Evaluating Seasonal Variablity As an Aid to Cover-Type Mapping From

PEER.REVIEWED ARIICTE

EvaluatingSeasonal Variability as an Aid to Gover-TypeMapping from Landsat Thematic MapperData in the Noftheast

JamesR. Schrieverand RussellG. Congalton

Abstract rate classifications for the Northeast (Nelson et ol., 1g94i Hopkins ef 01.,19BB). However, despite these advances, spe- Classification of forest cover types in the Northeast is a difficult task. The conplexity and variability in species contposi- cific hardwood forest types have not been reliably classified tion makes various cover types arduous to define and in the Northeast. Developments within the remote sensing identify. This project entployed recent advances in spatial community have shown promise for classification of forest and spectral properties of satellite data, and the speed and cover types throughout the world, These developments indi- potN/er of computers to evaluate seasonol varictbility os an aid cate that, by combining supervised and unsupervised classification techniques, increases in the accuracy of forest to cover-type mapping from Landsat Thematic Mapper (ru) classifications can be expected (Fleming, 1975; Lyon, 1978; dato in New Hampshire. Dato fron May (bud break), Sep- the su- tember (leaf on), and October (senescence) were used to ex- Chuvieco and Congaiton, 19BBJ.By combining both plore whether different lea.f phenology would improve our pervised and unsupervised processes,a set of spectrally and informationally unique training statistics can be generated. ability to generate forest-cover-type maps. The study area covers three counties in the southeastern corner ol New' This approach resr-rltsin improved classification accuracv (Green Hampshire. A modified supervised/unsupervised approach due to the improved grouping of training statistics was used to classify the cover types. A detailed accuracv os- and Teply, 1991). This studv will utilize tlvt satellite data sessment was perfornted to evaluate the clossification. The and a combined classification approach to help determine if results indicate that specific northeast hardwood species con it is possible to discriminate between specific northeastern be identified ond that tinte of the yeor can significanily offect hardwood forest types. the cover-type classification accuracy. In addition, few studies to date have employed satellite imagery taken during autumnal senescence.Autumn data sets have been shown to increase accuracies in hardwood Introduction forest type delineations when applied to aerial photography A relevant and accurate fbrest cover-tvoe and/or land-classi- (Ecler, 19Sg). It is also possible that imagery acquired in the fication system is essential to providing information for effec- spring, at or shortly after bud break, may provide advan- tive management of natural resources.Research aimed at tages for specific hardwood species delineation. Therefore, developing methods for reliably classifying forest cover and/ this study compared classification accuracies for three tem- or habitat types dates back many decades and continues to poral data sets (autumn, spring, and summer) to determine (Heimburger, this day 1934; Westveld, 1952; Damman, 1964; if seasonal variabilitl' significantlv affects classification ac- Pfisteret aL.,1977:Eyre, 1980; Leak, lgaz; Smalley, 1986). curacY. Three approaches to the classification of habitat types have been outlined: biophvsical, forest type, and forest type/forest ClassificationSystem soils classifications (Leak, 1982). The single factor important For information generated bv satellite imagerv to be useful, a to all three approaches is the classification of forestland into classification svstem which utilizes this information must be specific cover types. Satellite imagery has been demonstrated developed. If the results of a classification are to be of value to be a cost effective method fbr classifying forest cover to potential users, it is also important that the classification types throughout the world (e.g.,|oyce, 1978; Kushwaha, scheme be well defined, reievant, understood, and accepted. 1990; Green, 1990; Schardl et aL.,1990;Congalton et o1., To deveiop a relevant and useftii classification, project objec- 1993t. tives and data limitations must be determined. In the United States,remote sensing projects involving One limitation of tM satellite imagery when compared to the classification of forest types have typically focused on aerial photographv is that spatial resolution cannot be con- forest tvoes in the South or West. Recent advances in tech- trolled. To identify individr,ral tree species r-rsingaerial pho- nology *hich have improvecl the spatial, spectral, and radio- tography, a scale of 1:8,000 or larger seems appropriate metric properties of Landsat Thematic Mapper (rv) .satellite (Ciesla, 19sg). The small scale associated rvith Tu satellite imagery have shown promise for increased success in accu- Photogrammetric Engineering & Remote Sensing, Department of Natural Resourc;es,University of New Hamp- Vol. or, No. 3, March 1995, pp. 321,-327. shire, 215 JamesHall, Durham, NH 03824. 0099-11 1219s/6013-321s3.00/0 f.R. Schriever is presentlv with Pacific Meridian Resources, O 19S5 American Societv for Photogrammetry 5200 S.W. Macadam Avenue, Suite 570, Portland, OR 97201. and Remote Sensrnu

PE&RS PEER REVIEWED ARIICTE

imagery limits its usefulness to classification of forest stands can be termed a purely single- or multi-factor system.AI- with a minimum of several acres rather than individual trees. though site factorsare not consideredin defining or identify- Therefore, forest cover type classification systems employed ing the types, the descriptions"give recognition to the in studies utilizing satellite imagery must define cover types ecologicalfactors that helped to createthe types and will at the siand level. continue to influence their development" (Eyre, 1980).Com- An important consideration in choosing a forest-type bining theseclassification schemes (see below) should facili- classification svstem is if existing or potential (climax) vege- tate applications which utilize satellite data for forest tation t1,pes are to be described. Methods for gathering infor- cover-typemapping in the forestscommon to the Northeast. mation to determine potential vegetation types typically In addition, forest-typerecognition coupled with soil maps involve examination of understorv vegetation, regenerating can provide information which indicates likely habitats,cli- tree species, and/or examination of soil conditions (Leak, max species,and martagementIimitations and/or potential 1982). Befort (ts8o) was able to utilize understory vegetation (Leak,1982). This informationwill be valuableto foresters, and regeneration for aerial identification of habitat types in wildlife ecologists,and other natural resourcemanagers. northern Idaho and eastern Washington using very large scale photography. Dense canopy stlucture typical of the Forest Types Northeast, and the small scale of satellite imagery, limits the Cover usefulness of this technique for classification of northeastern Coniferous cover tvoes. . WP - Easternwhite pine comprisesa maiorityof the stocking Two approaches that describe the existing vegetation are (>70 percent) and characteristically occurs in pure stands. single-factor (Bailey et al,, "lg7B) or dominance type, and On lighter textured soil its'associates include red pine, pitch multi-factor classifications. The single-factor classification pine, quaking and bigtooth aspen, red maple, pin cherry, and white oak. On heavier soils associates are birches(paper, method is based on a single measure, for example, dominant sweet, and gray), black cherry, white ash, northern red oak, vegetation. The mr-rlti-factorclassification takes into account sugar maple, basswood, hemlock, red spruce, balsam fir, several distinguishing characteristics.Habitat types, for ex- white soruce. and northern white cedar. ample, can be identified by soils, landforms, and chronose- . WH - Eistern white pine and eastern hemlock, in combina- quences of vegetation (Leak, 1gB2). tion, comprise the largest proportion of the stocking, but nei- Multi-factor classifications can result in the proliferation ther species alone represents more than half of the total. The of categories.For example, five elevation classes,five soil- combination rarely exists without associates and red maple is tvpe classes,and five cover-types classeswill result in 125 a very common one. Other common associates include paper possibie classifications. Congalton (1991) has stressed the im- birch, northern red oak, beech, sugar maple, yellow birch, gray birch, red spruce, white ash, and balsam fir. portance of employing a mutually exclusive and totally ex- . HE - Eastern hemlock is pure or provides a maiority of the the haustive classification approach for determining accuracy stocking (>70 percent). Common associatesare eastern white of remotelv sensed classified data. For the multi-factor classi- pine, balsam fir, red spruce, sugar maple, beech, yellow fication syitem to approach a mutuallv exclusive and totally birch, northern red oak, white oak, yel1ow poplar, basswood, exhaustive condition, it must be subjectively simplified into black cherry, red maple, and white ash. categorieswhich have only a few distinguishing characteris- o OC - Other conifer species comprise a majority (>70 percent] oI the stocking.ln lhe strrdyarea lhe most common speciesis By way of corrtrast,a single-factor ciassification system red pine. However, red spruce and any other conifer species is mutr.rally exclusive, totally exhaustive, easy to define and in the study area which comprises a majority of the stocking mav be included in lhis category. apply, and is very objective. This approach can also be hier- Mixed archical in nature when more general categoriesare desired. . MX - White Pine/Red Oak/Red Maple - Eastern white pine In addition, a single-factor classification approach can be and northern red oak are the most important species in this used to determine a varietv of site characteristicsbecause forest cover type, although red maple is always present. several resource values appear to be highly correlated (Bailey White ash is often a maior associate. Other trees commonly et c.|.,1,978; Leak, 1982). found are eastern hemlock, birches (paper, yellow, and Last, and perhaps most important, is that the classifica- sweet), black cherry, basswood, sugar maple, and beech. tion chosen must be acceptable and relevant to potential Deciduous o - users. The land classification scheme most popular within RM Red maple comprises a maiority of the stocking. Most common associatesinclude red spruce, balsam fir, white the remote sensing community is the one developed by An- pine, sugar maple, beech, yellow birch, eastetn hemlock, pa- derson et o.l.[1s76). This svstem is described in detail in per birch, aspen, black ash, pin cherry, northern red oak, and various manuals and texts iJ"nr",r, 1983: Campbell, rgBZ; black cherrv. Lillesand and Kief'er, 1987) and is pertinent for a variety of . OAK - White oak, black oak, or northern red oak comprise a project objectives. However, projects attempting to determine majority of the stocking. One or more species of hickory are if rtu satellite imagery can reliably classify specific forest consistent components but seldom make up over 10 percent cover types requires a classification scheme which describes of the basal area. Other associates may include sugar and red forest cover types in greater detail than the one proposed by mapies, white and green ash, American and red elm, bass- Anderson t1976). wood, black cherry, American beech, and hemlock. o BH - Beech must comprise at least 25 percent of the stocking The forest cover-type classification currentlv used in the and may be associatedwith any of the above listed hardwood State of New Hampshire is the Society of American Foresters species. (sar) classification scheme (Eyre, 1sBO).In addition, Leak (1982) has identified seven forest tvpes representing major stand conditions in New Hanrpshire. These approaches de- StudyArea scribe forest cover tvpes in greater detail than the one devel- The study area is located in southeastern New Hampshire oped bv Anderson et ol. (tgza). However, neither approach and includes portions of Strafford, Merimack, and Rocking-

322 PE&RS PEER.REVIEWED ARIICLE

tober 1989 (#520621,4513),B september 199O (#523821,4466)' and 13 May 19BB(#5153414570). All data were acquired from EOSAT CorPoration' Acquisitionof Refetence Data The forest cover-type reference data utilized in this stuCy were acquired froni three different sources: the State of New Hampshire, UNH Woodlands Office, and additional field work conducted as part of this project. Guidelines for defining cover types ar-eoutlined in the sAF publication (Eyre- 1980J, and werf the standard utilized by all three sources' Specific 0102030 utilized fbr this project are defined above' Part of l# cover types to help develop training areas Milos the reference data was utilized arrd the remainder was used to perform an accuracy assess- Reference data provided by t*ot o*to ment for resulting classificationJ. $ the State included two wildlife management units and three State parks and totaled 1'5,775 acres. The University data orovided an additional 1,800 acres. Due to the lack of certain cover types, it was necessary to supplement the existing reference database through per- sonai field work. The assistance of University of New Hamp- shire Cooperative Educators in forestry for Strafford and Rockineham Counties was essential to the completion of this studv. To identify, delineate, and locate these areas in the field, National High Altitude Photography (Nunp) color-infra- red (ctn) photos, topographic maps, o.rthophotos,and tree farm manigement maps were utilized' Once the areas were (nnn) point cruise of the area MEBRIMRCK located, a tlen basal area factor was performed. Resulting inventory data provided the means for cbver typing specific-forest stands in accordance with the previously defined cover tYPes.

MancheslerO Analysis Anaiysis of the three TM satellite data setswas divided into five maior steps:(1) derivingnew bands,(2) delineating training areas,(s) generatingstatistics an-d.spectral pattern analysi"s,(4) classifyingthe images,and (s) assessingthe ac- Figure1. Locationof the projectstudy area. All image proCessingfor this study was performed using'nnDRs"*r"ty. Veision 7.5 sof[wareon a 486 personalcom- ^outer IERDAS,1991). In addition to the originalraw bands 1 to 5 and 7' two tvoes of derivativebands were utilized in the analysis'The of the threevisible ham counties and encompassesover 260,000acres (Figure fitit *^. a principle componentanalysis between the three 1). Historically, much of the areawas converted from forest bands. Beciuse oTthe strong correlation of thesebands land to agricuiture during the eighteenthand nineteenth cen- visible bands,the first princlpal component visible band variability (85 to turies (Irlind, 19S2).These farms have largely reverted to explains a large portion of th-e principal componentof the forest as agriculture moved west beginning in the late nine- 95 percent).T*heiefore, the first be used as a substitute for teenth ceniury. Forest types in the area range from early suc- ,risible bands (pcr) can potentially similar to a panchro- cessionalstages to mature forests.Topography for the area is thesebands and can be thought of as of the three bands' The sec- relatively flai with a range in elevation-of sea level in the matic rendition of a combination Band ratios 4/3 (R4/3) Great Bay areato 1,,41,3feel above sea level on Fort Moun- ond method was band ratio analysis. sensitive to changesin tain in the Nottingham Mountains. and s/+ (R5/4)have been shown to be (Peterson al.' Therefore, The study arei was chosen to help the Uliversity of vegetationcharacteristics et -1'sB6)' mentioned ratios and also in- New Hampshlre (uNs) Woodlands Office in the development thi"sstudy employed the above This resulted in a final image of a geographic information system (cts) cover-type data cluded a"715(r:7l;\ ratio band. ^ 1 to 5 and 7; the first prin- laverl T[e Woodlands Office managesover 1,800 acresof containing ten bands: raw bands (pcr);and ratios R+/s, Universitv owned woodlands in the area.In addition, the ciple componentfor the visible bands areawas chosento maximize available state and university R5/4,and R7l5. utilized in this foresttvpe existingreference data sets(over 17,000acres), The training area delineation technique training area poly- becauseof its close proximity to campus which easedthe study was a traditional approach-whereby Criteria im- collection of field data, and to provide a variety of forest gotrj u." digitized on the image display device' representa- types. iortant to ihe selection of training areas include cover the Thiee Landsat TM satellite imageswere used in this iion o. distribution of the areas for each class throughout areas on the image dis- study. Correspondingscene dates and n numbers are 23 Oc- image, the ability to locate training

323 PE&RS PEER REVIEWED ARIICTE

play device (this generally required stand size to be a mini- duced classeswhich did not meet a pre-determined minimum of ten acres), and areas must represent normal mum criteria for similarity. If classes with unknown labels conditions for the class which they represent. were still present, visual inspection, analyst expertise, and/or Generation of statistics and spectral pattern analysis was personal field work were used for final labeling. After all sta- performed to meet two specific objectives: (t) to determine if tistics were labeled and similar statistics were merged, the fi- training areas were acceptab)efor the final classification pro- nal classification was oerformed. cess, and (2) to reduce the data to only those bands neces- The classification algorithm utilized was a maximum- sarv for the final classification. A statistical analysis of each likelihood classifier with a first-pass parallelepiped optimiza- potential training area was performed. The statistics file pro- tion set at two standard deviations. Initial classifications duced in the training area selection provides a covariance were run without the first-pass parallelepiped optimization. matrix and standard deviations for each band, as well as a A comparison of runs with and without the optimization histogram lbr each band. It is important that training areas showed virtually no difference. Because the parallelepiped used in the final classification display unimodal histograms optimization cuts computer processing time and does not as- and have relatively low standard deviations and low values sume a normal data distribution, it was utilized. The main in either major diagonai of their covariance matrices. advantase of the maximum-likelihood classifier is that it To help speed up computer processing time and reduce takes the variability of classesinto account by using the co- data redundancy, it is valuable to reduce the data to only variance matrices of classes. those bands which maximize class separability. The graphi- Accuracy assessmentis an essential component of the cal and statistical spectral pattern analysis techniques em- classification process. Therefore, a complete accuracy assess- ployed include spectral pattern plots (Stenback and ment was performed on all three classifications generated Congalton, 1990), divergence analysis, and ellipse plots as a during this project, Over 335 polygons or forest stands were final diagnostic. It should be noted that only the statistics utilized as reference data for testing the accuracy of all ten generated trom the training areas during the supervised ap- classes in each classification. None of the training data poly- proach were utilized for the spectral pattern analysis' This gons utilized in the classification was used in performance of was essential for determining which bands best discrimi- the accuracy assessment. To determine which class reference nated between forested classes. data polygon was assigned in the final classification, an ER- Spectral pattern plots are a simple graphical technique DAS program called sutvltvtARYwas run on the reference poly- in which the average value for each band for each category gons and the final classified images. This program provides a in the classification is plotted. The graphs can then be stud- cross tabulation between the reference data polygon and the ied to see which bands provide the best visual separation be- classified image. Analyst judgment (in accordance with the tween categories. Divergence analysis is a mathematical guidelines outlined in the classification system) was then technique which computes a statistical distance between cat- used to determine which class was assigned to specific poly- egories in a classification based on variances and covariances 8ons. as well as average values. The divergence analysis provides a Descriptive statistics calculated from the error matrices more robust analysis than the spectral pattern plots. Finally, include user's error (error of commission), producer's error ellipse plots can be used to further verify the spectral pattern (errors of omission), overall accuracy (total percent correct), anall,sis by plotting, in two bands, the signatures of the cate- and normalized accuracy. Analytical techniques utilized dis- gories in the classification and looking for signature overlap. crete multivariate technioues and include normalization or Ellipse is best performed after the number of bands has been standardizing the matrices and performance of a KAPPA anal- limited because the plots are performed on two bands at a ysis. Normalization allows for an effective comparison be- time. Once the appropriate bands are selected, the classifica- iween e..or matrices because the normalized value takes into tion of the data may proceed. account both user's and producer's etror, allowing for direct Traditional approaches to classification include super- comoarisons w.ithin individual cells. The r,q,ppninalvsis vised and unsupervised techniques. Because both techniques r-rtiliied in this study provides information about a single have inherent advantagesas well as disadvantages,many matrix and facilitates a statistical comDarison of several ma- combined approaches aimed at maximizing the advantages trices. For a complete description of these accuracy assess- while minimizing the disadvantageshave been developed. ment techniques, see Congalton (1991). The combined approach utilized for this study was developed by Chuvieco and Congalton (1sBB).This approach util- Results izes the mean data values (training area means or cluster Three spectral pattern analysis techniques were employed in means) generated from both the supervised and unsupervised this study. The first technique utilized was spectral pattern classification approaches. Mean values for each band are in- plots. The statistics for all training areas in each class were put into another clustering routine which begins merging sta- merged, and the mean digital number (oN) values for each tistics which are similar. The final output is a dendrogram class were graphed and visually inspected to determine indicating the numeric distance between statistics which which bands maximized separability (Figure 2). This step were merged or grouped together. serves as an excellent preliminarv check as to which classes The information from the dendrogram is used in labeling can be discriminated. However, beca,lse there is little indica- the unknown clusters as well as in reducing the number of tion as to the loss or gain of separability by the addition or training areas. When training areas of known classes(super- subtraction of bands and the standard deviation and covari- vised areas) were grouped with training areas of unknown ance statistics are not considered, it is difficult to assess classes(unsupervised areas),it was possible to label the un- which bands maximize separability from this technique known spectrally unique classes. In addition, some training alone. areas displaying spectrally similar characteristics could be A more robust method for determining which bands merged into one class, thereby reducing the total number of maximize separability is through signature divergence. The training areas. This process was repeated until results pro- DIVERGEcommand in ERDASoffers both transformed diver-

324 PE&RS PEER.REVIEWED ARIICTE

cific hardwood soecies. For the October imase the hardwood species American beech, northern red oak, and red maple had user's and producer's accuracies of 69 percent and 73 percent, 83 perient and g1 percent, and 85 percent and 94 120 percent, respectively (Table 1). Table 2 presents the results of the accuracv assessment 100 for each date of imagery using all three measuies of accu- q) racy: overall accuracy, KHAT accuracy, and normalized accu- f (u racy. These three measures produce identical results in that 80 thev all rank October with the highest accuracy Ievel and z September as the lowest. Table 3 shows the results of the o KAPPAanalysis. Both the October and May classification c (g 60 were significantly better than the September classification at o the 99 percent and 95 percent level, respectively. However, although the October classification has slightly higher accuracies, it is not significantly different from the May classification. These findings seem to indicate that successessimilar to

TABLE1. ERRonMnrRrx roR rne CLnssrncATroNoFrHr OcroseRlvnce. 5 7 VisPCa/3 514 715 ReferenceData Octobe r Band ROW WH HE OC RM NF MX OAK BH rorAr ,l TrainingArea c +5 J I 59 *wP -f HE )eWH +OC *RM -;MX +OAK +BH I a 7 t4 9 9 39 s 4 4 Figure2. Spectralpattern plot of the mean digital number s oc (DN)values for the forestclasses in the Octoberdata i z 5 I z 10 t 5C t. RM i I 3+ 36 c NF a 53 l 54 t MX gence and /efferies-Matusita distance formulas for calculating 8 t2 5 I 48 A 80 - i spectral distances between classes.Both formulas produced o OAK identical +J J 47 results; that is, onllr fsul bands of imagery were n BH needed for the final classification. The addition of a fifth ll t5 band did not increase forest class separabilitv for anv of the coL 50 30 30 2A 40 56 50 52 l6 255 imagery. while utilizing only three bands would have re- TOTAL sulted in a loss of forest class separability. The ellipse program, which graphs any user specified combination of OVERALLACCURACY =255/344 = 7 4% training area statistics two bands at a time, was utilized as a final check to verify the final band combinations. PROOUCER'SACCURACY USER'SACCURACY VEGETATIONTYFES Results of the spectral pattern analysis indicate that TM WP = 43150 = 86 % WP = 43159= 73% raw bands 3, 4, and 5 are valuable for forest cover-tvoe clas- WP = white pine wH =r4l30 = 47% wH =14139= 35% WH = white pire, hemlock sifications in the Northeast. In addition, the value oi deriving HE = hemlak HE=4/30 = t3% HE =4/4=r00% new bands (in this case, principle component analysis of the OC = other conifer oc =5/20 - zs% oc =S/'to = 50% visible bands and band ratio techniques) was demonstrated. RM = red maple RM = 34140= 85 % RM =34/36 = 94% NF =rcn-forested Each of the images used in the final classification contained MX = white pire, red oak, NF - 53755 = 95 % NF = 53/54 = 98 % at least one derived band. For the September. October. and red maple MX = 48150 = 96 % MX = 48/80 = 60 % OAK = oak(black,red, white) May data, the bands utilized were [pir,4,5, and Rs/4];[3,4, oAK=43/52 = 83 = 43/47 = 91 % BH = bech, hardwood s, and R+/gl;and fpcr, q, s, and R4/3],respectivelv. In this OAK =11/16= =11/15= study. bands 1,2. and R7lswere not utilized.ThL fact that BH 69% BH 73Va the R7ls band was not utilized agrees with results from Gal- lup (1991) who found that, in the Northwest, band ratio z/s did not aid in maximizing separability of forest cover types. Tnare 2. CoN,4pARrsoNoF ALr THREEAccuRAcy MensuRes roR tHr All images were classified in nnoas z.s using the MAX- CLnsstrtcnttons. CLASprogram (maximum-likelihood classifierl with a two- Overall KHAT Normalized standard-deviation parallelepiped optimization option. To lmage Accuracv Accuracy Accuracy produce the final classified image, two iterations of the com- September bined classification approach (Chuvieco and Congalton, 62% 56"/u 54Yo October 74o/o 70,'1, 64o/o 1988) were run on the data. In two iterations, it appears this May 69"/" 64"1, 60% classification approach can begin to distinguish between spe-

PE&RS PEER.REVIEWED ARIICTE

Tnere3. REsuLTSor rne KAPPAAnnLvsts ron CovpnRtsollor TM ble as pure standsin the referencedata in order to avoid Clqsstrtcnttotts. problemsassociated with mixed covertypes. The problem which still remains for classificationof Comparison Z Statistic NortheasternNoflneastern cover types whichwnrcn utilizeutllrze satellitesatelllte imageryrmaSery aP-ap- - October versus September 3.665* pears to be associated with the complex mixed forest typestypes. October versus May 1.568 easternwhite pine/easternhemlock, and easternwhite -2.O97* The May versus September pine/northern red oak/red maple mixed types accounted for * sisnificant at 957olevel most of the off-diagonalerrors. This makes sensebecause ** significantat 99%,level there are so manvmany factorsrs which can affect the spectralrereflec- tance of mixed types. These types can be hard to distinguish in the field. and subiective techniques are often used in cre- ating the final delineation of stand cover-type maps. Ques- aer- those found in classification of hardwood species using tions regarding the legitimacy of the mixed types and how to (Eder, ial photography acquired during autumnal senescence compensate for these problems are difficult to answer, but utilize rses) may also be expected from classifications which certainly relate back to how we define and establish the cri- at or satellite data. It also appears that satellite data acquired teria for mixed stand designations. shortly after bud break may provide advantages over summer The only criterion utilized for designation of individual New data sets for classification of hardwood species found in stands in this study was species type. While employing the Hampshire. combined classification approach, it became very clear that additional criteria were needed to maximize the data poten- Summaryand Discussion tial of satellite imagery as weII as the diagnostic value of the This study had two primary obiectives. The first was to de- combined classification approach. The combined approach termine if seasonal rrariablitity makes a significant difference showed that statistics generated through the unsupervised in accuracy of TM cover-type classifications in New Hamp- approach were measuring more variables than were being ex- shire. The second was to determine if advanced techniques plained by the supervised cover-type data' Therefore, future and methodologies for classification of remotely sensed satel' studies should strive for reference data that contain more in- lite data would enable the classification of specific hardwood formation regarding specific forest types. Some of the data species found in the Northeast. that would be useful include stand age, size class, structure, To meet the first obiective, rv satellite data sets from overall stocking levels, and a specific breakdown of species various times during the year were acquired. The fact that percentages within the mixed forest types. October classification accuracy was the highest did not come Finallv, it should be noted that methods aimed at com- as a surprise. It was hypothesized that the difference in hard- bining imagery and utilizing additional data enhancement wood foliage reflectance characteristics (e.g.,leaf biomass, techniques should help increase classification accuracy. Ad- water moisture, and chlorophyll absorption) between species ditional techniques include principal component analysis would be at a maximum during autumnal senescence. How- utilizing a wide variety of combined data, spatial filtering of ever, the fact that the May classification was also shown to the final classified image, use of combined and individual be significantly better than September was interesting. Be- multitemporal data sets for generating additional vegetation cause some species break bud sooner than others, chloro- indices and band ratios, utilization of ancillary data, and phyll absorption rates, water moisture levels, and leaf edge analysis investigations. AIso, because it appears that biomass levels should be distinctly different between species certain image dates may map specific species more effec- in May. It is also likely that understory reflectance character- tively, a combining of the classified images may also im- istics associated with species such as red maple that typi- prove classification accuracies. cally lose there leaves before 23 October or species that break bud at or after 13 May aided in the discrimination of Acknowledgments specific species. The authors would like to thank Jessica Burton and Casey This project also seems to indicate that specific hard- Moffitt for their help in completing this project. Acknowledg- wood speiies can be discriminated by classification of TIr't ment is also given to the University of New Hampshire Agri- imagery. It also clearly demonstrates the value of an itera- cultural Experiment Station for funding provided to this tive, hierarchical classification scheme combined with a study under Mclntire-stennis proiect #MS-32. Finally, Drs' assessment.Many resource managers may complete accuracy ]ames L. Smith and fohn G. Lyon are acknowledged for their queslion the value of data found to have an overall accuracy fine suggestionsto the manuscript. of z4 percent. However, when project obiectives are defined, classesmay be aggregated,thereby improving expected accuracy dramatically. For example, a wildlife project interested References in mapping hardwood mast production may aggregate both This aggregation the American beech and red oak classes. Anderson, ]. R., E. E. Hardy, J. T. Roach, and R. E. Witmer, 1'976.A will result in a producer's and user's accuracy of 90 percent Land IJse and Land Cover Classificotion System for IJse with Re- and gB percent, respectively. note Sensor Dofo, Professional Paper 964, United States Geolog- This study utilized a fairly extensive reference data set. ical Survey, Washington, D.C., 28 p. However, even with the extensive reference data available, Bailev, R. G., R. D. Pfister, and J. A. Henderson, 1978. Nature of projects involving a larger databaseand a wider variety of Land and Resource Classification - A Review, lournal of For' -over types should be employed. It would be beneficial to esfry, 76:650-655. have a siudy area that encompassed a wider variety of cover Befort, W., 1986. Large-Scale Sampling Photography for Forest Habi types with at least 50 accuracy assessmentreference stands tat-Type Identification, Photogrammetric Engineering & Remote each. tt is also important that the species present are availa- Sensing, 52[4J:101-108.

326 PE&RS PEER.REVIEWED ARIICTE

National Aero- Campbell, 1,987.Introduction to RemoteSensing, The Guilford ScannerDala, NASA ReferencePublication 1015, I.8., p' Press,New York, sst p. nautics and Space Administration, Hottston, Texas, 43 Detec- Chuvieco,E., and R. Congalton,1988. Using Cluster Analysis to Im- Kushwaha, S. P. S., 1990. Forest-Type Mapping and Change prove the Selectionof Training Statisticsin ClassifyingRe- tion from Satellite Imagery, ISPRS Journal of Photogrammetry motely SensedDaIa, PhotogmmmetricEngineering & Remote and Remote Sensing, 45:775-1'81' Sensing,56(9):L27 5-7287 . Leak, W. 8., 1982. Habitat Mapping and lnterpretation in Neu' Eng- Ciesla.W, M., 1989.Aerial Photosfor Assessmentof ForestDecline 1and, USDA Forest Service Research Paper NE-496, 28 p' - A Multinational Overview, lournal of Forestry,87(2):37-41'. Lillesand, T. M., and R. W. Kiefer, 1987. Remote Sensr'ng and Image Congalton,R. G., 1991.A Reviewof Assessingthe Accuracyof Clas- Interpretation, Second Edition, John Wiley and Sons, New York, iifications of Remotely SensedDaIa, Remote Sensing of Environ- 727 p. Commrrnities in the ment, 37:35-46. Lyon,' lohn G., 1.578.An Analysis of Vegetation Congalton,R., K. Green,and J. Teply, 1993.Mapping OId Growth Lower Columbia River Basin, Pecora IV Symposium, Sioux Forestson National Forestand Park Lands in the Pacific North- Falls, South Dakota, 7 p. west from Remotely SensedDaIa, Photogrammetric Engineering Nelson, R. F., R. Latty, and G, Mott, 1984. Classifying Northern For- & RemoteSensing, 59[4J:529-535. ests Using Thematic Mapper Sin'rulator Data, Photogramnetric Damman,A. W., 1964.Some ForestTypes of Centtal New'foundland Engineering and Renote Sensing, 50(5):607-617' and Their Relation to EnvironmentalFactors, Forest Science Peterson, D. L., W. E. Westman, N. f. Stephenson, V. G. Ambrosia, J' MonographB,62 p, A. Brass, and M. A. Spanner, 1986. Analysis of Forest Structure Eder, J. I., 1ss9. Don't Shoot Unless its Auturnn, lournal of Forestry, Using Thematic Mapper Simulator Data, IEEE Transoctions on 87(6):50-51. Geoscience and Remote Sensing, GE-2411):1'1'3-120' ERDAS, 't991.EIDAS V.7.5 SystemGuides, ERDAS, Inc., Atlanta, Pfister, R. D., B. L. Kovalchik, S. F. Arno, and R. C. Presby,'1'977' Georgia. Forest Habitat Types of Montona, USDA Forest Service General Eyre, F. H. (editor), 1980.Forest Cover Types of the U.S. and Can- Technical RePort INT-34, 1'74P. ado, Societyof American Foresters,Washington, D.C.' 148 p. Schardt, M. K., A. Schurek, and R. Winter. 1990. Forest Mapping Map Sheet 1:200,000 Fteming,M. D., S. Berkebile,and R. M. Hoffer,1975. Compute.r- Using Satellite Imagery. The Riegensburg l. and Remote Aided Analysis of Landsat-1 MSS Data: A Comparison of Three as Exlample, ISPRS lournal of Photogrammetry Approaches, Including a "Modified Clustering" Approach, LARS Sensing, 45:33--46. Information Note 072475,Purdue University, West Lafayette,In- Smalley, G. W., 1986. Classification and Evaluation of Forest Sites diana,pp. 54-61. on the Northern Cumberland Plateou, USDA Forest Service Gen' Gallup, B., 1SS1.Semi-Automof ed Training Area Selectionand a Tech. Rep. 50-60, 7a P. Nonparametric Classifier Compared to Traditional Digital Satel- Stenback, J., and R. Congalton, 1990. Using Thematic Mapper Im- lite bata Classification of Fotest Types in Northern California' agery to Examine Forest Underslory, Photogranmetric Engineer- M.S. Thesis,University of California,Berkeley, California, 17o p. ing & Remote Sensirrg,56[9):1285-1290. '1952. Green, K., 1990. Mapping Forest Vegetation: National Forest and Westveld, Marinus, A Method of Evaluating Forest Site Quality Forest Ser- ParkLands in Oregonand Washington,ERDAS Monitor, Sum- from Soil, Forest Cover, and Indicator P/anfs, LISIIA mer 1990,15 p, vice Station NE Station Paper 48, 12 p Much Re- Green,K., and J. Teply, 1991.Old Growth Forest:How (Received 2L May 1993; accepted 10 August 1993; revised 12 Octo- mains, Geo Info Systems,1(4):22-31. ber 1993) Heimburger,Carl C., 1934.Forest-Type Studies in the Adirondack University Agricultural Experiment Region,Memoir 165, Cornell |ames Schriever Station,Ithaca, N.Y., 122 p. james Schriever received a B'S. degree in for- and T. Lillesand,1988. Assessment of Hopkins,P. F., A. Maclean, estry from the State University of New -York under Lake Thematic Mapper Imageryfor ForestryApplications College of Environmental Science and forestry PhologrammetricEngineering & RemoteSens- StatesConditions, at Sviacuse in 1988, and an M.S. degree in Re- rng,54(11:61-68. souice Administration and Management' spe- Irland, L. C.,1.582.Wildlands and Woodlots- A Story of New En' cializing in remote sensing and Gis, from the-University of gland's Forests,University Pressof New England,Hanover, New , New Hirnpshire in 1992. He is currently with Pacific Merid- Hampshire,217 P. ian Resouices where he is involved in several remote sensing R., 1936.Introductory Digital Image Processing.A Remote Jensen,J. projects covering a variety of applications. His pri- SensingPerspective, Prentice-Hall, Englewood Cliffs, New Jer- and cIs integration of several forms of ancil- sey,379 p. mary interest is in the data to produce GIS systems which are applications A. T., 1978.Procedures GatheringGround Truth Informa- lary Ioyce,' for managers successfully complete tion a Supewised Approach to a Computer-Implemented- oriLnted for ielping resource for constraints' Land Cover ehssification of Landsat-AcquitedMultispectral and comply with regulatory and environmental

327 PE&RS