseThe SOFIA database covered the state in a network of 3200 1 acre (2.5 ha) plots located at theintersection of a 3 mile (4.8 km) grid system (Figure 2.4.). The database is updated on a five yearcycle. The latest update occurred in 1988. Landcover attributes for SOFIA plots were unloaded into

11

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 DataDigital stand data were available for two USDA Forest Service Ranger Districts: Buffalo RangerDistrict (Ozark-St. Francis NF) and Jessieville Ranger District (Ouachita NF) (Figure 2.5.). Standdata were formatted as ARC/INFO GIS data sets in Arkansas State Plane projection. Stand coverageswere 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 RangerDistrict stand data. More USDA Forest Service ranger districts in both Ozark-St. Francis NF andOuachita NF became available in digital form in 1995. Those and other new data were laterincorporated to refine the classification. Those refinements are described in the next section: MapEditing.

Figure 2.4. SOFIA plot distribution.

· CERDS (Computerized Environmental Resource Data System)

12

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

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

State-wide CompositeAR-GAP�s composite landcover map involved producing a single legend of consistent UNESCObased map classification categories from 10 individually classified TM scenes. The number and typeof vegetation categories for each interpreter's final classification differed due to the combination ofchanging vegetation structure across diverse physiographic regions, the extent and quality ofavailable ground-reference data, and interpreter bias. Resultant differences, therefore, required cross-walking individual interpreter classification into one master classification list. For each interpreter, alist of total vegetation categories was compiled. Then, common categories were matched. Uniquecategories were maintained in the master list while some closely related categories were collapsedinto 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 assignedto the patched landcover map with a legend reflecting the UNESCO based natural vegetationclassification system for Arkansas Gap Analysis (Foti et al.1994). Then, the map was reviewed forimmediately apparent thematic errors. Errors attributed to categorical differences across sceneboundaries and mislabeled general landcover categories were corrected by further reclassing.

2.3.4. Map Editing

Map RevisionTwo additional data sets were acquired afterinitial classification. Eight digital USDAForest Service ranger districts and TimberManager Inventory (TMI) plot data from theArkansas Game and Fish Commission wereadded to the ground truth database. DigitalCISC stand data were now available over 10ranger districts (Figure 2.5., Table 2.3.).Those data were used to improve theassignment of spectral classes toinformation classes. CISC data werecompared to the classes in the spectraldiversity image. Area weighting algorithmswere employed (Edwards et al. 1995) torefine information class assignments. TMIdata were correlated to spectral data on thebasis of coincidence of the two data sets. Where inconsistencies in classification were discovered

4

25 10

5

6

6

7

9

3

1

8

Figure 2.5. USDA Forest Service Ranger Districts.

13

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

Categorical Expansion Of The Landcover DataGap analysis has traditionally ignorednonnatural landcover categories such as"Agriculture: Crops", "Agriculture: Pasture","Urban: Residential", and "Urban:Commercial" preferring to group these classessimply as "Agriculture and Urban." It wasobvious that differentiating these classes wouldimprove the utility of the landcover productboth for gap analysis and general usage. Urbancategories were mapped using a combinedgeographic information system and spectralapproach. US Census Bureau TIGER(Topologically Integrated Geographic Encodingand Reference) incorporated boundaries wereused to �mask� or confine spectral analysis.Within these areas, commercial and residentialcategories were mapped on the basis of theirunique spectral character. Because there areoften 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 aroad, and possess a candidate "Urban: Residential" spectral signature.

Separating "Agriculture: Pasture" from "Agriculture: Crops" was accomplished on the basis ofgeneral STATSGO landform regions. Because 1:250,000 scale STATSGO boundaries are generallyintended for regional usage, their spatial precision fluctuates. As a result, these categories represent aless accurate model of their spatial distribution which can be useful in the terrestrial vertebratemapping 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 spatialdiscrimination across the state. Initially, many medium sized streams and ponds were not fullyresolved by the Landsat 30 meter pixel resolution. TIGER water helped provide a more realisticdepiction of water in the state. There may however be some unwanted consequences that areseasonal in nature. For example, water levels in reservoirs are dynamic and therefore their spatialextents 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 overstatewater area at some locations.

2.3.5. Aggregation

To meet the national spatial scale requirement, AR-GAP produced a final 100 ha minimum mappingunit (MMU) landcover product with allowable 40 ha inclusions for special features (e.g., lakes,

regnaR regnaR regnaR regnaR regnaRtcirtsiD

ecivreStseroFADSU ecivreStseroFADSU ecivreStseroFADSU ecivreStseroFADSU ecivreStseroFADSUtseroFlanoitaN

paM paM paM paM paMxednI.oN

BAYOU FNsicnarF.tS-krazO 3

BUFFALO FNsicnarF.tS-krazO 1

COLDSPRINGS FNatihcauO 4

FOURCHE FNatihcauO 5

JESSIEVILLE FNatihcauO 2

MENA FNatihcauO 6

POTEAU FNatihcauO 7

STFRANCIS FNsicnarF.tS-krazO 8

SYLAMORE FNsicnarF.tS-krazO 9

WINONA FNatihcauO 01

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

14

palustrine landcover) from the base 30 m resolution landcover data. Aggregated 100 ha AR-GAPlandcover will provide the basis for broad ecoregional terrestrial vertebrate modeling effortsdescribed 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 andShea (1992) presented a response to "why generalize" through their framework of philosophicalobjectives which included elements related to computational efficiency, application needs, andtheoretical concerns. Lagrange and Ruas (1994) explained that generalization is basically conductedthrough two scenarios according to need: model and cartographic generalization. Modelgeneralization redefines (mapped) data reality into a new (scale dependent) perspective andcartographic generalization alters data for eventual cartographic representation. Li and Su (1995)insisted that model generalization is simply an interpretation of purpose. In their estimation, scalepredominates generalization. Accordingly, model and cartographic generalization are translated totheir simplest form: "After transformation in the scale dimension, spatial reality in a form of digitaldata is generalized (simplified)" ( Li and Su 1995, 46). AR-GAP implemented those concepts byaggregating the 30 m landcover data in the "scale dimension" to 2 ha, 10 ha, 40 ha, and 100 haMMUs and converting the aggregated 100 ha raster data to a vector data structure for subsequent linegeneralization (Figure 2.6.).

Many gap analysis projects are challenged by aggregation of TM derived digital landcover data (with30 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 onapproaches 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. Thisapproach assumes that the data undergoing aggregation were constructed from multispectral remotesensing techniques. First, the method derived a binary similarity cell matrix based upon multispectraland classified image inputs. To overcome memory limitations, the state was divided into sevensubsections. Interfaces were written to the Montana program to derive similarity matrices for thosesections 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 fromeach aggregation pass. The interfaces reclassed only those categories which were present in thesection (then restored the original category numbers at the end of the process), constructed GRASSGIS supporting files, and did other miscellaneous tasks. With these interfaces and 100 megabytes ofavailable RAM, six of the seven sections were processed in one day. Testing was necessary to ensurethat parameters would not exceed memory requirements. At each larger aggregation unit the programwas slower than the previous level, which would be expected. Aggregation levels were 2, 10, 40, and100 ha (Figure 2.6.). On some of the wider (more columns) sections, additional aggregations at the60 and 80 ha level were required to further reduce the number of polygons so that the availablememory was not exceeded. To maximize the potential user base within the state the full range ofMMUs 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 outputof 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

15

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 complexand dependent upon GIS mapping strategies, assumptions, classification structures, and the characterof the phenomena underlying the abstraction.

To move the 100 ha product through the entire "model generalization" process, a second step wasneeded 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 competitiveaggregation of those categories and the natural and agronomic landcover classes. The secondaryaggregation procedure was accomplished with ARC/INFO polygon eliminate. The program mergedpolygons smaller than 40 ha with the adjacent polygon with which they shared the longest commonborder. Based solely on the geometric relationship of neighboring polygons, this aggregationapproach in some instances lead to conversion of water to land. As a result, the desired outcome maynot 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 themotivations for generalization are becoming better defined, it seems that the implications ofgeneralization on modeling and decision making are unclear and require further investigationespecially in the context of spatially and thematically hierarchical data structures in an operationalenvironment (Richardson 1994).

2.3.6. Final Map Editing

Cartographic Adaptation Of The Landcover DataAR-GAP benefited from early lessons in cartographic communication and miscommunication. Theprocess of portraying the landcover database in hardcopy map form was iterative and providedvaluable learning experiences toward production of the final map. Burt (1995: 154) discussed theconcept that "clarity of understanding can be a function of experience" suggesting that, "the moreways an image is experienced the clearer its meaning becomes." For the Arkansas project, experienceand experimentation in design truly played a functional and valuable role in improving cartographic

16

realization of the landcover map.

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

The second iteration of the landcover map totally avoided the difficult issue of linking the mappedimage to the legend. Taking into account our original goal to simply illustrate the data, this mapincluded the mapped categories in the legend, however, colors were not associated with thosecategories. 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 wasrequired to think through this problem.

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

A similar problem was encountered in the process of linking landcover (habitat) to gap analysisterrestrial vertebrate mapping models. To build these relations, an organizational tree was formulatedthat more clearly expressed the hierarchical structure and relationships among mapped landcoverclasses. 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 multiplehierarchies of the UNESCO/TNC based Arkansas landcover classification system. Thus, the treestructure 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 successfullycommunicate the structural composition of the classification system and to facilitate comprehensiblelandcover category to color hue allocation at the formation level. Clearly, the project benefited fromthe dynamic interaction and creativity derived from a multidisciplinary team with diversebackgrounds in geography, anthropology, forestry, biology, and cartography assembled together at thesame location. Keates (1993) has lamented that the cartographic design process is often limited notby technological factors but rather by conceptual barriers like the extensive color choices available tomodern cartographers that potentially obscure other map design alternatives outside the colordomain. Similarly, Kuchler (1967: 143) presented potential ambiguities of vegetation type

17

description in map legends and noted that "it is easy enough for the mind to get into a groove, but itis difficult to get from there into another one." The AR-GAP team's strongly cooperative workingrelationship united many innovative ideas that facilitated cartographic production of the finallandcover 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 ofcolors (and categories) was brought about by the need to generalize for graphical clarity, selection ofcolors was equally critical in maintaining map clarity and consistency. Immediately recognizable arethe 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 urbanfeatures and the surrounding phenomena. Colors that implied additional ecological meaning inlandcover categories were also considered. Palustrine map units, for example, were shown in hues ofblue in an effort to impart the physical reality of a flooding regime (or moisture) to those mappedcategories. This technique conforms to the traditional Gaussenian approach of attributing ecologicalsignificance to color (Kuchler 1967), which has been suggested for small scale (1:200,000 to1:1,000,000) maps (Gaussen 1953). Coincidentally, the color scheme in the AR-GAP landcover mapclosely 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 wasillustrated on a single map sheet using the 2 ha landcover map product at 1:600,000 scale. Marginstatistics were calculated based upon the ungeneralized 30 meter product listing statistics for each ofthe 37 categories although the mapped image presented a total of 15 natural landcover and 5 othercategories. A beige background was used to improve the map's figure to ground relationship and torefine the clarity of accompanying text. Upon completion, the map was mailed to state gap analysiscooperators for comments. One criticism of the map was that the legend used binomial Latinscientific names for the alliance level map units rather than common names. This confirmed ourassumption that the map would be read by a diverse audience without specialist knowledge of thespecific map content. Although these units were not explicitly portrayed on the map, they are part ofthe digital database. A second map was constructed that translated the legend from scientific namesto 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 accuratelyas possible. Doing so, however, requires significant investment in project time and effort.

Line Smoothing and TilingOnce the 30 meter data were aggregated to the 100 ha MMU, the next challenge was to vectorize theraster-based aggregation output to meet the national digital submission standard. The visualcharacter of the resultant polygons were defined by the regular grid of the input raster data thatdefines the raster. Landcover data for Benton Co., Arkansas were selected for testing line smoothingalgorithms both in ARC/INFO and in Intergraph Map Finisher. Different algorithms were initiallycompared for computational efficiency. Map Finisher moving averages method had superiorprocessing speed. That algorithm also appeared to have greater capacity for maintaining the generalform of the original polygons. AR-GAP proceeded to test different input parameters required foroperation of the moving averages algorithm to generate the most aesthetically pleasing output. Itwas determined that equal weighting parameters produced the most effective results.

18

National Gap standards required the final landcover product to be in an ARC/INFO vector fileformat. All previous product development had occurred in the raster domain. Raster to vectorconversion was required. While the output minimum mapping unit remained 100 ha, the baseresolution was 30 meters. The resulting vectorized 100 ha landcover map maintained the rasterlegacy in its very irregular grid shaped pattern. The 30 m resolution created the appearance of a verydetailed and precise boundary shape. In reality, those boundaries (100 ha MMU) are not as detailedas the raster might indicate. AR-GAP considered the representation of these boundaries inconsistentwith the scale at which the data had been aggregated. A more generalized depiction of the data wasrequired to insure scale consistencies at the required 100 ha MMU level in a vector format. Toachieve the desired generalized product, AR-GAP smoothed the derived linework that resulted fromthe 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 whilemaintaining most of the original character of the polygon.

Line smoothing implementation revealed many unforeseen pitfalls. First, moving the data fromGRASS to MGE posed an initial hurdle. Because of its complexity and size (data volume), the entirestate consisting of over 19,000 landcover polygons could not be transferred as a single file. Effortswere made to split the state in half along the Arkansas River, a natural feature that essentially bisectsthe state. Still, the data volume continued to overwhelm the data translation software. Finally, thelandcover 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 tileswere vectorized (r.poly) and then output to MGE ASCII Loader format through v.out.mgal.sh(developed by Malcolm Williamson - CAST). The tiles were individually migrated into MGE via

Texarkana

Idabell

De Queen

Magnolia El Dorado Crossett

CamdenHope Dumas Greenwood

Arkadelphia Malvern De Witt

Mena Lake Ouachita Little Rock Brinkley

Memphis WestFort Smith Russellville Conway Searcy

JonesboroStilwell Fly GapMountain

Mountain View Batesville

Cherokee Village

Blytheville

ParagouldBull ShoalsLake

HarrisonFayetteville

Helena

Clarksdale

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

19

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 werelocated in the adjacent polygon after smoothing. To retain proper polygon attribution, thosecentroids needed to be relocated to their appropriate polygon. The lines were first smoothed, then alist file containing the centroid locations was created. The list file allowed the operator to reviewcentroid placement by cycling through the queued list file. Duplicate centroid locations werereadjusted upon visual inspection when they were found to jeopardize polygon attribution. After therelocation process, topology was built for the tile. Other topological errors, however, still existed inareas characterized by narrow polygon features with few vertices defining that narrow peninsularshape. In these situations, line smoothing would collapse the feature on itself subsequently creating asecond polygon without a centroid These errors were unique since the point at which the linecrossover occurred lacked any spatial connectivity. As a result, they were corrected by applying linecleaning (intersection processor) tools to flag false intersections. Since intersection processor is amemory 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 quadrangleboundaries to surrounding quadrangles. This was a manual process that proceeded in Microstation.To guide edgematching and minimize potentially confusing situations, every polygon centroid waslabeled according to its landcover code. Original vectorized polygons including all details of theoriginal 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 anARC/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 MAPJOINprocess were manually corrected in ARCEDIT. Approximately 20 edgematching error polygonswere removed from the coverage using ELIMINATE with a area threshold of 40,000 square metersyielding a total of 19,748 polygons.

2.4. Results

AR-GAP landcover classification included 36 landcover classes (Table 2.4.) with 31 naturallandcover 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 occupiedmore than 10% of the land area of Arkansas. Dominant natural landcover classes (greater than 5% oftotal 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 - mixedhardwoods; 9.03%), and P.1.B.3.c.VII (Quercus phellos; 5.41%). Sparse landcover classes (less than1,000 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 (Mixedshrub species; 741 ha), P.5.A.4.a.I (Tall grass; 214 ha), and P.5.A.4.b.III (Arundinaria gigantea; 785ha). Minor representation of those classes may be connected in part to combined circumstancesof sparse source data for classification of those categories and differential effects of aggregation ofinitial 30 meter classification to a 100 ha MMU (with allowable inclusions of 40 ha for Palustrineclasses). Changes in area due to aggregation are discussed in the following section.

20

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 inarea with the aggregation to a 100 ha MMU (with allowable inclusions of 40 ha for Palustrineclasses). 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%) andT.1.B.3.a.II (Quercus alba; -3.05%), lost more than one percent. One category, T.2.B.4.a.I (Quercusspp. - Carya texana), disappeared from the map. While very few classes experienced more than a

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

21

one percent gain or loss, even smaller changes attributed to the aggregation procedure appeared tohave a significant effect on the distribution of AR-GAP landcover classes. Twenty two classes haveless 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. Percentchange 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 provideddemonstrative evidence of selective discrimination in the decision to aggregate at a specificthreshold. More work is needed to fully evaluate the outcome of the 100 ha aggregation and explorewhether 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 predictedterrestrial 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 whichmodels and potential management decisions may be based.

Error matrices are one commonly used method for assessment of landcover maps produced byanalysis of remotely-sensed digital multispectral satellite data (Congalton 1988, Lark 1995, Verbylaand Hammond 1995). The error matrix is a two-way contingency table that compares classifiedpixels to a set of independently collected reference data (Janssen and van der Wel 1994). While errormatrices provide one approach to map validation, it should be recognized that there are potentialproblems with this methodology. Lark (1995), for example, indicated that map accuracy may bedescribed by a variety of error rates which change in accordance with specific mapped categories andaccuracy requirements of different map users. The error matrix samples and reports those particularerror rates (probabilities) as determined by map classification system, of sample designmethodology, and spatial data complexity (Congalton 1988). Those map accuracy factors form anintricate interface among remote sensing, cartography, and scale. Puech (1994) demonstrated thatclassification of pixels often reveals local landscape anomalies rather than discrete homogenous mapunits because the physical components described by the remote sensor are fixed by its spatial andspectral resolution characteristics. As a result, Puech (1994) suggested that the error matrix may notbe a sufficient measure by which a classification should be evaluated. Still, some comparisoncriterion of mapped landcover to ground reality is needed and error matrices provide a logical pointto begin that exercise (Congalton 1991).

AR-GAP developed and combined two assessment strategies. First, an opportunistic approach reliedon reserving a portion of the original training data for accuracy assessment (USDA Forest ServiceCISC and AGFC TMI). Second, a cooperative field based data collection effort (Rapid AssessmentTrack (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

22

tree species. Most of the state's physiographic regions are represented by the network of AGFCWildlife Management Areas (WMA). Each WMA is covered with a regular grid of TMI inventoryplots at roughly 100 m intervals. Plots were converted from site to raster data on an aligned 30 mresolution to match the AR-GAP landcover map. Opportunistic assessment of the first-cut landcovermap with AGFC TMI data provided initial area indications of high agreement, low agreement, andareas 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 asDzur et al. 1996. Collecting unbiased independent reference samples across a 13,798,200 ha studyarea such as Arkansas becomes problematic. Because of their size, state-wide natural landcover mapsoffer many difficult logistical issues for systematic thematic accuracy assessments. Among theseissues are limited access to private lands, time, and cost constraints. These broad issues characterizethe problem of implementing an accuracy assessment on a state-wide scale. While those issues arereadily apparent and relatively easy to isolate, the challenge remained to devise available options thataddressed those broad constraints. AR-GAP required an implementation plan - a "Rapid AssessmentTrack" (RAT) - for gap analysis landcover map accuracy assessment that minimized direct datacollection time, costs, and land access limitations.

Perhaps the most influential factor on the magnitude of logistical obstacles lies in the character of thesample design used for accuracy assessment. Janssen and van der Wel (1994) suggested that pooraccessibility and limited budgets may result in a cluster-based sampling approach. While clusteredsampling might reduce the cost to travel between sample units, field crews still needed to visit andrecord sample site information. Field crews for accuracy assessment were solicited from AR-GAPcooperators gathered at the third annual state-wide gap analysis meeting in February of 1995 in LittleRock (Dzur et al. 1996). Here, Arkansas Forestry Commission (AFC) demonstrated a willingness toparticipate 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-developedorganizational structure. Furthermore, AFC provided a knowledgeable and well-equipped (vehicles,etc.) personnel base trained in land navigation and natural landcover recognition. This state-widepersonnel network facilitated simultaneous sample data collection across the entire state, thusminimizing collection time.

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

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

Sample Site SelectionA critical RAT objective was to employ random sampling at clustered locations. This objective

23

framed another compelling reason to involve AFC in the sampling effort. Commanding state-widepresence and authority, AFC field crews gained ready access to lands otherwise inaccessible due tothe geography of land ownership. AFC participation helped enable a random site selection design.

Sample Site ProcessingSample 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 theclustering of random samples. A total of 100 USGS 7.5 minute quadrangles (Figure 2.9.) wererandomly 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 regionwas randomly assigned 10 quadrangles.

Once identified, quadrangles were further subdivided into two road-based strata within which a totalof 20 1-ha sample points per quadrangle were distributed. Each quadrangle was buffered by 360 m toensure that roads on surrounding quadrangles did not influence random site selection. Then, the1:100,000 scale USGS Digital Line Graph (DLG) road network was buffered by 300 m. Within thisbuffer zone, 10 sites were randomly selected. Next, 10 additional sites were systematically locatedadjacent to one another based upon a single randomly placed starting point selected outside the roadsbuffer. These 10 off-road sites were positioned so as not to touch the roads buffer and to lie on a 1km transect oriented north for easy compass location by the ground crew. A GRASS GIS UNIXscript automated random site selection processing for the 100 quadrangles.

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

Field MaterialWhile AFC provided the labor force required to visit those sample sites, CAST needed to ensure aquality data collection effort. Accurately locating sample points in the field was the central concernand most difficult expectation. Since it was impractical to equip AFC with Global PositioningSystem (GPS) equipment, an alternative was to provide AFC crews with clearly delineated anddetailed field map sheets. Designing the most effective field map for AFC personnel requiredintimate understanding of AFC land navigation training guidelines. AFC outlined these guidelines

24

and recommended a ranked priority of routine AFC land reference features for inclusion on fieldmaps. 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 milebarscale 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, mapclassification keys, and disposable cameras.

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

Numerous 1:100,000 scale reference layers were then printed over the raster background. Thesereference layers appeared on both map sets and included: TIGER roads (with varying line weightsand styles for distinct road classes), TIGER railroads, DLG hydrology, DLG contours, and DLGPLSS data (with labelled sections). USGS Geographic Names Information System (GNIS) populatedplace names and their site locations were printed and oriented such that the names would always becontained entirely within the map area. A white dot surrounded by a black ring marked the 20 sitesobtained from the site selection procedure. Sites were given an alphabetical (A-K) letter designationfrom north to south where K always indicated the 1 km transect. Margin information containedbarscale, state locator map, north arrow, quadrangle name and unique number identifier, regionalcoverage (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 mapswas maintained at AFC headquarters in Little Rock, AR. The remaining maps were bundled by AFCdistrict for distribution to district foresters and then to actual field crews. To avoid potential bias indata collection, no explanation or meaning was attributed to the pattern of color portrayed by theLandsat TM background layer. The black-and-white map doubled as field data sheet. On the reverseof 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 wasrecognized that some sites may have been previously inventoried. Therefore, the source columnlisted three data survey alternatives: on-site visit (V), inventory records (I), or personalcommunication (P). At the top of the field data sheet, name and phone number of the field observerwere recorded. Upon RAT completion, black-and-white map/field data sheets were returned toCAST.

Classification Key DevelopmentA UNESCO/TNC based vegetation classification scheme was adopted for AR-GAP landcovermapping (Foti et al. 1994). The UNESCO/TNC classification system is characterized by its

25

Figure 2.10. Color AFC Field Map.

26

Figure 2.11. Black and White AFC Field Map.

hierarchical structure (Jennings 1993). To help field crews match ground data to this classificationsystem a coded semi-dichotomous key was assembled from the classification system (Appendix11.3.). The key was designed to lead the user through a possible 35 main entries with multiplecontrasting subentries defining particular landcover criteria to identify the correct landcover mapclass. Flowing linearly in accordance with the hierarchical structure of the UNESCO basedclassification, the key first delineates class membership at a system level (e.g., Terrestrial, DevelopedCover, Palustrine, Riverine, etc.). The two-page (double-sided) key reads from left to right startingwith landcover code, followed by a code description, the equivalent gap analysis classificationnotation (e.g., T.1.A.9.b.I - Pinus echinata), and a �Go To� column with a number designating thenext 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

27

personnel start with code 1a. As 1a - "primarily water (bottomlands)" does not apply, proceeding to1b - "primarily land" satisfies the present conditions and points to 4. The first entry (4a), "areasdominated by trees with a total canopy cover of 61% or more, tree crowns usually interlocking", fitsthe 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 thesite characteristics. Then, 7a was recorded on the field data sheet. Instructions to field crews were torecord the code about which they had the most confidence.

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

Figure 2.12. Example AFC Field Photo.

RAT Collection ResultsA total of 90 field data sheets (quadrangles) were returned. From those quadrangles, AFC collected atotal of 1777 field sample points. Two quadrangles lacked the off-road points because they were notprinted on the field maps. Only 3 sites were not collected because the landowner would not allow theAFC crew to enter the property. Of the total sample sites, 1601 were actual on-site visits, 156 werebased on inventory records, and 20 were based on personal communication.

28

A total of 23 cameras were returned. From those cameras, 412 photos were developed. Althoughevery roll was not completely developed, most photos appeared very informative. Evenunderexposed photos provided informational clues related to canopy closure. Expected benefitrelative to a total cost of $330 for camera purchase and film developing justified the experimentalcamera concept.

ConclusionsRAT's primary accomplishment was implementing a cooperative state-wide systematic datacollection effort. As noted in the methods section, two major elements guided successful RATimplementation. First, the sample design achieved immediate reductions in data collection timethrough stratification on quadrangles and roads. Second, clear and concise field materials supportedAFC field observations. Field materials were generated to assist sample site location, description,and documentation. Furthermore, field materials formed the vital link between sample design andfield data collection. As potential users of the Arkansas natural landcover map, AFC recognizedlegitimate value in RAT objectives and contributed a proportional commitment to the successfulcompletion of RAT. Although AFC's participation in RAT minimized direct costs to AR-GAP, AFCcosts were not minimal. In fact, Robert McFarland, AFC Assistant State Forester, estimatedmanpower costs alone at one half man year (over $10,000). Still, RAT demonstrated that partnershipsstrengthened by well-developed planning can overcome major impediments to state-wide accuracyassessment 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 thelandcover map was conducted using AGFC Timber Manager Inventory (TMI) digital database, AFCfield data collection, and a 5 percent sample of the 20% reserved USDA Forest Service CISC standdata. Sampling the stand data only served as a necessary data reduction technique to facilitate theconversion of those sites into a manageable reference data set. To boost sample sizes for mostcategories, 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 AssessmentTools.

Accuracy was tested on level 5 (alliance) map categories. Output of the pixel to pixel comparisonswas arranged in a standard error matrix (Appendix 11.6.). Aggregated error matrices were alsoproduced for each level in the AR-GAP classification hierarchy (Appendix 11.18.). To visualizethose results, a set of binary maps (Figures 2.13. through 2.18.) was created for each level. The mapsillustrate the area of the state that met an overall map accuracy criterion of 75 percent. All individualmap category accuracy highlighted in this section are described as the percentage of pixels in eachcategory of the AR-GAP landcover map that are actually that category on the ground (as determinedby 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 matrixdivided by the total number of reference pixels (Senseman et al. 1995). Other measures of accuracymay be evaluated by consulting the error matrices (Appendix 11.7. - 11.8.). Agriculture (crops) andurban areas were unaffected by the classification hierarchy and were mapped respectively at 75% and67% accuracy. Since agricultural landcover constitutes a large portion of Arkansas� land area, mostof the eastern third of the state (known as the Delta) remains in all binary accuracy maps.

29

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. Overallaccuracy 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 greateraccuracy threshold (Figure 2.14.). Mainly evergreen forest was 76% accurate while bottomlandforest maintained 78% accuracy. Mainly deciduous or mixed forest was 67% accurate due to thedifficulty of classifying woodland versus forested areas.

At level 3, woodland and forest are separated in the classification structure. As a result, 64% of thestate was displayed (Figure 2.15.) at 75% or better accuracy. Temperate lowland and submontanebroad-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 estimatedat 49%.

Very little change in accuracy occurred between level 3 and level 4 (formation). Here, 62% of thestate was expressed as being at least 75% accurate (Figure 2.16.). With some minor separation in thebottomland forest classification at level 4, it was discovered that Cold-deciduous alluvial forest was76% 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 detailedlevel 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 - mixedhardwoods; white oak - mixed hardwoods was 79% accurate. Other notable level five categoriesinclude P.1.B.3.c.I, Quercus lyrata, overcup oak mapped 55% accurate and P.1.B.3.c.III, Quercusfalcata 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 orregional analysis. The AR-GAP landcover map was not designed for analyses at scales finer than1:100000. Some landcover classes presented problems for mapping from Landsat TM. Bottomlandhardwoods have similar spectral signatures, occur in similar environments and are often in mixedstands. We used AGFC TMI data to help refine bottomland classification. Landcover is a dynamicfeature that can change rapidly across the landscape which adds to the difficulty of makinggeneralizations 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 andlimited ground truth input. Most forest regeneration areas were classified with Agriculture: Pasture.

Gap analysis uses landcover distribution to define wildlife habitat for predictive terrestrial vertebratemodeling. Gap analysis landcover mapping standards required large area coverage on a clearlydefined time/financial budget. Budgetary elements combined with scale and an objective to producethe most detailed Arkansas landcover map ever promoted an operational atmosphere within aresearch environment. AR-GAP investigated methods for large area vegetation cover mapping.Concentrated efforts were made to incorporate unique research methodologies in the AR-GAPproject. Applied techniques included tasseled-cap transformation of TM data, image segmentation by

30

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 transformedand 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 withunderlying inherent qualities of TM data and data used to interpret spectral clusters.