Mapping the Forest Type and Land Cover of Puerto Rico, aComponent of the Caribbean Biodiversity Hotspot

E. H. HELMER, O. RAMOS, T. DEL M. LÓPEZ, M. QUIÑONES, AND W. DIAZ

International Institute of Tropical Forestry, USDA Forest Service, P.O. Box 25000,Río Piedras, Puerto Rico 00928, USA. ehelmer@fs.fed.us

ABSTRACT.—The Caribbean is one of the world’s centers of biodiversity and endemism. As in similarregions, many of its islands have complex topography, climate and soils, and ecological zones change oversmall areas. A segmented, supervised classification approach using Landsat TM imagery enabled us todevelop the most detailed island-wide map of Puerto Rico’s extremely complex natural vegetation cover.Many Caribbean forest formations that are not spectrally distinct had distributions approximately separableusing climatic zone, geology, elevation, and rainfall. Classification accuracy of 26 land cover and woodyvegetation classes was 71 % overall and 83 % after combining forest successional stages within imagemapping zones. In 1991-92, Puerto Rico had about 364,000 ha of closed forest, which covered 41.6 % of themain island. Unlike previous island-wide mapping, this map better identifies the spatial distributions offorest formations where certain groups of endemic species occur. Approximately 5 % of Puerto Rico’s forestarea is under protection, but the reserve system grossly underrepresents lowland moist, seasonal evergreenforests.

INTRODUCTION

Mapping the distributions of vegetationtypes and human land uses provides criti-cal information for managing landscapes tosustain their biodiversity and the structureand function of their ecosystems. Becausevegetation type can link to species compo-sition or habitat types, vegetation mapsprovide crucial information for biodiver-sity conservation planning. Spatially ex-plicit inventories of vegetation types andland cover permit comparisons betweenparticular vegetation distributions and dis-tributions of land cover, land-cover change,expected climate changes, and protected ar-eas. Although the interpretation of satelliteimagery is useful for mapping vegetationtype (Scott et al., 1993, Muchoney et al.,2000), it poses important challenges tomapping vegetation in regions with com-plex topography and climate.

This study focuses on satellite-imagebased mapping and related challenges forthe complex Caribbean island of PuertoRico. The subject requires attention becausetropical areas with complex topography areparticularly important for biodiversity con-servation. Such tropical mountainous areas

comprised 11 of 25 “biodiversity hotspots”identified by Myers et al. (2000) after evalu-ating species endemism and habitat losscaused by human disturbance. Sevenhotspots included or consisted of complextropical islands. In fact, “almost all tropicalislands fall into one or another hotspot”(Myers et al., 2000), because high speciesendemism combines with proportionallyextensive habitat loss. The Caribbean re-gion is one of the “hottest hotspots,” where11.3 % of the region’s original primary veg-etation contains 2.3 and 2.9 % of the world’sendemic plants and vertebrates, respec-tively (Myers et al., 2000). Unfortunately,“lack of species and ecosystem inventorydata currently hinders development ofbiodiversity conservation strategies, whichare needed at a time of intense develop-ment pressures” (Oldfield and Sheppard,1997). Addressing this problem requiresovercoming the remote sensing and map-ping challenges that occur in the Carrib-bean, where diverse biotic, climatic, andtopographic conditions combine to pro-duce a landscape with varied vegetationcommunities.

Mapping natural vegetation and land

Caribbean Journal of Science, Vol. 38, No. 3-4, 165–183, 2002Copyright 2002 College of Arts and SciencesUniversity of Puerto Rico, Mayaguez

165

use with satellite imagery in complex tropi-cal areas is difficult. First, ecological zonesand illumination angles change over smallareas, leading to spectral confusion inwhich varied vegetation communities havesimilar spectral signatures. Second, theranges of forest successional stages pro-duced by human disturbance compoundspectral confusion (Helmer et al., 2000), buttheir distinction is important for evaluatingecosystem processes. Third, persistentcloud cover often requires combining satel-lite images from different dates, thus fur-ther complicating image-based mapping.

A final challenge to mapping tropical is-land vegetation is that spatial scale andclass resolution become critical. The spatialand class resolutions of recent global andsub-continental ecoregion and land-covermaps (e.g. Dinnerstein et al., 1995; Hansenet al., 2000; Loveland et al., 2000) are toocoarse for biodiversity conservation ontropical islands. For example, while map-ping ecological zones of Latin America andthe Caribbean, Dinnerstein et al. (1995)grouped over 10 major Puerto Rican eco-logical zones into two: moist and dry for-ests. In Puerto Rico, generalizing foresttypes extends the apparent distribution ofendemic species with narrow distributions.For example, the vegetation of serpentinesoils includes many endemic species (e.g,.Garcia, 1991; Cedeño-Maldonado andBreckon, 1996). Existing global- or subcon-tinental-scale maps of Caribbean vegeta-tion zones or land cover do not recognizethese areas as distinct.

The most recent land-cover map of Pu-erto Rico is based on aerial photos taken in1977 and 1978 (Ramos and Lugo, 1994). Themap includes mangrove forests plus fourcover-based woody vegetation classes. TheU.S. National Atlas map of forest covertypes (USDA and USGS, 2001) used thisdata set for Puerto Rico and groups thosecover classes into just one: tropical broad-leaf evergreen forest.

Previous maps for the island have notdepicted land cover in conjunction withforest type beyond cover-based classes.However, ecological zonation systems havedefined potential distributions of vegeta-tion type. Published diagrams of zonation

systems include Puerto Rico Forest Regions(Little and Wadsworth, 1964) and Vegeta-tion Zones (Dansereau, 1966). Publishedmaps of ecological zones available in a geo-graphic information system (GIS) includeHoldridge life zones (Holdridge 1967; Eweland Whitmore, 1973) and Holdridge lifezones superimposed on generalized geol-ogy (Figueroa Colón, 1996). In addition,Areces-Mallea et al. (1999) categorized in-sular Caribbean vegetation types for con-servation efforts based on US Federal Geo-graphic Data Committee (FGDC, 1997)standards. In this hierarchical system, lifeform and percent cover define vegetationclasses, leaf phenology defines subclasses,and leaf morphology defines groups withinsubclasses. Groups contain formations, andwithin formations there are floristically de-fined levels of alliance and association. Al-though some physiognomic factors maydefine vegetation formations, most forma-tions are groups of vegetation units broadlydefined by environmental factors, includ-ing disturbance (FGDC, 1997).

The objectives of our study were to 1)develop and evaluate a practical approachto satellite-image based mapping of foresttype and land cover for Puerto Rico, in away that simultaneously addresses forestsuccession, and 2) illustrate the usefulnessof the resulting map through comparingthe island-wide protected extents of the for-est formations mapped.

MATERIALS AND METHODS

Study area

Puerto Rico’s land area is approximately8900 km2. From northeast to southwest, inthe direction of the trade winds, coastalsemi-deciduous forests occur in bands andpatches along the north and east coasts. Asthree major cordilleras force moisture-carrying trade winds to higher altitudes,moist, submontane and lower montane wetand rain forests occur, including cloud for-est formations. In the southern and west-erly directions, rainfall decreases in the rainshadow of major cordilleras, resulting inmoist followed by increasingly drier forestformations. The largest climatic zone in-

E. H. HELMER ET AL.166

cludes moist broadleaf evergreen forest.Forests have developed over alluvial, sedi-mentary, volcanic, limestone, and serpen-tine substrates.

Summary of Approach

We used the vegetation type classifica-tion given by Areces-Mallea et al (1999).First, we hierarchically related its woodyvegetation types to environmental factor-based ecological zones and rainfall. Thisstep enabled us to identify satellite-imagemapping zones that corresponded withgroups of woody formations and landcover and that were necessary for mappingforest successional stages (Helmer et al.,2000). Reference observations of thesewoody vegetation formations and landcover provided data for building separatediscriminant function models that classi-fied satellite imagery from each mappingzone. This approach limited labeling ofwoody vegetation to one of a pre-deter-mined group of 2 to 5 formations within theboundaries imposed by the ancillary datafor each image mapping zone. Because for-est formations were usually distinguishablein aerial photos, however, a randomizedaccuracy assessment permitted estimationof errors including those that this limitationintroduced. A summary of protected ex-tents of forest formations also evaluatedwhether the image segmentation approachusefully mapped the complex vegetationformations on the island.

Vegetation classification system

In seeking a vegetation classification sys-tem appropriate to the island-wide scale ofour analysis, and that would use Landsatimagery, we adapted the hierarchical veg-etation classification system of Areces-Mallea et al. (1999). Since resources wereinadequate to collect sufficient field-derived reference observations to map flo-ristically-defined vegetation types, wesought to map woody vegetation to its for-mation level. Aerial photos could supportreference data for formation-level map-ping. Environmental factors, including dis-turbance, largely define these formations,

and most defining environmental factorswere directly or indirectly identifiable inaerial photos and often in satellite imagery.The formations in Areces-Mallea et al.(1999) related hierarchically to a simplifica-tion of Figueroa Colon’s (1996) GIS overlayof life zone (Holdridge, 1967; Ewel andWhitmore, 1973) and generalized geology(Krushensky, 1995) (Table 1). Aggregatingthe geoclimatic zones within a GIS resultedin the satellite-image mapping zones in Fig-ure 1, which would help overcome themapping challenge of spectral confusionbetween vegetation formations and landcovers with similar spectral signatures.Mapping to the formation level also al-lowed us to add environmentally-definedformations that Areces-Mallea et al. (1999)did not include, such as serpentine forma-tions and two forest successional stages forseveral geoclimatic conditions. Moreover,its hierarchical structure permitted later as-signment of each class to the more generalclassification systems used in the globalForest Resources Assessment (FRA) (FAO,1996, 1998) and Global Observation of For-est Cover (GOFC) programs (GOFC DesignTeam, 1999).

We aggregated Figueroa Colon’s geocli-matic zones, as detailed in Table 1, if theycontained the same group of forest forma-tions and land covers. For example, sedi-mentary-moist, intrusive-volcanic, and ex-trusive-volcanic moist zones formed oneimage mapping zone because they con-tained mainly seasonal evergreen forestformations and similar land uses. We alsoaggregated wet, rain, and lower montanewet and rain forest zones on alluvial, sedi-mentary, and volcanic substrate under theassumption that satellite imagery plus digi-tal elevation data could more precisely de-lineate the distributions of cloud forest for-mations. Modifications to the geoclimaticzones included subsetting by rainfall (usingCalvesbert, 1970) the moist zone on karstsubstrate in the northwest. Most of thenorthern karst vegetation falls in the moistlife zone, but forest vegetation graduallydries toward the coast to the point wheresemi-deciduous forest dominates karst sub-strates instead of the semi-deciduous andseasonal evergreen mosaic of more humid

MAPPING FOREST TYPES AND LAND COVER 167

TABLE 1. Satellite image mapping zones for Puerto Rico and associated woody vegetation formations.

Satellite image mapping zone1,2 Woody vegetation formations2

Dry forest—Alluvial Lowland dry semi-deciduous forest or woodland/shrubland

Tidally and semi-permanently flooded evergreensclerophyllous forest

Dry forest3—Volcanic, Sedimentary, Limestone Lowland dry semi-deciduous forest or woodland/shrubland

Lowland dry mixed evergreen drought-deciduousshrubland with succulents

Dry and moist forest—Serpentine Lowland dry and moist, mixed seasonal evergreensclerophyllous forest with succulents

Moist forest—Alluvial Lowland moist evergreen hemi-sclerophyllousshrubland

Lowland moist seasonal evergreen forest or forest/shrub

Lowland moist coconut palm forestSeasonally flooded evergreen forestTidally and semi-permanently flooded evergreen

sclerophyllous forestMoist forest—Volcanic and Sedimentary Lowland moist seasonal evergreen forest or forest/

shrubLowland moist semi deciduous forest4

Moist forest with rainfall <1500 mm yr−1—NorthernLimestone5

Lowland moist semi-deciduous forest or forest/shrub

Moist forest with rainfall >1500 mm yr−1—NorthernLimestone5

Lowland moist and wet, seasonal evergreen andsemi-deciduous forest or forest/shrub

Wet and lower montane wet forest—Serpentine Submontane and lower montane wet evergreensclerophyllous forest or forest/shrub6

Wet and rain forest, lower montane wet and rainforest—Volcanic, Sedimentary and Alluvial

Submontane wet evergreen forestActive sun/shade coffee, submontane/lower mon-tane wet evergreen forest/shrub, other agricultureSubmontane/lower montane wet evergreen forest/

shrub, active/abandoned shade coffeeLower montane wet evergreen forest7—tall cloud for-

estLower montane wet evergreen forest7—palm and

elfin cloud forestLower montane wet evergreen forest—elfin cloud

forest

1Aggregated from Geoclimatic Zones in Figueoa Colón (1996), which overlay Holdridge life zones (Ewel andWhitmore 1973) onto generalized geology (Krushensky, unpubl.). Volcanic refers to intrusive/plutonic andextrusive/volcanoclastic geology.

2Forests are subtropical sensu Holdridge (1967) and broadleaf unless otherwise indicated; lowland refers toforests from 0 to 400 m elevation. Both forest/shrub and woodland/shrubland refer to stands with a) 25-60%cover of trees with distinct canopies and an under story of shrubs, seedlings or saplings, or b) dense shrubs,seedlings or saplings, as indicated by a matrix of woody vegetation or a smooth canopy.

3The Dry-Volcanic/Sedimentary/Limestone Zone included southern limestone areas in the drier part of themoist forest zone.

4Coastal areas in southeastern Puerto Rico.5Northern Limestone refers to limestone areas north of the Central Cordillera with well-developed karst

topography and areas at the Cordillera’s southern edge.6Includes forest in the rain forest zone sensu Holdridge (1967).7Includes forest in the lower montane rain forest zone sensu Holdridge (1967).

E. H. HELMER ET AL.168

FIG. 1. Satellite image mapping zones used for segmenting classification.

MA

PPING

FOR

EST

TY

PES

AN

DL

AN

DC

OV

ER

169

TABLE 2. Relationships between woody vegetation formations mapped in this study, formations in Areces-Mallea et al. (1999), forest zones in Dansereau (1966), and forest regions in Little and Wadsworth (1964).

Woody vegetation formation,this study

Woody vegetation formation,Areces-Mallea et al. (1999)1

Vegetation zone2

Dansereau (1966)

Forest region3

Little andWadsworth

(1964)

Forest and shrubland—Dry andDry/Moist

Lowland dry semi-deciduousforest or woodland/shrub-land4

Lowland semi-deciduous wood-land; Lowland and Lowland/submontane drought-decidu-ous woodland

SDF DCF, DLF

Lowland dry mixed evergreendrough-deciduous shrub-land with succulents4

Mixed evergreen drought-de-ciduous shrubland with succu-lents

SDF DLF

Lowland dry and moist,mixed seasonal evergreensclerophyllous forest withsucculents5

— SDF LCF

Forest and shrubland—Moist andMoist/Wet

Lowland moist seasonal ever-green forest or forest/shrub6

Lowland seasonal evergreen for-est

RF MCF, (low elev.LCF, LLF)

Lowland moist semi-decidu-ous forest or forest/shrub

Lowland semi-deciduous forest;Evergreen shrubland

SEF, HSC, LRF MLF, MCF

Lowland moist coconut palmforest

Lowland seasonal evergreen for-est developed from Cocos nu-cifera woodland plantation

RF MCF

Lowland moist evergreenhemi-sclerophyllous shrub-land

Evergreen hemi-sclerophyllousshrubland

LIT MCF

Lowland moist and wet, sea-sonal evergreen and semi-deciduous forest or forest/shrub7

Lowland semi-deciduous forest;Lowland seasonal evergreenforest; Evergreen shrubland

SDF, HSC MLF

Forest—Wet, Rain, Lower montaneWet/Rain

Submontane and lower mon-tane wet evergreen sclero-phyllous forest with succu-lents8

— LMRF, MF LCF

1Formations based in part on Dansereau (1966), and all are broadleaved unless otherwise indicated. Forest/shrub formations would be termed as disturbed (Areces-Mallea, pers. comm.). Dash (—) indicates formationsnot in Areces-Mallea et al. (1999).

2LIT—Littoral, RF—Lowland rainforest, SEF—Seasonal evergreen forest, HSC—Hill scrub, SDF—Semi-deciduous forest, LMRF—Lower montane rainforest, MF—Montane forest, MSC—Montane scrub.

3MCF—Moist coastal forest, MLF—Moist limestone forest, DCF—Dry coastal forest, DLF—Dry limestoneforest, LCF—Lower cordillera forest, UCF—Upper cordillera forest, LLF—Lower Luquillo forest, ULF—UpperLuquillo forest.

4Also see Murphy et al. (1995).5Also see Garcia (1991) and Ewel and Whitmore (1973).6Also see Aide et al. (1996), and Chinea (2002).7Also see Chinea (1980), and Rivera and Aide (1998).8Also see Caminero Rodriguez (1991), Cedeño-Maldonado (1997), Cedeño-Maldonado and Breckon (1996),

and Ewel and Whitmore (1973).9Also see Ewel and Whitmore (1973), Scatena and Lugo (1995).10Also see Weaver (1995).

E. H. HELMER ET AL.170

regions. We have seen no attempt to mapthis feature of the landscape. A secondmodification was extending the dry semi-deciduous zone in the southeast to includeareas east of Guyama (based on Little andWadsworth, 1964).

As mentioned above, we added fivewoody vegetation formations that structur-ally defined a second, younger forest suc-cessional stage, within all but the dry/moist serpentine and alluvial satellite-image mapping zones. Disturbance timingand intensity drive these younger forma-tions. We defined these stands as havingeither 25-60 % cover of trees with crownsvisibly distinct in aerial photos that co-dominate with shorter woody vegetation,

or young, dense stands of shrubs, seedlingsor saplings (as indicated by a continuousbut smooth canopy with more than ap-proximately 40 % cover). Different standstructures within this range are difficult todistinguish in satellite imagery becausetheir spectral signatures are often similar(Foody et al., 1996; Steininger, 1996; Helmeret al., 2000). We used the term forest/shrub(or woodland/shrubland within the drymapping zones) to identify these forma-tions. For example, we termed youngerstands in moist zones as lowland moist sea-sonal evergreen forest/shrub. By contrast,stands with more than 60 % cover of treeswith canopies distinct in aerial photos weretermed lowland moist seasonal evergreen

TABLE 2. Continued.

Woody vegetation formation,this study

Woody vegetation formation,Areces-Mallea et al. (1999)1

Vegetation zone2

Dansereau (1966)

Forest region3

Little andWadsworth

(1964)

Submontane evergreen wet andrain forest9

Submontane rainforest LMRF LLF, LCF

Active/abandoned sun andshade coffee; submontane/lower montane wet forest/shrub; other agriculture9

Inga vera-Erythrina poeppigiana/Cafea arabica forest; commercialplantations

LMRF, MF LLF, ULF LCF,UCF

Lower montane wet evergreenforest—tall cloud forest10

Montane rainforest MF ULF, UCF

Lower montane wet evergreenforest—palm and elfin cloudforest10

Montane rainforest (palm);Cloud forest (elfin); Evergreenshrubland (highest peaks)

MF, MSC ULF, UCF

Forest—FloodedTidally and semi-permanently

flooded evergreen sclero-phyllous forest (Mangrove)

Tidally (Mangrove) and semi-permanently flooded ever-green sclerophyllous forest

LIT DCF, MCF

Seasonally flooded evergreenforest (Pterocarpus officinalis)

Seasonally flooded rainforest(Pterocarpus officinalis)

LIT MCF

1Formations based in part on Dansereau (1966), and all are broadleaved unless otherwise indicated. Forest/shrub formations would be termed as disturbed (Areces-Mallea, pers. comm.). Dash (—) indicates formationsnot in Areces-Mallea et al. (1999).

2LIT—Littoral, RF—Lowland rainforest, SEF—Seasonal evergreen forest, HSC—Hill scrub, SDF—Semi-deciduous forest, LMRF—Lower montane rainforest, MF—Montane forest, MSC—Montane scrub.

3MCF—Moist coastal forest, MLF—Moist limestone forest, DCF—Dry coastal forest, DLF—Dry limestoneforest, LCF—Lower cordillera forest, UCF—Upper cordillera forest, LLF—Lower Luquillo forest, ULF—UpperLuquillo forest.

4Also see Murphy et al. (1995).5Also see Garcia (1991) and Ewel and Whitmore (1973).6Also see Aide et al. (1996), and Chinea (2002).7Also see Chinea (1980), and Rivera and Aide (1998).8Also see Caminero Rodriguez (1991), Cedeño-Maldonado (1997), Cedeño-Maldonado and Breckon (1996),

and Ewel and Whitmore (1973).9Also see Ewel and Whitmore (1973), Scatena and Lugo (1995).10Also see Weaver (1995).

MAPPING FOREST TYPES AND LAND COVER 171

forest. Although the FGDC system classi-fies woody vegetation with 25-60 % treecanopy cover as woodland, we used theterm forest/shrub for younger stands inhumid zones because 1) shrub species canco-dominate with tree species, and 2) moststands will develop closed canopies. Fordry forests we used the term woodland/shrubland for stands with structures simi-lar to those described for forest/shrub; wefelt that this term described more appropri-ately the structural range of younger standsin the dry region because grazing or burn-ing can maintain the zone’s woody vegeta-tion in a relatively open or shrub-domi-nated state.

We added Holdridge life zones to forma-tion names because they played an impor-tant role in defining woody vegetation for-mations. Following Areces-Mallea et al.(1999), we also added the term lowland toformations occurring in the Holdridgebasal altitudinal zone (basal is not usuallyincluded in life zone names), and we addedthe term submontane to formations occur-ring in wet life zones. To facilitate interpre-tation of the woody vegetation formationslisted in Table 1, we related them to thosein Areces-Mallea et al. (1999) and to twoecologial zonation systems for the island(Table 2). For floristic information on theseformations, readers should consult Areces-Mallea et al. (1999) and the bibliographythey used to build their classification. Inaddition, Little and Wadsworth (1964)compare Upper Cordillera with Upper Lu-quillo forest (Table 2). Floristic informationfor the formations that we added, whichAreces-Mallea et al. (1999) do not describe,is available from Chinea (1980), Aide et al.(1996), Rivera and Aide (1998), and Chinea(2002) for moist and wet secondary forestsover limestone and volcanic substrates.Garcia (1991), Caminero-Rodriguez (1991),Cedeño-Maldonado and Breckon (1996),Cedeño-Maldonado (1997), and Ewel andWhitmore (1973) contain additional infor-mation on forests developed over serpen-tine substrates.

Image data

We used Landsat TM imagery dated 7February 1991, 24 December 1991 (both

Path 4/Rows 47-48), and 19 August 1992(Path 5/Rows 47-48). Georeferencing in-cluded terrain correction and pixel re-sampling to 30 m resolution. Transformingall images to the TM Tasseled Cap bright-ness, greenness, and wetness indices (Cristet al., 1986) enhanced visual interpretationof the imagery because these bands haveecologically meaningful interpretations,can be viewed in three-dimensional colorspace, and minimize topographic effects(Cohen et al., 1995). They derive from aguided principle components analysis thatcondenses the six visible bands in LandsatTM imagery into three bands. Furthermore,Helmer et al. (2000) showed that the TMwetness band enhances distinction of tropi-cal forest successional stage in complextropical regions. Reference data derivedfrom photo-interpreting NASA high defini-tion Aerochrome-IR photos (scale 1:32,000;dated January and February of 1991) andfrom field visits made in 1999 and 2000.

Segmented approach to image interpretation

Geographic data ancillary to satellite im-agery alleviates spectral confusion whenmapping natural vegetation types (Strahler,1981; Skidmore, 1989; Ma et al., 2001), landcover (Vogelmann et al., 1998), or forestsuccessional stage (Helmer, 2000). Wechose a segmented approach to image in-terpretation to help overcome the challengeof similar spectral signatures among differ-ent vegetation formations and land cover.In addition, work in montane Costa Ricahad indicated that information on ecologi-cal zones would benefit efforts to addressforest successional stages. After collectingreference observations for each image map-ping zone, we segmented Landsat TMscenes into the image mapping zones inFig. 1, and we separately built and appliedwithin each mapping zone a discriminantfunction model that assigned a woody veg-etation formation or land cover to each ofthe satellite image pixels within that zone(commonly referred to as a supervisedmaximum likelihood classification). Thisapproach limited labeling of woody vegeta-tion formation within each zone to one ofthe major formations within that zone

E. H. HELMER ET AL.172

(Table 1). Digital elevation data with 30-mresolution (Gesh et al., 2002) served as afourth data band in the image mappingzone combining submontane with lowermontane zones.

We made many reconnaissance surveysduring 1999-2000, before the developmentof reference observations (i.e., training datafor building satellite image classificationmodels). These surveys and extensive con-sultation with experts on Puerto Rico’s for-est vegetation (John Francis, J. Danilo Chi-nea, Juan Ramirez, Julio Figueroa-Colón,Sandra Molina, Ariel Lugo, and Fred Sca-tena) enabled us to recognize various forestformations in the aerial photos. Forest for-mations were usually clearly distinguish-able in the photos, and often also in thesatellite imagery, because their geology,canopy texture, color tone, and other at-tributes were recognizable. Exceptions oc-curred in transitional areas between low-land moist seasonal evergreen forest andsubmontane wet evergreen forest, and intransitional areas between lowland dry/moist semi-deciduous forest and lowlandmoist seasonal evergreen forest. Field sur-veys frequently focused on these transi-tional areas.

Training data for building satellite imageclassification models derived from aerialphotos and from our reconnaissance sur-veys. They consisted of 70-100 aerial photo-derived point observations for woody veg-etation formation and land cover withineach zone. Training data included high-density urban areas, sunlit greened-up pas-ture, sunlit senescent pasture, shadowedpasture, and sunlit and shadowed woodyvegetation formations within each zone.For example, in the higher elevation imagemapping zone, we collected separate sets ofreference data for submontane wet ever-green forest, three classes of lower montanewet (cloud) forest, and two spectral classesthat represented mixtures of coffee cultiva-tion (Coffea sp.) and submontane and lowermontane wet forest/shrub. We did not in-clude training data for coffee cultivation inmoist areas because currently it is not ex-tensive there. Another example is the moistalluvial zone, in which training data distin-guished between lowland moist seasonal

evergreen forest, mangrove, seasonallyflooded forest dominated by swamp blood-wood (Pterocarpus officinallis), and aban-doned closed-forest coconut palm planta-tions. Except for dry zones, pasturesincluded urban grasslands and could haveup to 25 % tree cover or about 40 % shrubcover as long as they were apparently ac-tive (as indicated by a matrix of grass).Grass-dominated lands with up to about 25% tree cover are generally difficult to reli-ably distinguish in satellite imagery fromgrass-dominated areas with no trees. In thedry zones, pastures with up to 60 % woodyvegetation cover can still be active but areindistinct in aerial photos and satellite im-agery from recovering vegetation. Conse-quently, the forest formation of lowlanddry semi-deciduous woodland/shrublandincludes pastures, which still likely un-dergo grazing, that have 25-60 % woodyvegetation cover (Table 1).

Manual editing and interpretation

Another challenge in mapping complextropical areas is that the spectral signaturesof several agriculture classes are often simi-lar to forest or urban areas. This occurred inPuerto Rico even after segmenting the im-agery. Consequently, we manually inter-preted the satellite imagery (using aerialphotos as a guide) for sugar cane, plantain,banana, mango, and citrus plantations.Similar manual interpretation included re-cently burned and bare agricultural fields,which were spectrally similar to urban andbulldozed areas; emergent wetlands, whichhad highly variable spectral signatures; co-conut palm plantations; and pineappleplantations. Manual recoding was also nec-essary for scattered moist semi-deciduousforest stands along the southeast coast andfor confusion between urban and pastureareas. In the dry forest zone, we manuallycorrected for confusion between pasturewith less than 25 % tree or shrub cover andmixed drought-deciduous shrubland withsucculents. Finally, we manually inter-preted the few stands of lowland moisthemi-sclerophyllous shrubland that werelarge enough to be clearly visible in the im-agery. The manually interpreted agricul-

MAPPING FOREST TYPES AND LAND COVER 173

ture formed less than 6 % of the total areamapped, while all other manually inter-preted or recoded areas combined formed1.6 % of the total mapped area.

Cloudy areas

Clouds cause difficulties in satellite-image based mapping of complex tropicalareas. We used unsupervised clustering toseparate clouds, cloud shadows, and waterfrom the remainder of each Landsat TMscene. Successive reclustering of confusedclasses, followed by combining subclassesof these elements, produced contiguous ar-eas of clouds and cloud shadows. We thenfiltered contiguous groups of less than 11pixels, which usually did not representcloud elements or water (ERDAS, 1999).Adding a 5-pixel buffer around the remain-ing contiguous cloud elements allowedmasking of cloud and cloud shadow edgesthat remained confused with non-cloud el-ements.

Following separate classifications of theimage mapping zones in each eastern Pu-erto Rico scene, cloud-free portions of theclassified February scene replaced cloud-covered areas in the December scene. Sub-sequent mosaics of image mapping zonesand scenes resulted in an island-wide map.For remaining areas of cloud elements inthe imagery (less than 5 % of the total maparea), we overlaid data from the 1977-78land-cover map (Ramos and Lugo, 1994)that we reclassified to the land-cover andnatural vegetation classes mapped for1991-92. Reclassification by the geoclimaticzones that we used to segment the imageclassification enabled us to translate intoforest formations the forest and shrubclasses from the 1977-78 data. High- andmedium-density canopy forests were re-classified to forest, while forest with low-density canopy cover and shrub were re-classified to forest/shrub or woodland/shrubland. Agriculture in the wet imagemapping zone in the west became themixed class that included coffee, wet ever-green forest/shrub, and other agriculture(the data did not distinguish between dif-ferent types of coffee cultivation).

Accuracy Assessment

The 1991 NASA aerial photos were mostsuitable for accuracy assessment becausethe large study area relative to staff re-sources made reliance on field data imprac-tical (Zhu et al. 2000). Another advantage ofusing aerial photos was that accessibilitywould not affect the probability of an ob-servation’s selection. Finally, the satelliteimagery was dated 1991-92 and the projecttook place between 1999 and 2001. Conse-quently, field-based accuracy assessmentwas problematic for estimating accuracy ofland cover, because some land-coverchanges could have occurred.

A random sample of over 250 9-pixelclusters (3 by 3 pixel groups), stratified bythe class of each cluster’s center pixel, pro-vided data for accuracy assessment. The ac-curacy assessment included manually in-terpreted or recoded classes. Locatingclusters in the original satellite imagery fa-cilitated their identification in aerial pho-tos. We photo-interpreted the woody veg-etation formation or land-cover class ofeach pixel within each cluster, recordingthe predominant as well as other classesthat mixed pixels apparently reflected. Us-ing these observations, we estimated theoverall percentage of correctly classifiedobservations, the user’s and producer’s ac-curacy for forest formations and landcover, and the Kappa coefficient (Landisand Koch, 1977), which measures accuracythat accounts for chance agreement be-tween classes in an error matrix. Producer’saccuracy is the proportion of correctly-classified accuracy assessment observa-tions, while user’s accuracy estimates theproportional assignment of pixels to a cor-rect class.

Forest formation was not always distinctin aerial photos of transitional areas be-tween, for example, moist seasonal ever-green and wet evergreen forest. Wherephoto interpretation or our knowledge oftransitional areas was insufficient to iden-tify forest formation, we recorded the mostlikely forest formation and noted the tran-sition type. Staff resources were insufficientto field-check the 5 % of accuracy assess-

E. H. HELMER ET AL.174

ment observations located in transitionalareas between forest formations.

Global forest cover classification systems

To facilitate using the data in global-scaleforest cover evaluations, we identified theclass of each forest formation for two clas-sification systems used in global forestmonitoring efforts: the GOFC (GOFC De-sign Team, 1999) and FRA systems (FAO,1996, 1998) (Table 3). The GOFC systemclassifies upland forest according to fourvariables: leaf type (needle leaf, broadleaf,mixed), leaf longevity (evergreen, decidu-ous, mixed), canopy cover (10-25 %, 25-40%, 40-60 %, 60-100 %), and canopy height(0-1 m low shrub, 1-2 m tall shrub, >2 mtrees). This system recognizes flooded for-est as a separate class.

For the GOFC system, we classified Pu-erto Rico’s moist and dry broadleaf semi-deciduous forests as broadleaf mixed for-ests. We distinguished semi-deciduousforest from the mosaic of semi-deciduousand evergreen forests that occur in the karstarea. We also grouped each forest/shrubclass with its corresponding forest forma-tion because the forest/shrub areas gener-ally have over 60 % total woody vegetationcover.

In the FAO (1998) system, natural closedforest has tree cover in the various under-stories and under growth of over 40 % anddoes not have a continuous dense grasslayer. Although our forest/shrub had treecover of 25-60 %, it does not have a con-tinuous grass layer and its canopy couldreach at least 5 m on maturity, so it is clas-sifiable as closed forest. Most forest/shrubis not forest fallow because shifting cultiva-tion is not significant in Puerto Rico. Be-cause the woodland/shrubland formationincludes some open forest or tree/shrubmixtures with a grass layer, we disaggre-gated it to 70 % closed forest, 10 % openforest, and 20 % shrubs (“other woodedland”). Using our accuracy assessmentdata, we disaggregated the two forest/shrub/coffee classes to 40 % agriculture(active sun/shade coffee and other agricul-ture) and 60 % closed forest (formerly for-est/shrub) for the first class, and 80 %

closed forest (forest/shrub, abandonedshade coffee and forest) with 20 % agricul-ture (active coffee shade) for the secondclass. We based our estimates of totalclosed forest and coffee agriculture areason these disaggregation percentages.

Summary of protected woody vegetationby formation

To illustrate the usefulness of the forestformation and land-cover map that re-sulted from satellite image interpretation,we quantified island-wide protected extentof each forest formation. A simple GIS sum-mary (ERDAS, 1999), which overlaid forestformation with a digital map of protectedareas (Dragoni, 2002), output the aerial ex-tent of each forest formation within Fed-eral, Commonwealth, or private reserveland.

RESULTS

Mapping forest type and land cover inPuerto Rico

We mapped 21 forest formations plusland cover (Fig. 2) and estimate that Pu-erto Rico (plus the islands of Icacos andPiñero) had about 364,000 ha of closed for-est in 1991-92, which composed 41.6 % ofits land area (Table 3). The second mostabundant land cover was pasture andgrassland (36.7 % of the land). Agricultureand coffee agriculture composed 5.9 % and2.4 % of the area, respectively, and 10.5 % ofthe land reflected urban and developedland cover. Remaining land covers eachcompose <1 % of the island area.

The spatial resolution of Landsat TM wasgenerally adequate for mapping Carrib-bean-island woody vegetation formations.Exceptions included some upland coastalformations too narrow to be distinct in theimagery. The Puerto Rico map includesonly a few stands of one coastal shrublandformation that we manually interpreted be-cause it mainly occurs in narrow, coastalbands. The spatial resolution of LandsatTM also contributed to difficulty resolvingelfin from palm cloud forests wherepatches were no more than a few

MAPPING FOREST TYPES AND LAND COVER 175

TABLE 3. Forest type, land cover, protected areas and proportions, and forest class for two global classificationschemes (GOFC and FAO) for the islands of Puerto Rico, Icacos and Piñero. Numbers identify classes inAppendix A. Total area of closed forest was about 363,650 ha, or 41.6% of the main island’s area. Total shrublandarea (OW) was 5,734 ha.

No. inApdx A Formation and Land Cover

Area(ha)

Protected

GOFC1 FAO2(ha) (%)7

Forest and shrubland—Dry and Dry/Moist1 Lowland dry semi-deciduous forest3 16,342 1,248 7.6 4 CF2 Lowland dry semi-deciduous woodland/shrubland3 22,957 2,044 8.9 3 Mix4

3 Lowland dry mixed evergreen drought-deciduousshrubland with succulents

1,052 250 23.8 6 OW

4 Lowland dry and moist, mixed seasonal evergreensclerophyllous forest with succulents

3,535 1,527 43.2 1 CF

Forest and shrubland—Moist and Moist/Wet— Lowland moist evergreen hemisclerophyllous shrub-

land90 45 49.8 2 OW

5 Lowland moist seasonal evergreen forest 55,383 1,093 2.0 6 CF6 Lowland moist seasonal evergreen forest/shrub 68,348 393 0.6 6 CF7 Lowland moist coconut palm forest 448 77 17.2 6 CF8 Lowland moist semi-deciduous forest 5,130 423 8.2 4 CF9 Lowland moist semi-deciduous forest/shrub 2,081 99 4.8 4 CF

10 Lowland moist and wet, seasonal evergreen andsemi-deciduous forest

25,872 2,933 11.3 5 CF

11 Lowland moist and wet, seasonal evergreen andsemi-deciduous forest/shrub

26,411 857 3.2 5 CF

Forest—Wet, Rain, Lower montane Wet/Rain12 Submontane and lower montane wet evergreen

sclerophyllous forest3,064 2.453 80.1 6 CF

13 Submontane and lower montane wet evergreensclerophyllous forest/shrub

1,956 1,109 56.7 6 CF

14 Submontane wet evergreen forest 49,790 6,027 12.1 6 CF15 Active sun and shade coffee, submontane and lower

montane wet forest/shrub, agriculture26,879 521 1.9 6 Mix5

15 Submontane and lower montane wet evergreen for-est/shrub, active/abandoned shade coffee

51,926 1,173 2.3 6 Mix6

16 Lower montane wet evergreen forest—tall cloud for-est

21,299 9,530 44.7 6 CF

17 Lower montane wet evergreen forest—palm and elfincloud forest

3,051 2,082 68.2 6 CF

17 Lower montane wet evergreen forest—elfin cloud for-est

1,104 626 56.7 6 CF

Forest—Flooded18 Tidally/semi-permanently flooded evergreen sclero-

phyllous forest36,838 2,709 39.6 7 CF

19 Seasonally flooded rainforest 319 202 63.3 7 CF

1GOFC classes: 1) Broadleaf evergreen tall shrub, 10-40%; 2) Broadleaf evergreen tall shrub, 60-100%; 3)Broadleaf mixed tree, 25-40%; 4) Broadleaf mixed tree, 60-100%; 5) Broadleaf evergreen and mixed mosaic tree,60-100%; 6) Broadleaf evergreen tree, 60-100%; 7) Flooded forest.

2FAO classes: OW—Other wooded land (shrubs); OF—Open forest; CF—Closed forest; AG—Cultivated land.3Includes Icacos island, which contains 44 ha of class 1, and 5 ha of class 21, and Piñero Island, which contains

49 ha of class 1, 11 ha of class 2, 45 ha of class 18, and 14 ha of class 22.4Disaggregate: 70% CF, 10% OF, 20% OW.5Disaggregate: 40% AG, 60% CF (16,127 ha), based on proportions of sub-types in accuracy data: 17% active

sun coffee, 17% and 7% active and abandoned shade coffee, 48% forest/shrub).6Disaggregate: 20% AG, 80% CF (41,541 ha), based on proportions of sub-types in accuracy data: 2% active sun

coffee, 13% and 9% active and abandoned shade coffee, 49% forest/shrub, 20% forest).7Percent protected for combining classes 1 and 2-18.4%, classes 5 and 6-1.2%, classes 8 and 9-7.2%, classes 10

and 11-7.2%, classes 12 and 13-63%, and classes 14 and 15-6.0%.

E. H. HELMER ET AL.176

pixels wide. Also, in more humid parts ofthe northwest karst zone, semi-deciduousforest occurs on the sides and tops of thesteep and narrowly dissected limestonehills (haystack hills, mogotes). However,seasonal evergreen forest occurs in theclosely spaced valleys between these hills.We called these areas “seasonal evergreenand semi-deciduous forest.” While geologydistinguished this mosaic of two forest for-mations, Landsat TM seemed too coarse toresolve the two formations. Additionalwoody vegetation formations that occurredin patches smaller than a few pixels, andthat were impractical to map using LandsatTM imagery, included sinkhole swamp for-ests, cactus scrub (Areces-Mallea et al.,1999), and gallery forest in the dry/moistserpentine zone.

Accuracy assessment yielded a kappa co-efficient of agreement of 0.69 ± 0.01 (Ap-pendix A), which indicates substantialagreement (Landis and Koch, 1977). Thiskappa value estimates agreement aftercombining all manually defined agriculture

and hay/pasture into one class, called ag-riculture/hay/pasture, because their sepa-rate classification accuracies were low, andconsidered observations classified to agri-culture/hay/pasture when labeled pasturein reference data. We attribute low accu-racy estimates for these classes to differ-ences between photo interpreters. Many ar-eas of former sugar cane and pineapple aregradually shifting to hay or pasture, whichcauses difficulty in interpretation. In addi-tion, for the above kappa estimate we com-bined the mixed palm and elfin forest classwith the elfin forest class. The proportion ofcorrectly-classified reference observations(producer’s accuracy) was low for urban/bare areas (Appendix A). We believe thesource of this error was mixed pixels, be-cause removing such pixels from a sepa-rately calculated error analysis increasedproducer’s accuracy to 79 %. Many mixedurban-green vegetation pixels mapped topasture.

The major error source in woody vegeta-tion formation derived from confusion be-

TABLE 3. Continued.

No. inApdx A Formation and Land Cover

Area(ha)

Protected

GOFC1 FAO2(ha) (%)7

Emergent Wetlands, Agriculture and Non-vegetated20 Tidally flooded evergreen dwarf-shrubland and forb

vegetation52 0 0.9 — —

21 Other emergent wetlands (includes seasonallyflooded pasture)

5,798 1,187 20.5 — —

26 Salt and mud flats 533 276 51.922 Pasture/grass 321,011 3,451 — — —23 Agriculture/hay/pasture 51,607 703 — — —24 Urban and bare 91,799 1,013 — — —25 Sand and rock 2,422 366 15.1 — —26 Quarries and salt mining 249 44 — — —— Water 6,348 1,379 21.7

1GOFC classes: 1) Broadleaf evergreen tall shrub, 10-40%; 2) Broadleaf evergreen tall shrub, 60-100%; 3)Broadleaf mixed tree, 25-40%; 4) Broadleaf mixed tree, 60-100%; 5) Broadleaf evergreen and mixed mosaic tree,60-100%; 6) Broadleaf evergreen tree, 60-100%; 7) Flooded forest.

2FAO classes: OW—Other wooded land (shrubs); OF—Open forest; CF—Closed forest; AG—Cultivated land.3Includes Icacos island, which contains 44 ha of class 1, and 5 ha of class 21, and Piñero Island, which contains

49 ha of class 1, 11 ha of class 2, 45 ha of class 18, and 14 ha of class 22.4Disaggregate: 70% CF, 10% OF, 20% OW.5Disaggregate: 40% AG, 60% CF (16,127 ha), based on proportions of sub-types in accuracy data: 17% active

sun coffee, 17% and 7% active and abandoned shade coffee, 48% forest/shrub).6Disaggregate: 20% AG, 80% CF (41,541 ha), based on proportions of sub-types in accuracy data: 2% active sun

coffee, 13% and 9% active and abandoned shade coffee, 49% forest/shrub, 20% forest).7Percent protected for combining classes 1 and 2-18.4%, classes 5 and 6-12%, classes 8 and 9-7.2%, classes 10

and 11-7.2%, classes 12 and 13-63%, and classes 14 and 15-6.0%.

MAPPING FOREST TYPES AND LAND COVER 177

FIG. 2. Map of Puerto Rico natural vegetation and land cover.

E.H

.HE

LM

ER

ET

AL

.178

Fig.2live

4/C

tween forest successional stages within thesame ecological zone. Combining forestwith forest/shrub or woodland/shrublandwithin each image mapping zone signifi-cantly improved agreement to a kappa co-efficient of 0.79 ± 0.01; for forest formations,it also yielded user’s and producer’s accu-racies between 62 % and 92 %. Additionalerror derived from confusion between ad-jacent forest formations, such as confusionbetween lowland moist seasonal evergreenand submontane wet evergreen forest, andmisclassification as pasture of some recov-ering or previously little disturbed shrubvegetation on limestone hilltops. Spectralsignatures similar to pasture apparently re-flected poor soils, where potential biomassaccumulation or recovery rates are low. Wealso noted confusion between pasture andthe higher-elevation mixed class that in-cluded active sun/shade coffee, submon-tane/lower montane wet forest/shrub, andother agriculture.

Protection of Puerto Rico’s forest formations

Protection of natural vegetation forma-tions ranges from about 0.6-2 % for lowlandmoist, seasonal evergreen forest formations(not including abandoned coconut palmplantations), to 56.7-80.0 % for submontaneand lower montane wet evergreen sclero-phyllous forest formations on serpentinesubstrates. Percentages of protected forestwere lowest for moist seasonal evergreenforest formations. The extensive woodyvegetation mosaic of active and abandonedshade coffee and forest/shrub has a pro-tected proportion of 1.9-2.3 %. The overallprotected proportion of Puerto Rico, Icacosand Piñero is about 5.2 %, including Fed-eral, Commonwealth and private reservelands.

DISCUSSION

Approach to image interpretation

Our approach to satellite-image basedmapping of land cover and forest type ad-dressed several related challenges in com-plex tropical areas. A major challenge iscomplex topography and geology, its asso-

ciated variation in vegetation formations,and the resulting spectral confusion be-tween vegetation types. First, separatingreference observations into sunlit vs. shad-owed slopes likely improved land coverdistinction (Helmer et al., 2000). Second,segmenting the image interpretationhelped overcome the problem of spectralconfusion among land covers and foresttypes; for example, the segmented ap-proach allowed distinction between drysemi-deciduous woody formations andyoung forest or pasture occurring in wetterzones. Incorporating geology into the im-age mapping zones separated the uniqueforest formations developed on serpentineand karst substrates. Submontane wet ev-ergreen forests on serpentine, for example,would have otherwise been confused withtidally flooded forested wetlands. Rainfallidentified areas where limestone hill veg-etation becomes mostly lowland moistsemi-deciduous forest or forest/shrub, asopposed to a mosaic with seasonal ever-green forest. Third, adding an elevationband to the submontane-lower montaneimage mapping zone allowed distinctionbetween cloud forest formations vs. dis-turbed forest at lower elevations, anddense-canopied submontane vs. lowermontane wet cloud forests. The accuracyassessment results helped to validate oursegmented approach to mapping forest for-mation, but they also indicated that map-ping of forest successional stages was lessreliable. Finally, manually delineatingsome agriculture relieved its confusionwith forest or urban areas.

With regard to challenges of spatial scaleand class resolution, mapping floristically-defined forest communities, remnant old-growth forest, or certain forest types thatoccur in patches smaller than a few pixels,would require more field surveys andcould benefit from imagery with finer spa-tial resolution than Landsat TM.

The segmented approach permittedmapping forest formations simultaneouslywith land cover. Unlike previous island-wide mapping, this map better identifiesthe spatial distributions of forest for-mations where certain groups of endemicspecies occur. Previous work has mapped

MAPPING FOREST TYPES AND LAND COVER 179

ecological zones or land cover plus cover-based woody vegetation classes and man-grove. Although distinguishing forest suc-cessional stages was less successful, themap resulting from the approach presentsthe most detailed picture to date of PuertoRico’s natural vegetation and land cover. Adrawback of our approach is that the imagemapping zones not only segmented ourclassification, but also defined the extentsof some formations and therefore couldonly approximate the boundaries betweenthem. The resulting visible and evenboundaries between zones could more re-alistically depict patchy ecotone shifts inforest formation. The approach also re-sulted in misclassification of coffee cultiva-tion that extended into moist forest zones(see Appendix A). Integrating spectral datawith continuous climate and elevation datausing machine-learning algorithms (e.g.Muchoney et al., 2000), along with incorpo-rating more extensive field-derived refer-ence observations from transitional areas,could yield maps that more realistically de-pict patchy and uneven transitions betweenforest formations.

Protection of natural vegetation inPuerto Rico

Our detailed vegetation map identifiedan uneven distribution of reserve areasacross forest formations. It indicates thatonly 1.2 % of lowland moist seasonal ever-green forest or forest/shrub are protected(Table 3). Generally, these forests occur atlower elevations where rates of land-coverconversion to urban and developed landsare highest (López et al., 2001). Some foresttypes in Puerto Rico are well-protected, in-cluding 45 to 68 % of cloud forest types and43 to 80 % of the sclerophyllous forests thatdevelop on serpentine substrates (Table 3).However, quarries in serpentine areas oc-cur immediately adjacent to reserve areasand therefore require monitoring. Exten-sive wetland draining and forested wetlandclearing has occurred for agriculture orother uses (Lugo and Brown, 1988). Exclud-ing riparian wetlands, about 20 % to 63 %of remaining forested or non-forested wet-lands receive protection (Table 3).

Acknowledgments.—This research wasfunded in part by a U.S. National Aeronau-tics and Space Administration - Institu-tional Research Award to the University ofPuerto Rico (Grant NAG8-1709, UPR Sub-contract 2000-000946). We thank RichardWebb and John Parks, U.S. Geological Sur-vey-Puerto Rico; Carmen Santiago, U.S.Natural Resources Conservation Service-Puerto Rico; NASA Global Observation ofForest Cover program; Alexis Dragoni,DNER; two anonymous reviewers and theeditor, and John Francis, Alberto Areces-Mallea, J. Danilo Chinea, Julio Figueroa,Ariel Lugo, Sandra Molina, Juan Ramirez,Fred Scatena, Alberto Rodríguez and IvanViscens, for guidance on Puerto Rico’snatural vegetation.

LITERATURE CITED

Aide, T. M., J. K. Zimmerman, M. Rosario, and H.Marcano. 1996. Forest recovery in abandonedcattle pastures along an elevational gradient innortheastern Puerto Rico. For. Ecol. Manage. 28:537-548.

Areces-Mallea, A., et al. 1999 (active January 2002). Aguide to Caribbean vegetation types: classificationsystems and descriptions. N. Panagopoulos (Ed.),The Nature Conservancy International Head-quarters, Washington, D.C., 166 pp. http://edcsnw3.cr.usgs.gov/ip/tnc/

Calvesbert, R. J. 1970. Climate of Puerto Rico and theU.S. Virgin Islands. Climatology of the UnitedStates No. 60-52. U.S. Department of Commerce, 29pp.

Caminero-Rodríguez, G. 1991. Estudio de la veg-etación en afloraciones de serpentinita en elBosque Estatal de Maricao, Puerto Rico. M.S. The-sis, University of Puerto Rico, Mayagüez, PuertoRico. 118 pp.

Cedeño-Maldonado, J. A. 1997. Vascular flora of theRio Maricao watershed Maricao CommonwealthForest, Maricao, Puerto Rico. M.S. Thesis, Univer-sity of Puerto Rico, Mayagüez, Puerto Rico. 157 pp.

Cedeño-Maldonado, J. A. and G. J. Breckon. 1996. Ser-pentine endemism in the flora of Puerto Rico.Carib. J. Sci. 32:348-356.

Chinea, J. D. 1980. The forest vegetation of limestonehills of northern Puerto Rico. M.S. Thesis CornellUniversity. Ithaca, New York., 70 pp.

Chinea, J.D. 2002. Tropical forest succession on aban-doned farms in the Humacao Municipality of east-ern Puerto Rico. For. Ecol. Manage. 167:195-207.

Cohen, W. B., T. A. Spies, and M. Fiorella. 1995. Esti-mating the age and structure of forests in a multi-ownership landscape of western Oregon, U.S.A.Int. J. Rem. Sens. 16:721-746.

E. H. HELMER ET AL.180

Crist, E. P., R. Laurin, and R. A. Cicone. 1986. Vegeta-tion and soils information contained in trans-formed thematic mapper data. Proceedings,IGARSS ‘86 Symposium, Zürich, Switzerland, 8-11September 1986, pp. 1465-1470. ESA SP-254, Paris.

Dansereau, P. 1966. Studies on the vegetation ofPuerto Rico. 1. Description and integration of theplant-communities. Institute of Caribbean Science,Mayagüez, Puerto Rico, Special Publication 1:1-45.

Dinnerstein, E. D. et al. 1995. A conservation assess-ment of the terrestrial ecoregions of Latin Americaand the Caribbean. The World Bank, Washington,D.C. 129 pp.

Dragoni, A. 2002. Digital map of protected areas ofPuerto Rico. Puerto Rico Department of Naturaland Environmental Resources, San Juan, PuertoRico.

ERDAS, Inc. 1999. Earth Resource Data Analysis Sys-tem Field Guide, Fifth Edition. ERDAS, Inc., At-lanta, Georgia, 672 pp.

Ewel, J. J. and J. L. Whitmore. 1973. The ecological lifezones of Puerto Rico and the U.S. Virgin Islands.USDA For. Ser. Res. Pap. No. ITF-18, Inst. Trop.For., 72 pp.

FAO. 1996. Forest resources assessment 1990, surveyof tropical forest cover and study of change pro-cesses. FAO Forestry Paper 130, Rome, Italy, 152pp.

FAO. 1998. FRA 2000: Guidelines for assessments intropical and sub-tropical countries. FAO workingpaper No. 2, Rome, Italy, 45 pp.

FGDC. 1997 (active January 2002). National vegetationclassification standard. Vegetation Subcommittee,Federal Geographic Data Committee, FGDC-STD-005, June 1997, U.S. Geological Survey, Reston, Vir-ginia, 58 pp. http://biology.usgs.gov/fgdc.veg/

Figueroa Colón, J. 1996. Geoclimatic regions of PuertoRico (map). U.S. Dept. Interior Geol. Surv. WaterRes. Div. San Juan, Puerto Rico.

Foody, G. M., G. Palubinskas, R. M. Lucas, P. J. Cur-ran, and M. Honzak. 1996. Identifying terrestrialcarbon sinks: classification of successional stages inregenerating tropical forest from Landsat TM Data.Remote Sens. Environ. 55:205-216.

Garcia, G. R. 1991. Relaciones taxonómicas entre laflora endémica de serpentina en Susúa, Puerto Ricoy Río Piedras, Gaspar Hernández, Repúlica Do-minicana. M.S. Thesis, University of Puerto Rico,Mayagüez, Puerto Rico, 137 pp.

Gesch, D. et al. 2002. The national elevation dataset.Photogramm. Eng. Rem. S. 68:5-11.

GOFC Design Team. 1999 (active January 2002). AStrategy for Global Observation of Forest Cover,Draft for comments, Version 1.2, January 4, 1999,http://www.gofc.org/gofc/docs/docs.html

Hansen, M. C., R. S. Defries, J. R. G. Townshend and R.Sohlberg. 2000. Global land cover classification at 1km spatial resolution using a classification tree ap-proach. Int. J. Rem. Sens. 21:1331-1364.

Helmer, E.H. 2000. The landscape ecology of second-

ary forest in montane Costa Rica. Ecosystems 3:98-114.

Helmer, E. H., W.B. Cohen and S. Brown. 2000. Map-ping montane tropical forest successional stageand land use with multi-date Landsat imagery. Int.J. Rem. Sens. 21:2163-2183.

Holdridge, L. R. 1967. Life zone ecology. Tropical Sci-ence Center, San Jose, Costa Rica. 206 pp.

Krushensky, R.D. 1995. Generalized geology map ofPuerto Rico. U.S. Geol. Sur., San Juan, Puerto Rico.

Landis, J. R. and G. C. Koch. 1977. The measurementof observer agreement for categorical data. Biomet-rics 33:159-174.

Little, E. L. and F. H. Wadsworth. 1964. Common treesof Puerto Rico and the Virgin Islands. USDA ForestService Handbook No. 249, Washington, D.C., 548pp.

López, T. del M., T. M. Aide and J. R. Thomlinson.2001. Urban expansion and the loss of prime agri-cultural lands in Puerto Rico. Ambio 30:49-54.

Loveland, T. R., et al. 2000. Development of a globalland cover characteristics database and IGBP DIS-Cover from 1 km AVHRR data. Int. J. Rem. Sens.21:1303-1330.

Lugo, A. E. and S. Brown. 1988. The wetlands of Ca-ribbean islands. Acta Cient. 2:48-61.

Ma, Z., M. M. Hart and R. L. Redmond. 2001. Mappingvegetation across large geographic areas: integra-tion of remote sensing and GIS to classify multi-source data. Photogramm. Eng. Rem. S. 67:295-307.

Muchoney, D., et al. 2000. Application of the MODISglobal supervised classification model to vegeta-tion and land cover mapping of Central America.Int. J. Rem. Sens. 21:1115-1138.

Murphy, P. G., A. E. Lugo, A. J. Murphy and D. C.Nepstad. 1995. The dry forests of Puerto Rico’ssouth coast. In: A. E. Lugo and C. Lowe, C. (eds.)Tropical forests: management and ecology. Ecol.Stud. 112:178-209.

Myers, N. R. A. Mittermeier, C. G. Mittermeier, G. A.B. da Fonseca, and J. Ke. 2000. Biodiversityhotspots for conservation priorities. Nature 403:853–858.

Oldfield, S., and C. Sheppard. 1997. Conservation ofbiodiversity and research needs in the UK depen-dent territories. J. Appl. Ecol. 34:1111-1121.

Ramos, O. M. and A. E. Lugo. 1994. Mapa de la veg-etación de Puerto Rico. Acta Cient. 8:63-66.

Rivera, L.W. and T. M. Aide. 1998. Forest recovery inthe karst region of Puerto Rico. For. Ecol. Manage.108:63-75.

Scatena, S. N. and A. E. Lugo. 1995. Geomorphology,disturbance, and the soil and vegetation of twosubtropical wet steepland watersheds of PuertoRico. Geomorph. 13:199-213.

Scott, J.M., et al. 1993. Gap analysis: a geographic ap-proach to protection of biological diversity. Wild.Monogr. No. 123. Suppl., J. Wildl. Manage. 57:1-41.

Skidmore, A. K. 1989. An expert system classifies eu-calypt forest types using thematic mapper data

MAPPING FOREST TYPES AND LAND COVER 181

and a digital terrain model. Photogramm. Eng.Rem. S. 55:1449-1464.

Stehman, S. V. and R. L. Czaplewski. 1998. Design andanalysis for thematic map accuracy assessment:fundamental principles. Rem. Sens. Environ. 64:331-344.

Steininger, M. K. 1996. Tropical secondary forest re-growth in the Amazon: age, area and change esti-mation with Thematic Mapper data. Int. J. Rem.Sens. 17:9-27.

Strahler, A. H. 1981. Stratification of natural vegeta-tion for forest and rangeland inventory usingLandsat imagery and collateral data. Int. J. Rem.Sens. 2:15-41.

USDA Forest Service and US Geological Survey. 2001(active May 2002). Forest cover types of the United

States. National Atlas of the United States, http://nationalatlas.gov/atlasftp.html.

Vogelmann, J. E., T. Sohl and S. M. Howard. 1998.Regional characterization of land cover using mul-tiple sources of data. Photogramm. Eng. Rem. S.64:45-57.

Weaver, P. L. 1995. The colorado and dwarf forests ofPuerto Rico’s Luquillo Mountains. In A. E. Lugo,and C. Lowe (Eds.) Tropical forests: managementand ecology. Ecol. Stud. 112:109-141.

Zhu, Z., L. Yang, S. V. Stehman and R. L. Czaplewski.2000. Accuracy assessment of the U.S. GeologicalSurvey regional land-cover mapping program:New York and New Jersey region. Photogramm.Eng. Rem. S. 66:1425-1438.

E. H. HELMER ET AL.182

APPENDIX A. Error matrix for 2303 accuracy assessment observations for 26 classes. Observations within the four- to six-cell shaded blocks collapse to one cellfor the 19 classes that result from collapsing forest successional stages within 6 classification zones. Overall accuracies (Po) and kappa coefficients of agreement,respectively, are 71% and 0.69 ± 0.01 for 26 classes and 83% and 0.79 ± 0.01 for 19 classes. The variance estimates for kappa do not account for lack of independenceamong secondary sampling units within clusters (Stehman and Czaplewski 1998).

Class

Reference

User’s accuracy %

Number of classes

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 26 19

1 41 7 — — — — 16 — — — — — — — — — — 6 — — — — — — — — 5976

2 14 52 1 — — — — — — — — — — — — — — 1 — — — 12 — — — — 653 4 7 32 — — — — — — — — — — — — — — — — — — 4 — 6 — — 604 — — — 40 — — — — — — — — 5 — — — — — — — — 5 — 7 1 — 695 — — — — 31 35 1 3 — — — — — 6 — — — 3 1 — 1 5 3 — — — 35

706 — — — — 3 41 — — — — — — — 8 9 — — — — — — 4 — 2 — — 617 — — — — 7 — 43 — — — — — — — — — — — — — — — — — — 1 848 — — — — — — — 80 5 1 — — — — — — — — — — — — — — — — 93

989 — — — — — — — 16 26 — — — — — — — — — — — — 1 — 1 — — 59

10 — — — — — 3 — — — 30 18 — — — 5 — — — — — — — — — — — 5473

11 — — — — 9 4 — — 5 17 28 — — — 4 — — — — — — 5 — — — — 3912 — — — — — — — — — — — 35 21 — — — — — — — — — — — — — 62

9113 — — — 4 — — — — — — — 11 20 — 1 — — — — — — 4 — — — — 5014 — — — — 2 2 — — — — — — 3 45 22 — — — — — — 4 — — — — 58

9215 — — — — 2 1 — — — — — — — 22 156 — 1 — — — — 5 — 4 — — 8216 — — — — — — — — — — — — — 10 25 53 6 — — — — — — — — — 5617 — — — — — — — — — — — — — — 10 21 56 — — — — — — — — — 6418 9 — — — 4 — 3 — — — — — — — — — — 41 — — — — — 2 — — 7019 — — — — — 8 — — — — — — — — — — — — 21 — — 1 — — — — 7020 — — — — — — 1 — — — — — — — — — — — — 18 — — — — — — 9521 — — — — — — — — — — — — — — — — — 2 — — 61 — — — — 5 9022 — 13 — — 2 8 — 4 11 2 1 — 1 2 8 — — 1 — 1 4 135 9 21 — 2 6023 — — — — — — — — — 1 — — — — — — — — — 3 — — 338 12 — — 9524 — — 3 1 — — 3 — — — 3 — 2 — 1 — — 9 — — — 21 2 86 — — 6625 — — — — — — — — — — — — — — — — — 1 — — — — — 2 56 — 9526 — — — — — — 1 — — — — — — — — — — 2 — — — 1 1 9 8 65 76

Producersaccuracy%

19classes

60 66 89 89 52 40 63 78 55 59 56 76 39 48 65 72 89 62 95 82 92 65 96 57 86 89 Po = 71%

26classes

78 68 85 92 89 82 Po = 83%

MA

PPING

FOR

EST

TY

PES

AN

DL

AN

DC

OV

ER

183