l PEER.REVIEWED ARIICTE

lmprovedFotest Glassification in the NorthernLake States UsingMulti-Temporal Landsat lmagery

Peter T. Wolter, DavidJ. Mladenoff,George E. Host, and ThomasR. Crow

Abstmct to the MSSsensor (Williams et al., 1,984:Toll, 1985),The ad- (band to 1.75pm, Forest classifications using single date Landsat rM data have dition of two middle infraredbands 5, 1..55 contentof been only moderutely successful in separating cover and band 7,2.o8 to 2.35pm),sensitive to moisture forest (Tucker, Hunt ef al,, tgBZ; types in the northern Lake States region. Few regional vegetation 1980; Ripple, 1986; forest to im- classificotions have been presented that achieve genus or Hunt and Rock, 1gB9;Wolter, 1990),has been shown prove (Toll, 1sB5; Bensonand species level accuracy. We developed a more specific forest classificationresults forest L990; Moore and cover classification using rtr't data early summer in con- DeGloria, 1985; Stenbackand Congalton, from of the TM junction with uss dates to capture phenological changes Bauer, 1990).However, the increasedresolution four classificationsof suffi- of different species. Among the 22 types classifed, sensorhas not resulted in forest cover forest (i,e., to warrant practical use multi-temporal image analysis aided in separating 1.3types. cient detail Anderson Level III) (Skidmore and Of greatest significance, trembling aspen, maple, nor-th- of this technologyby forestlandmanagers multi-temporal or multi- ern rcd oak, northern pin oak, black ash, and tamarack were Turner, 19BB).Classifications using potential for higher forest classifi- successfully clossified. The overoll classification occuracy phenological imagery have cation precision over single-dateclassifications (Schriever was 83.2 percent and the forest classification accuracy was and Congalton,1993). In this paper we develop an applica- 80.1 percent. This approach may be useful for broad-scale tion of this multi-phenological approachto classify dominanl forest cover monitoring in other areas, particularly where an- cillary dato layers are not available. forest specieswithin northern Lake Statesconditions. 0bjectives Introduction The oblectivesof this study include Forest cover type mapping in the northern Lake States(Min- nesota,Wisconsin, and MichiganJ using spacebornesensors o Developinga forest classificationwith dominant tree species has been a forest managementgoal since the launch of Land- Ievel precision within northern Lake Statesconditions, phenologicalevents sat-1 on 23 1972. Forest classificationsof large regions o Using ltss digital data to capture major July of hardwood forest cover tyPes, with Anderson Level III precision (Anderson et aL,,L976) are o Reassessingthe utility of vss data for multi-temporal or especially neededto assistlandscape-scale analysis and man- multi-phenologicaiforest classifications,and agementobjectives (Mladenoff ef 01.,1993; Mladenoff and o Determiningthe practicality of a layered classificationap- Pastor,1993). Unfortunately, detailed level III forest cover proach utilizing image ratioing and ratio differencingtech- mapping efforts using a single date of LandsatMultispectral niques for multi-temporal image analysis. Scanner(lrss) data have been largely unsuccessful(Mead and Meyer, 1977;Roller and Visser,1980; Downs, 1gB1), Background Moore and Bauer (rggo) concluded that forest heterogeneity There are few accountsof researchwhere TM data have been in northern Minnesota and the suboptimal spectraland radio- used to classify northern mesic and boreal Lake Statesfor- metric resolution of the MSSsensor preclude detailed classifi- ests.Studies that used Tv or Thematic Mapper Simulator cation. The Thematic Mapper (rv) sensorsaboard Landsats4 (rus) data in this region have covered small areas(Shen ef and 5 (launched in 1982 and 1.984,respectively) provide en- o/., tgB5; Hopkins et ol., 1.sBB;Moore and Bauer,1990; Bol- hanced spatial, spectral,and radiometric resolution superior stad and Lillesand, 1992) relative to the 34,225-Wrfcoverage of a full Landsat scene. Using airborne TMSin northern Min- P.T. Wolter and G.E.Host are with the Natural ResourcesRe- nesota,Shen ef o1.(1985) achieved 84.2 percent accuracyfor searchInstitute, University of Minnesota, Duluth, MN 55811. five forest species:red pine (Plnus resinosa),iack pine (P. banksiana),black spruce (Piceamafiana), paper birch (Be- D. f. Mladenoff, formerly with the University of Minnesota, tula papyfifero),papynlera), and tremblitremrtltng aspen lropulu(Populus tremuloides). Duluth, Natural ResourcesResearch Institute, is now with Theii 23-km'zstudy area was ideal as it contained mostly the Wisconsin Departmentof Natural Resources,and Univer- sity of Wisconsin-MadisonDepartment of ,Madison, PhotogrammetricEngineering & RemoteSensing, wI 53706. Vol. ot, No. 9, September1995, pp. 7129-1.1-43.

T.R. Crow is with the USDA Forest Service,North Central 0099-111 2/95/6109-1 129$3.00/0 ForestExperiment Station, Forestry SciencesLaboratory, O 1995 American Society for Photogrammetry Rhinelander,WI 54501. and RemoteSensing

PE&RS PEER.REVIEWED ARTICTE

pure, spatially homogeneouscover types. Furthermore,be- forestedarea near Ottawa, Canada.For discrimination of co- causethe TMSinstrument was flown at an altitude of 72OO niferous forest, deciduous forest,and agricultural land using m, atmosphericaffects may have been negligible. Buchheim a maximum-likelihood decision rule. thev found that three et al. (1985),using simulated spor (SystemeProbatoire d'Ob- dates of MSSimagery from |une. September,and October servation de Ia Terre) data of a 60-km' study area in north- provided the best results (84 percent overall classificationac- western Wisconsin, were able to visually discriminate curacy) over all other single- or multiple-date classifications Anderson level III (species)forest types. However, while tested.The OctoberMSS scene captured peak senescencefor computer-aidedclassification (maximum likelihood) was ex- most hardwoods while the |une MSSscene captured hard- cellent at level I (96 percent overall accuracy)and level II flush. Kalensky and Scherk (tgzs) concluded that, (91 percent overall accuracy),level III classificationprecision although the October,fune, and SeptemberMSS scenes indi- (90 percent overall accuracy)was limited to lowland types: vidually produced low overall classificationaccuracies (67 white-cedar-balsamfir (Thuia occidentolis-Abiesbalsamia), percent, 69 percent, and 81 percent, respectively),their col- tamarack (Larix laricino), and black spruce. Hopkins ef 41. lective use mitigated the effectsof individual image noise. (rgaa) used Tru data of a 15-km' study area in northwestern Beaubien (1979)concluded that comparing or superimposing Wisconsin and reported Anderson Level III accuracyfor red MSSimages taken at different seasonsprovides better con- pine and jack pin-ebut only Level II accuracyfor remaining trast among certain types of vegetationin easternQuebec. forestclasses. Hopkins ef o1.,(1988) and Shen ef 01.(1985) Using rrassdata for a classification of the Crater Lake Na- consideredconditions in their respectivestudv areasunre- tional Park region, Walsh (1980)found that Septemberim- presentativeof typical northernLike Statesfoiest cover. agery provided more information than summer MSSdata due In northern Wisconsin, Bolstad and Lillesand (19s2) to the phenological condition of vegetationand the Iower combined a priori information (soil and terrain position) and sun an8le. TM data in a maximum-likelihood classificationof two study Conversely,Kan and Weber (1978) determined that there areas(each 300 km'z).Pooled classificationaccuracy for one was no clear benefi,t of multiple-date classifications over sin- of the areasreached 94 percent for northern hardwoods, red gle-dateclassifications using MSSdata for nine broad vegeta- pine, jack pine, pine/haidwood. upland brush, lowland coni- tion communities acrossthe United States(including central fer, lowland brush, Sphagnum,lowland vegetation,crop/pas- hardwoods, northern hardwoods, northern conifers, and bo- ture, soil/urban, aquatic vegetation,and water, an increaseof real). Nelson et al. (1984)stated that senescentimagery 24 percent compared to the same area classified without o should be avoided for forest classifications in New England. priori knowledge. However, genus or specieslevel discrimi- Their imagery recorded the later stagesof senescencewhere nation was not obtained for most forest types. The ancillary many forest standswere leafless.Similarly, Toll (1985)in a data they used were manually digitized or scannedfrom 1: Maryland study used both November (uss and ru) and luly 24,000-scaleU. S. GeologicalSurvey maps (terrain position) (uss and ru) data in a comparisonbetween classificationpo- and 1:2o,0oo-scaleU. S. Soil ConservationService soil sur- tentials of the two sensors.ToU (1985)concluded that the vey maps. Unfortunately, these types of ancillary data are November TM classificationaccuracy was not significantly not yet conveniently available in contiguous digital form for better than MSSaccuracy. This result was attributed to fall many areas(Mladenoff and Host, 1994).To manually digitize color variability and foliar loss in the November imagery, or scan these data layers for large regional classifications Their November results also suggestthat MSSdata may be would be an enormoustask. Multi-temporal image analysis superior to TM data when analyzing senescentimagery. That provides additional forest cover informition without reiiance is, the greaterpixel size of MSSdata (79 by 56 m) could alle- on human-derived ancillary data. viate some of the spatial/spectralheterogeneity caused by au- Changesin spectralreflectance caused by phenological tumn senescence. differencesamong temperateforest tree speciesmay allow The identification of tree specieson aerial photographs for Anderson level III forest cover tvpe classificationon a re- using phenological aids has been studied in great detail gional scale.Large seasonalvariations in forest speciesspec- (Sayn-Wittgenstein,1961). Eder (fge9) used true color aerial tral responsein the visible portion of the electromagnetic photography of autumn senescenceto map hardwood forest spectrum(Miller et o1,,1991.;Eder, 1989; Schwaller and speciesin the Medford RangerDistrict of the Chequamegon Tkach, 1985) and phenological differencesin senescence National Forest in northern Wisconsin. He found best separa- among tree species(Ahlgren, 1957; Sayn-Wittgenstein,1.961; tion between sugar maples (Acer saccharum)and mixed as- Eder, 1989) presentunique forest classificationopportunities. pen/paper birch standswas achievedby acquiring photogra- The accumulation or unmasking of pigments such as antho- phy during the peak of sugarmaple senescence.Eder cyanins (responsiblefor scarletto red leaf coloring), carote- (personalcomm., 1992) notes that paper birch, trembling as- noids (orangeto yellow coloring), tannins (brown coloring), pen, and bigtooth aspen (P. grandidentato) will remain green and xanthophylls (yellow coloring) following the denaturing for approximately one week after peak sugar maple senes- of are responsiblefor spectral change(Goodwin, cence.Thereafter, paper birch tended to color a few days 1958;Moore, 1965;Sanger, 1.971.;Boyer ef o1.,L988). Boyer prior to trembling aspenand bigtooth aspen (Ahlgren, 1957; ef d1.(1988) point out that tree speciescharacterized by se- J.J.Eder, personal comm., 1992).Conversely, black ash (Frax- quential decline (such as Quercuspolusfris) may inus nigral lose their leavesprior to peak sugar maple be significantly different spectrally from tree speciesexhibit- senescence(Eder, 1989),providing a window that can last as ing synchronouschloroplast decline. long as two weeks (personalobservations). Kalensky (192+)states that significant improvements in Schriever and ConealtonCongalton [tgg3](1993) used tTM imagery cover- multi-date image classificationcould be made if the images ing threehree key dates to determine if phenologicalphenoloeical differences used were selectedon the basis of spectralpatterns rather could improve forest classification accuracv of a 1052-km'z than on the basis of image availability alone. Kalensky and region in southern New Hampshire. They performed separate Scherk (1975)analyzed single-stageclassification accuracies forest classificationson imagery from May ( break), Sep- for various combinations of spring to autumn vSS data for a tember (stablegrowing season),and October (senescence),

1130 PE&RS PEER.REYIEWED ARIICI.E

different dates (Saderand Winne, 1992).Some investigators consider image differencing and image ratio differencing techniquesfor changedetection relatively uncomplicated and somewhat more accuratethan comparing multiple classi- fications(Woodwell et d1.,1983; Singh, 1986). Furthermore, Sader and Winne (rSSz) suggestthat image ratioing and im- age ratio differencing techniquesare preferred over principal componentsanalysis (rcR) becausethe transformedresults of PCAare often difficult to interpret.

Materialsand Methods StudyRegion The study region encompassesan area of 28,000 km' or roughly 83 percent of a full LandsatTM scenein northwest- ern Wisconsin (Figure 1). The ChequamegonNational Forest is located approximately in the north-central portion of this region. This glaciatedlandscape is characterizedby gentle topographic relief with boreal forestsin the north along Lake Minnesota Suoerior where clav soils are often quite wet, northern mesic for'estson loamy mtraines which.tt"k" .,p the maiority of the central region, northern xeric forestsor pine barrenson sandy soils, and some areasof oak savannato the south (Curtis,1959; Pastor and Mladenoff,1992). This regionis a complex mosaic of successionalforest types due to wide- spread and destructivelogging practicesthat took place in the late nineteenth and early twentieth centuries (Mladenoff Figure1. Locationof study area in northwesternWiscon- and Pastor,1993; Mladenoff and Stearns,1993). sin. ChequamegonNational Forest (- - -). TMlmage Acquisition Constraints All Landsat data used correspondto Worldwide Reference System coordinatespath 26 row 28 which are centeredat ap- and found that the October sceneorovided the best discrimi- proximately 46"N, 91'26'W. TM image selectionwas based (Fogus nation among the hardwood speciesAmerican beech on severalconstraints: grandifolia), northern red oak (Q. rubra), and red maple (A. rubrum). May imagery was secondbest with an overall accu- . imagery at least 90 percent cloud free racy of 69 percent comparedto 62 percent for September . relative humidity less than 60 percent and 74 percent for October classifications.Schriever and . wind speedless than 30 km/h r within 6 - 21 Congalton(1993) suggestthat the successof the October and date |une June May classificationover the Septemberclassification is a func- Satellite image data acquired for this study are summarized tion of differential chlorophyll absorption,foliar moisture, in Table 1. The TM image selected was ID 5120016163ac- and forest canopy characteristics. quired on 14 June 1987. According to data gathered from Schriever and Congalton (1993)compared the results of three weather stations within the region, mean relative hu- three separateforest classificationswith no attempt to com- midity between 900 and 1000 hours on this date was approx- bine raw data from different dates in either a layered classifi- imately 48 percent and average wind speed was 24 km/h. cation (Weismilleref al., lgzz: Hixson ef a1.,1980; Lozano- Relative humidity was considered because incident and re- Garcia and Hoffer, 1985) or a single-stageclassification. flected visible radiation scattered by water vapor in the at- Lozano-Garciaand Hoffer (1985)state that layered classifica- mosphere could adversely affect classification precision tions applied to multi-temporal satellite data are more effi- (Potter and Shelton, 1974). Wind speed was considered be- cient and accuratethan single-stageclassifications. The cause excessive winds would expose abaxial surfaces of for- stepwise nature of layered classificationsallows the analyst est . The axial and abaxial leaf surfaces of many (1) to optimize the use of specific spectralbands and (2) to species have very different albedo values which may intro- choosethe best seasonfor the identification and classifica- duce problematic spectral variability (Kharuk, 1992). Finally, tion of individual cover types (Lozano-Garciaand Hoffer, date was important because forest tree species are best sepa- 1985). rated using remote sensing techniques with imagery gathered Previous work demonstratesthat temporal image differ- encing techniquesare powerful tools for characterizing changesin forest canopy characteristics(Vogelmann, 19BB; Tnerr 1. Suvtunv or lvncrRv Useo trurnr CLASSIFIcATIoN. Vogelmann and Rock, 1989).Vegetation indices such as the Sensor Date Season Phenology normalized differencevegetation index (Nlvt) derived foom remotely senseddata collected throughout a growing season MSS 10 May 1992 sprrng aspen leaf flush can enhancedifferences in vegetationphenology (Tucker ef TM 14 Jun 1987 eariy summer all Ieaves flushed MSS 13 Sep 1985 early autumn black ash leaf-off al., tgBS; Goward ef o/., tgB5; Loveland et al., Lgg'L;Samson, MSS 0B Oct 1980 autumn oak senescence 1993).MSS data have been used to discriminate major MSS 25 Feb lsBB wrnter tamarack leaf-off changesin green leaf biomassby combining NDVIlayers from

PE&RS PEER.REVIEWED ARTICI.E

Tnerr 2. Cusstrteo CovenTvprs ANDVALIDATIoN METHoD UsED. FoREsr 1988) was chosen to classify the leaf-off phenology of tama- Cusses lrucluoeSAF FoRrsr CrnssEsl, USFS Cusses2, and Classes rack (Table 1). Derived Using USFS Species Codes and DNR Forest Stand Information3. CovrR Tvprs CusstREo Ustno Murrt-TrttponAL h'fice ANALYSIS(r) ANDCovER SpectralCalibration and Geometric Registtation (tr)' TypESlNDrREcrLy l|pRoveo ASA RESULTor Murtt-TrvponALIMAGE ANALYSIS AII digital MSSand TM data were calibrated to reflectanceac- Foresttypes validated using USFS and DNR forest stand information cording to Price (1987)before geometriccorrections were made. We geometricallyregistered the TM imagery to UTM - - Pr red pinel Pt trembling aspen3 zone 15 coordinateswith a pixel size of 28.5 metres using Pb - jack pine' P - mixed asPenDl nearest-neighborresampling with second-orderpolynomial Pm - black sprucec' Abp - balsam fir-aspen' Pg - white sprucel Psh - E. white pine-hardwood' transformationequations. We achieveda -mean-squared msc - mixed swamp coniferol Bpc - paper birch-coniferol error (Rusn) of unit weight of approximately 0.35 pixels for LI - tamaracktl Toh - Northern white-cedar' the fit between the digital rv data and the 1:24,000-scale Fn - black asht' Tch - E. hemlock-yellow birch' USGStopographic maps using 26 evenly distributed ground - ar - Northern red oakr' Fnc black ash-lowland conifeil' control points gatheredfrom the USGStopographic maps. - Qe - Northern pin oakr' Qep Northern pin oak-pinel:r All MSSscenes were initially transformed(first order) As - sugar maplen' Pbo - jack pine-oakr' from S7-metrepixels to 28.5-metrepixels using nearest- Cover types validated by photo-interpretation and field verification neighbor resampling (Rrrasa: 0.0 pixels). Each MSSscene

SS - sparselystocked forest sh - and herb.n was then coregisteredto the TM digital data (bilinear interpo- up - urban or pavement ff - flooded forest Iation resampling)using 26 image-to-imagecontrol points of - grass-forb ow - open water per MSSscene with a second-ordergeometric model. AII RMS cf - clearedforestl S - Sphagnum sp, errors were less than 0.5 pixels for the fit between 28.5-metre MSSdigital data and rM digital data. An independent assess- ment of the MSScoregistration to the Tv digital data was per- formed by looking at 18 check points throughout the study early in the growing season(Kan and Weber, 1978; Shen et. area.All RMserrors were within 0.7 pixels (first order) for d]..19851. the fit between 28.5-metreMSS and TM data. MSSdata selection was basedupon availability of cloud free datesthat correspondedwith the unique phenological ForestClassification System windows of the target forest species(Table 1). The number of We choseto follow the Society of American Foresters() suitable dates was few due to the 16-day repeat cycle of Land- classificationsystem (Eyre, 19s0) to define most of the forest sat and to frequent cloud coverage.MSS digital data were types (Table 2). The sAF forest types used in this classifica- chosen over TM data primarily becauseMSS data are more af- tion are the same or similar to the types used by the United fordable. In addition, MSSdata are of sufficient resolution to StatesForest Service (usns)except for the jack pine-oak.We detect coarseforest canopy differencessuch as leaf-on versus included a trembling aspentype to subdivide the sAF de- Ieaf-off (Williams, 1975) associatedwith the phenology win- scription of mixed aspen.Wisconsin Departmentof Natural dows exploited in this study. Resources(lNn) forest stand information (Locey, 1990) and USFSspecies codes were used for forest tvpes that did not fit ForestPhenology and MSS lmage Acquisition either ihe sAF or usFStype descriptions (Table z). We used Peak fall color for sugarmaple at Park Falls, Wisconsin, DNRand USFSforest stand maps as well as 6 fune 19BBNa- roughly the center of our tM scene,is approximately 21 Sep- tional High Altitude Aerial Photography(uHap) as reference tember with an annual variance of + 4 days [].J.Eder, per- information in training and assessmentof the classification. sonal comm., 1952).One scenein the MSSarchive came near Stand maps are extremely useful as ground truth information this phenological constraint (Il asosoo1.621.4,L2 September (Kalenskyand Scherk, 1975)because they provided stand lo- 1985).Based on personal observations,peak fall color for red cation and an indication of stand composition (Shen ef. o1., oak tended to be about two weeks later than sugar maple for 1sB5). our region. There were no cloud-freeMSS scenes similar in age to our TM scenefor this oak phenology window. There- PreliminaryClassification and Validation fore, an older MSSscene was selected(lo zzoael61'sr,B Octo- We used rvr bands3 (0.63to 0.69pm),4 (0.76to 0.90pm), ber 1sB0) (Table 1). Conversely,the best phenologic statefor and 5 (t.ss to 1.75pm)to separateforest from nonforest as the classificationof trembling aspenis between trembling as- well as to stratify forestedregions into conifer, hardwood, pen leaf flush (first hardwood tree speciesto leaf out in and mixed conifer-hardwoodclasses (Figure 2). For northern spring) and leaf flush of other associatedhardwood species temperateforests, these spectralregions possesspractically such as sugar maple which leafs-out about one week later all the information contained in TM data. and afford the best (Sayn-Wittgenstein,1961). This condition was best met with symmetry between classificationaccuracy and processingef- sceneID 52s9216767acquired on 10 May 1992 (Table 1). ficiency (Nelson et al., 1984;Horler and Ahern, 1986; Moore Field verification in the north-central portion of the study and Bauer,1990; Bolstad and Lillesand,1992). While rv area on this date (approx. 48 km south of the northern edge bands 4 and 5 provide for the best discrimination among for- of the study region) revealedthat trembling aspenleaf flush est types, a visible band is also necessaryfor discriminating had begun while sugar maple had not. Although aspen and forestedfrom nonforestedtypes (Hopkins et al., L988). maple phenology were not observedin the southern portion We stratified nonforestedareas from forestedcover types of the study area on this date, reflectancevalues (10 May by applying a threshold classificationalgorithm on TM bands 1992 MSS)of known sugarmaple dominated stands in the 3, 4, and 5 (Figure 2). This method of classificationis similar southern region Ied us to believe sugar maple leaf flush had to a knowledge-basedclassification technique using TM phys- begun. Finally, a winter scene(ID 51.456't6221,25February ical principles describedby Civco (1989).Forested cover

LL32 I

PEER.REVIEWED ARIIC!E

lmageryUsed ThematicLayer Derived ProcessDescription

ThreshofdTM bands 3,4, and 5 basedon averageminimum and maximum refleclancefor coniferand hardwoodstands.

ThrEsholdTM bands3, 4, and 5 basEdon averagEminimum and maximum reflectance valuesfor mixedlorest slands.

Usingthe AndersonLarel ll TM classification, mask out all cover types frorn the Oct. NDGI image,except hardwoods,then thresholdthe NDGI imageto isolateoaks.

Mask out all non-oak types from Oct. MSS bands 1,2, and 4. Then classifynorthern red oak and northernpin oak using a maximum likelihood classificalion algorithm.

Mask all non-hardwoodtypes and oaks from ths Sept. NDGI image. Subtraa Sept. NDGIfrom the June NDG|to highlighlblack ash stands. Then thresholdto the differenceimage to classilyblack ash stands.

Mask non-hardwoodlypes, oaks, and black ash from the May NDVI image. SubtractJune NDVI fromthe May NDVlto highlighttrembling aspen stands. Then thresholdthe differencsimage to classifytrembling aspen stands.

Mask non-coniferfrom the Feb. NDVI image. SubtractFeb. NDVItrom the June NDVIto highlighttamarack stands. Thresholdto the differenceimage to classilytamarack slands.

Mixedcover types containinghardwood or coniler TM Juno lg87 Jadrpinc-elq MSS Ocl. 1980 pin oakpine, componentswith uniquephenology were left out MSS May 1992 ud of the aboveclassilication steps. Therefore, MSS F€b. 1988 blech a!h. procedures MSS S€p. 1985 bwland onilcr differencingand thresholding wsre repeatedlor mixedtorest types.

Remainingforest cover types not classiliedusing t4 Junc 1987 muhi-temporalimage analysis techniques were TM 2-94-5 stratitiedusing a maximumlikelihood classification algorithm.

Figure2. Diagramdescribing each step of the forestclassification, imageryused, and intermediatelayers generatecl.

types have relatively high reflectance values in the near in- reflectancein the red spectralregions compared with most frared, moderatereflectance in the middle infrared, and low nonforestedareas. However, some nonforestvegetation (e.g.,

PE&RS PEER.REVIEWED ARTICIE

Sphagnum) has reflectance values in the near infrared spec- comparisonsbetween datesof greatertemporal difference tral region similar to some forest species (Vogelmann and than 6.6 years are made (Figure 2). Moss, 1993). In contrast, middle infrared and visible reflec- Vegetationindices derived from the MSSdata and the tance tend to be lower for Spfiognum than for forested cover fune rM data are systematicallycombined utilizing subtrac- tVDeS. tion to highlight and classify specific forest cover types. For Spectral differences among these major classes (forest example, during early Septemberblack ash is the first hard- and nonforestJ permitted the use of a general rule: wood type to lose its leaves.The Nnvl derived from satellite data gatheredshortly black IF pixel reflectancein rrll band 3 is low (betweenlower and up- after ash leaf drop exhibits lower per thresholds),very high in't'ru band 4 (betweenlower and index values for black ash than other forest types. Classifica- upper thresholds),and moderatein rv band 5 (betweenlower tion of black ash using one date of imagery would be diffr- and upper thresholds),THEN the pixel most likely represents cult because(1) defoliated black ash standsand nonforested Iorestcover. wetlands are spectrally similar and (2) summer black ash standsand other hardwood cover types are spectrally simi- Bands 3, 4, and 5 average minimum and maximum reflec- lar. To enhanceblack ash stands,then, we subtract autumn tance values for 30 forested training areas containing known NDVIimage (leaf-offblack ash) from a summer NDVrimage conifer and hardwood stands were used to determine thresh- (leaf-onblack ash). High index values (summer leaf-on maple old values for the forest-nonforest classification. Once classi- or aspen)minus other high index values (autumn leaf-on ma- fied, a qualitative visual assessment of the classification was ple or aspen)results in a very low to negativedifference. On performed using 6 19BB color-infrared NHap. The dis- fune the other hand, a high index value (summer leaf-on black tinction between forested and nonforested cover rvDes was ash) minus a low index value (autumn leaf-off black ash) re- good. Confused cover types included sparsely wooded areas sults in a medium difference.A threshold applied to this dif- such as shelterwood clearcuts, apple orchards, and forested ferenceimage classifiesblack ash, areas flooded by beaver activity, all of which were classified into nonforest. We stratified forested land into conifer, hardwood, and ForestClassification mixed cover types (Anderson Level II) by again applying the DNRand USFSstand maps were used in combination with threshold classification algorithm on TM bands 3, 4, and 5 field observationsto verify that senescentforest types ob- (Figure 2). Hardwood tree species have greater reflectance served in the OctobertrlSS imagery were indeed northern red than conifer species in each of these spectral regions (Vogel- oak and northern pin oak (Q. ellipsoidalis) (Plates 1a and mann and Rock, 19BB) with near infrared providing the best 1b). Basedupon comparisonsmade with SeptemberMSS im- separability(Benson and DeGIoria.1985: Shen et ol.. tgas). agery,the October MSSimagery, and forest stand maps, we Shen ef o1. (rses) suggestedthat a threshold performed on determined that the oaks were the only hardwoods still hold- TMS near infrared reflectance would discriminate between ing their leaves.Because leaf-on and leaf-off stands of trees hardwoods and conifers. However, mixed conifer-hardwood have very different reflectancevalues in the near infrared stands have intermediate reflectance in Ttr,lbands 3. 4. and 5 and visible portions of the electromagneticspectrum (Wil- relative to pure stands. Therefore, we selected 30 mixed con- liams, 1975,Vogelmann and Rock, 1989),a vegetationindex ifer-hardwood forest stands to determine thresholds for the was chosento discriminate oaks from defoliated hardwoods Anderson Level II forest classification. Only those stand [e.g.,sugar maple, aspen, birch, and black ash). maps which corresponded to the same year as our TM data We separatedboth oak species(red and pin) from other wete used. Assessment of the hardwood, conifer, and mixed hardwood cover types by first masking all but pure hard- forest classification precision was performed qualitatively by wood forest types from the OctoberMSS data using the June visually comparing the classified data with independent TM hardwood, conifer, and mixed forest classification(Figure stands identified on both NHAP and forest stand maos. Corre- 2). This method of masking the October MSSdata (approx. spondence between ground truth information and the three 6.6 years older than the TM data) assuredthat most hard- coarse forest classes was very good. wood forest types between the two dateswere unchangedin We then classified nonforested areas using the unsuper- terms of dominant forest species.For example,what were vised classification algorithm ISODATA(ERDAS, 1991). The mature oak dominated standsin 1987 most likely were ma- resulting 50 classeswere visually interpreted using 6 fune ture oak dominated stands in 1980. Furthermore, clearcut oak 1g88 NHAP and recoded into eight classes (urban-pavement, standsidentified in 1980 would not have regeneratedback to cleared forest, sparsely stocked forest, flooded forest, shrub- oak sufficiently enough to be classified as mature hardwood herbaceous, grass-forb, Sphagnun spp., and open waterJ (Ta- torest in 1-987.Once masking was completed, we applied a ble 2). No further division of the nonforested classes was threshold to the normalized difference greennessindex (NlcI) pursued. image derived from the October vSS data to classify oak stands(Figure 2). Here the term "greenness"refers to the use of visible greeninstead of visible red reflectance:i.e., Multi-Temporallmage Classification Overview The remainder of the forest classification relies predomi- NDGr= [(MSS4- MSSI)/ (rvrss++ MSSI)+ 1] x 100 (1) nantly on layered image classification techniques (Figure 2). The 14 June 1987 TM image is the base image for this classi- We chosethe MSSgreen band (0.b0 to 0.60pmJover the fication. The greatest difference in image date relative to the MSSred band (0.60 to 0.70pm) for this vegetationindex be- base date is roughly 6.6 years (B October 1980 to 14 June causethe red band of the OctoberMSS data was corrupted by 1SB7) (Table 1). The greatest absolute difference in image a striping pattern that was not entirely regular. Upon visual date is approximately 11.6 years (Table 1). The layered clas- inspection, the MSSgreen band from this date had noticeably sification techniques described in this paper, at most, com- fewer problems of this nature. Becausegreen reflectanceis pare only small portions of data from one MSS date at each strongly correlatedto red reflectance(Badhwar and Hender- decision step to the base TM image (Figure 2). Therefore, no son, 1982; Badhwar et ol., 1,984;Hall et al., 1SS1),the infor-

LL34 PE&RS I

PEER.REVIEWED ARIICTE

Plate1. (1) are 14 June1987 rM data (bands5-4-3), column (2) areN4sS data (bands4-2-L), and column (3) are classifieddata. Row (a) showsmid-fall (8 October 1980) northernred oak senescenceas yellow(a2) and classifiedN. red oak as dark orange.Red oak standsare visually indistinguishable from surroundingsugar maple dominatednorthern hardwood stands in the TMdata (a1).(Row (b) showssenescent N. pin oak as brownish-yellow(b2) and classifiedN. pin oak as cyan(b3)' Row(c) showsthe locationsof three black ash stands highlightedby white lines representing summerleaf on blackash (c1),early fall (13 September1985) leaf-offblack ash standsas black(c2), and classifiedblack ash standscolored dark blue (c3).Row (d) showstrembling aspen and sugarmaple-dominated northern hardwood stands in sum- mer (d1),mid-spring trembling aspen leaf-flush (10 May 1992) as red (d2),and classi- fiedtrembling aspen stands in greenand northernhardwood stands in yellow(d3). Row(e) showssummer leaf-on tamarack as maroon(e1), winter (2 February1988) leaf-offtamarack as black (e2), and classifiedtamarack as dark red (e3).

mation derived from the NDGIwas expectedto be compara- checked against forest stand information which revealed ble to the information provided by the mvr (Equation 3). good discr-imination between the oak dominated stands and This intermediate classificationwas then qualitatively other hardwood stands.

PE&RS 1135 PEER.REVIEWED ARIICTE

Senescentnorthern red oak stands (yellow when bands ary MSSNDVI) (Figure 2). The tamarackclass was then over- 4,2, and 1 are displayedin RGB)were clearly distinguishable laid onto the conifers classof the master TIUclassification. from northern pin oak stands (brownish yellowl in the Octo- By applying image differencing techniques to o:rly pure ber MSSimageiy (Plates1a2 and 1b2, respectively).Separa- hardwood-and pure Coniferstands, remaining mixed conifer- tion of the two oak speciesfrom each other was accom- hardwood covei types which contained oaks,black ash, and plished by performing a supervisedmaximumlikelihood tamarackin combination with other specieswere missed. ilassification on the October vtSSdata (bands 1',2, and 4) Therefore,image differencing procedureswere repeatedsepa- with 15 training samplesfor each species(Figure 2). The red rately for the mixed hardwood-conifertypes (Figure 2). The oak and pin oak types were then overlaid onto the hard- resulting classifieddata [jack pine-oak, pin oak-pine,and class of the rv classification. black asl-lowland conifer) were then overlaid onto the Attempts to classify sugar maple dominated northern mixed conifer-hardwoodclass of the master ttrl classification. hardwoodi using autumn leaf color were not successful.Se- Remaining forest cover types (red pine, iack pine, black nescencein sugar maple had not progressedsuffrciently to spruce, white spruce (P. glauca),mixed swamp conifers,- be distinguished from other hardwood speciesin the 12 Sep- white pine (P. itrobus)-hardwood,balsam fir-aspen,hemlock- tember rbss N4ssimagery. However, field checking revealed yellowbirch (Tsuga canadensis-&.dleghaniensis), white-ce- were separatedusing that defoliated regions in the SeptemberMSS imagery were - dar-hardwood,and paper birch-conifer) black ash standsihat had dropped their leavesprior to Land- bands 2, 3, 4, and s-of the lune tv image employing tradi- sat overpass(Plate rc). To separateblack ash from remaining tional iterative supervisedtraining and classificationtech- hardwood cover types, all non-hardwood types and oaks niques (Figure 2). Training information was gathered_by were masked from the SeptemberMSS data (Figure 2). The ground-basedsampling and from DNRand usFSstand maps. black ash type was then classifiedas describedabove by up- Fifteen training polygons per remaining classwere used in plying a threshold to a differenceimage (JuneTtr'l NDGI minus this classification. SeptemberMSS NDGI): i.e., AcculacyAssessment rM NDGr: [(rrvr+- rM2) I (rxrS+ rM2) + 1] x 100 (2) Accuracy assessmentof the final classification(Tables 3 and 4) was performed using usFS and oNR stand information. The MSSdata, the SeptemberMss red band had Like the October stand information was preferred as referenceinformation for than did the green band; therefore,the more sensornoise classificationvalidation becauseit contains no bias that the the vegetationin-dex. Classified green band was used for investigators might introduce if conducting their own recon- onto the hardwoods class de- 6lack ash was then overlaid naissance(Bryant et al., 1,g1o).DNR and usps tabular forest data. The remaining hardwood standsto rived from the TM stand informition (rsas-r9BB),independent of dataused for sugarmaple dominated northern hard- be classifiedwere training, was randomly sampled to decide which forest and mixed aspen. woods,trembling aspen, standswould be used as referencedata for the classification made in the field at the time of Landsat Observationi accuracyassessment. of the tabular data were made the 10 May 1992 MSSimage was Queries overpassconfirmed that for each forest cover type based on primary type, secondary leaf flush (Pbte 1d). To sepa-- highlighting trembling aspen type, height, basal area,and harvest year.Tor example,five hardwood cover types, all rale tr6mbling aspenfrom other queries were used to select suitable mixed white pine-hard- oaks, and black ash were masked non-hardwoo-dfoiest types, wood stands.The first query selectedall white pine stands data. Trembling aspenwas separatedfrom from the May MSS with oak as a secondaryforest type. two added to the speciesby applying a Query sugar maple and other hardwood first query all white pine standswhose secondarytypes were image (May MSSNDVI minus thieshold to a difference |une eithei aspen,paper b1rch,or sugar maple dominated north- NDVI)(Figure 2). NDVIrather than NDGIwas used for the TM ern hardwoods. three selectedfrom the result of que- the red band from the May MSSim- Query difference image because ries one and two all standswith basal areas) 16.09 m'/ha. agedid not exhibit serioussensor noise (striping):i.e.' Query four choseflom the result of queries one through were at least 9.15 m in height. five MSSNDVI : [(MSS4- MSSz)/ (MSS4+ MSS2)+ 1] X 1OO (3) three standsthat Query ensured that the resulting stands from queries one through - - (ttvrS+ rM3) + 1] x 100 (4) rM NDVr l(rrras rM3) I four were uncut at the time of sensoroverpass (t+ June The trembling aspentype was then overlaid onto the hard- 19871. woods class of the rn classification. Once all potential reference stands were tagged, a ran- Sugarmaple dominated hardwoods and mixed aspen dom numbers generatorwas used to selectthose standsthat were the only hardwood types left to classify.Thirty training would be used as referencedata. When the use of usps and polygons pei type were used to train a maximum-Iikelihood DNRforest stand information was inappropriate (e'8.,flooded ilaislfication ofru bands2,3,4, and 5 (Figure2). Sugarma- forest,urban or pavement,Sphagnum, etc.), sites were ran- ple and mixed aspen were then overlaid onto the rM hard- domly selectedfrom interpreted aerial photography (6 June - woods class,thui completing the classificationof hardwood 19BBNHAP) or field checked (Table 2). A minimum of 30 ref- forest cover types. erencesample sites (greaterthan 2 ha per site) for each clas- Becausetamarack is a deciduous conifer, winter NDVI sified cover type was selectedwith the exception of jack (zo paper birch-conifer (zo sites).Individ- values were expectedto be lower than other conifer !yp"t. pine-oak sites)and Stand information and field observationsconfirmed that leaf- ual sites consistedof severalpixels of classifieddata rather off tamarackstands were visible in the FebruaryMSS imagery than single-pixel samplesas recommendedby Roller and (Plate 1e). Therefore,all but pure conifer standswere Visser(1980). masked from the fune TM and FebruaryMSS data using the TM conifers classas a template. Tamarackwas separated Resultsand Discussion : from black spruce and other coniferoustypes by applying a The overall classification accuracy was 83.2 percent (xunr threshold to a differenceimage (JuneTvt NDVIminus Febru- 82.5) (Table 3) while the accuracy for the forest classes was

1136 PE&RS PEER.NEVIEWED ARTICTE

or o, < O N N nco O < < do q o N N o @ o i n s rj oi d.j e q 6ad ( o - . N 6 b cn d co - co co 6 n o o to o 6 o o N co o N o o o i D< i - -- ill

:Fr 3 Y-

:.i

F o o

N €i o + +€

I

MN

: , p 9;O ..6 (oi o= I oi F id Nfi N- E id I dN Ndin N@- x N6 E ^u

k o6 N c? 1F s i-

E N E NN N Nd Nq E n IJ 6 Nii d oi crt o @ U N 6 @ hd

+\ $N o N N dd NN ,J o VN

O ^ro o)N

Nod) NN no o N

iN

+inN @- X L >'-' 6nd C4i;F

tL37 PE&RS PEER.REVIEWED ARIICTE

(O '- O O) sl O N N i 6 O sl stt i O c| c) , $

(') (O 3(d sl o)

Nts $ o

O mts N d

N i FN F d

d N(o c-,1 @. d o $i U J

F o T z N N U z tr N U o)oN n iN N^ o NY z -^N tr o? tl N. NE: E @ o dN Ntsid N oqv

E i cnlo N 6 (J *tl z .>: o q.i U NN N Ni !e F h a ffi U O G co< E Ndd cobFO x E F o bi g o ,;

E U A s + U J @ d- NN N ; F

Ni sl o? N 6

b co 6i N o b

NOO) NN o o 6 d N

HN .] o

N

PEER.REVIEWED ARIICTE

TABLE5. Ennon MnrRtxFoR ANDERSoN LevEr ll FonEsrCrnssEs. between1987 and 1992on public lands alone.Initially, we thought the NDGIdifference image (May MSSNDGI minus June Reference Data trembling aspen standsfrom the re- Row User's TM NbcI) used to stratify Classified cover types (mixed aspen and su8-ar.ma- Data Conifer Hardwood Mixed Total acculacy maining hardwood ole dominated northern hardwoods) had missed trembling Conifer 260 10 270 96.3% ispen standsharvested after 14 June1987' Closer inspection Hardwood 256 10 266 96.2o/o of the NDGIdifference image and the raw 1992 May MSSim- Mixed 18 1.7 287 322 89.1o/o agery revealedthat trembling aspen-stands harvested be- 858 Col. Total 278 273 307 trireen1987 and tg8g had apparently regeneratedsufficiently Producer's Diagonal enough to be classifiedas trembling aspenby the difference Accuracy 93.5% 93.8% 93.5% total: 803 imag6 technique.It was not possibleto determine ftom the Overall accuracy93.6% KHAT : s0.a Uay vss data where trembling aspenharvesting operations within this time period had occurred. Trembling aspen stands cut later (between1990 and 1992) were progressively 80.1 percent (rcuar : 76.0) [Table 4) and, for Anderson more distinguishablein the May 1992 MSSimagery. Obvi- Level II forest classes,93.6 percent (KHAT : 90.4) (Table 5). ously, most of the problems associatedwith ha-rvestingoper- The greatest amount of confusion occurred between for- ations would have been avoided if the tvtssand ttvl data were est types where only TM data were used for classification all from the sameyear. Imagery acquired within a single-year (e.g., black spruce, white spruce, mixed swamp conifers, would also eliminite any forest successionaleffects' Hall white-cedar-hardwood, balsam fir-aspen, and white pine- (1991)found that forest successionwithin time spansas lit- hardwood) (Tables 3 and +). The poor discrimination be- tle as 10 years can be significant fiom a remote sensingper- tween the black spruce type (user's accuracy 61 percent and soective. producer's accuracy 74 percent) and the mixed swamp coni- In addition to problems associatedwith harvestingoper- fer type (user's accuracy 71 percent and producer's accuracy ations, trembling aipen leaf flush apparently had not prog- 52 percent) may be because mixed swamp conifer types ressedfar enough in the extreme northern portion of the within this region are made up of predominantly black study region to-be detectedby the MSSsensor and was also spruce and balsam fir with associates of white-cedar and miss-edusing the NDGIdifference image technique, Ulti- ^^ timarack. Beaubien (1979) found black spruce and balsam fir mately, trembling aspen standsmissed using the Nnct differ- were very similar in terms of vss spectral reflectance, except ence image approach,because of harvestingoperations when balsam fir grew in older, pure stands. When black betweenisei hnd 1992or becauseof delayedphenology, spruce and mixed swamp conifer types are combined, the ag- were classified(maximum likelihood) using the 1'4June 1987 giegated accuracy becomes 81 percent for both user's and TM image(bands 2,3, 4, and s) into either sugar-mapledom- producer's while overall forest cover accuracy increases from inated northern hardwoods or mixed aspen stands.This was b0.1 percent to 83.2 percent (Table a). Error associatedwith a relatively safe assumptionbecause oak standsand black confusion between lowland conifer classes (i.e., black spruce, ash standshad been stiatified prior to the trembling aspen tamarack, mixed swamp conifers) and upland conifer classes step' of the classification(Figure 2). (i.e., red pine, jack pine, etc.) could be resolved with the in- Using the TM data alone to classify the remaining hard- corporation of digital wetlands data (Polzer, 1992) or digital wood stands as well as the problems associatedwith using to some of soil information [Bolstad and Lillesand, 1992). the 10 May 1992 MSSdata most likely contributed Substantial etrors also occurred between the white pine- the error r-eportedfor trembling aspen,mixed aspen,and hardwood type and balsam fir-aspen type (Table 4). Some of sugar maple becausethese types are notoriously similar in the error is due to the fact that image differencing proce- ter-msof iu spectral and radiometric resolution. Other cover (pap.e.r dures used to classifu balsam fir-aspen types were aban- types in which trembling aspen standswere-confused doned. This procedure was unsuccessful because the May biich-conifer, balsam fir-aspen,and one black ash stand) did trembling aspen leaf flush signature was eclipsed by the have a fair amount of trembling aspenwithin them except (trem- more dominant balsam fir signature. Therefore, classification for the black ash stand. When the two aspen classes the aggregated of this type was performed using the June TM data alone. Of bling aspenand mixed aspen)are combined, the 44 sites which indicated white pine-hardwood in the ref- ,rc"ft uttd producer's accuraciesbecome 89 percent and 79 erence data, 21 were misclassified. Nineteen of the 21, omll- ^percent, reipectively. ted sites went to balsam fir-aspen (Table a). But, of the 53 Higher ilassification accuracyresults were obtained for black balsam fir-aspen reference sites, 12 were inconectly classi- pin oaf (100percent user's and 82 percentproducer's), fied though none were omitted as white pine-hardwood mix' ish (aa p"."".tt user's and 98 percent -producer's),tamarack Six were omitted as jack pine-oak, one as pin oak-pine, three (os percbntuser's and 82 percentproducer's), black ash-low- as trembling aspen, and two as black ash. The misclassifica- land conifer mix (92 percent user's and 86 percent produ- trembling aspen cer's),red oak (Bs peicent user'sand B7 percentproducer's), tion of two black ash stands and three q? stands into mixed forest categories indicates errol with the and pin oak-pine mix (e0 percent user's and p-ercent initial Anderson level II forest classification [Table s)' producer's) (iables 3 and +). Overall, the results for red oak Accuracies for forest cover types classified using multi- ilassification are good, although some problem areaswere temporal image analysis are highest for non-aspen forest- noticed. Red oak classificationprecision was lower along the types (Table +). Trembling aspen stands (36) were classified northern tier of the study area due to delayed senescence wilh eO percent accuracy (Table +). But, of the 42 trembling causedby the temperaturebuffering effects-of Lake-Superior. aspen stinds selected from the reference information, only For example,when sugarmaple, aspen,and paper birch . - with 74 percent were correctly classified. Upon cbecking DNR and leaveshave fallen farther to the south, as was the case uSFS tabular stand information, we learned that over 900 as- the OctoberMSS data, oak standsadiacent to Lake Superior pen stands in the study region were scheduled for harvest remain fully green.At the sametime sugarmaples, aspens,

1139 PE&RS PEER.REVIEWED ARIICTE

and paper birch trees adjacentto Lake Superior are just be- would have improved precision of multi-temporal image ginning to show signs of fall color. The temperaturebuffering classifications.One obvious advantageof using multi-tempo- effectsof Lake Superior extend inland roughly 5 to 10 kilo- ral tv over MSSdata would be the potential for more accu- metres,forming a gradient of senescence.Within this buffer rate image coregistration.However. we question whether zone, discrimination between red oak, sugar maple, and as- repeatingthe proceduresusing exclusively TM data would pen types was poorer than the overall classificationprecision have increasedclassification precision enough to justify the suggests.Although no attempts were made to isolate and greatercost over MSSdata. Unfortunately, the MSSsensor quantify these effectson the classificationwithin this buffer was turned off on 19 October 1992. ending its life as both an zone, theoretically, digital climate date could have been used effective and affordableresource assessment tool. to addressthis problem (Hostef o1.,submitted ms., 1994.). Forest cover types classifiedusing only TM data exhib- ConclusionsandSu$estions fu Future Research ited mixed precision results (Tables3 and 4). The user's and Distinguishing among deciduous forest types in the Great producer's accuracy for red pine reached B6 percent and 94 Lakes region, especially the so-calledsugar maple dominated percent, respectively.Some errors of commission occurred northern hardwoods, has been very difffcult using single-date between red pine plantations with high basal area (2 62 mr/ image classifications.Using a layered, multi-temporal image ha) and white spruce plantations of similar density. Both of classificationapproach, we were able to separatetwo oak these forest types are characterizedas having very dark un- species- black ash and tamarack- and, most importantly, derstoriesdevoid of ground vegetation.On the other hand, separateaspen types from sugarmaple-dominated northern jack pine, sugar maple, hemlock-yellow birch, and paper hardwoods. It is apparentthat a layered multi-temporal ap- birch-conifer exhibited only moderateprecision with user's proach to the classificationof Landsat data, combined with a and producer's accuraciesof 79 percent percent, and 91 83 specilc knowledge of cover-typephenology, is not only -pos- percent and 84 percent, 87 percent and B0 percent, and 76 sible but is preferableto single-dateclassifications or to g1 percent and percent, respectively.Furthermore, white multi-date classificationswhere only a broad knowledge of spruce (user'saccuracy 91 percent and producer's accuracy forest phenology is incorporated into image acquisition. Us- 54 percent) and white-cedar (user'saccuracy 100 percent and ing a layered multi-temporal image classificationapproach, a producer's accuracy67 percent) exhibited poor agreement specieslevel forest classificationwas achievedwith an accu- with referencedata. Oddly enough,white-cedar did not have racy of 80.1 percent (rrnr : 76.0).Accuracy for forest clas- any errors of commission with lowland conifer types, al- sesaggregated to Anderson Level II (hardwood, conifer, and though severalerrors of omission with black spruce and mixed) was 93,6 percent (rsar : 90.4). Overall classification hemlock-yellow birch did occur. The lack of commission accuracywas 83.2 percent (KHAT: 82.b). errors with lowland conifer types is puzzling becausethe By incorporating specific knowledge of forest species black spruce and hemlock-yellow birch types within this re- phenology, it is possible to gion have often associatesof white-cedarwithin them and o vice versa. Develop a forest classification with dominant tree species level precision within northern Lake States conditions, Some of the forest cover tvpes not directlv classifiedus- . Use MSS digital data to capture specific phenology of forest ing ancillary MSs data most liiely improved in classification cover types, precision becausethey were adja-cent^toforest cover types o SuccessfuIly incorporate multi-temporal uss and TM data for that were classifiedusing multi-temporal image data. For ex- detailed forest classifications, and ample, classificationof the sugarmaple dominated northern . Use a layered classification approach exploiting image ratio- hardwood type was simplified becauseadiacent stands of red ing and ratio differencing techniques for multi-temporal im- oak, pin oak, and trembling aspen were subtractedfrom the age analysis. greaterhardwoods type, leaving fewer hardwood types with There clearly are advantagesto this layered, multi-tem- which sugar maple could be confused.Table 2 lists the five poral classificationmethod where phenological changesoc- forest cover types that benefittedfrom this indirect multi- cur acrosslarge regions.In many instances,spectral variabil- temporal image classificationmethod. ity within a single forest type over large regions is great due It is likely that some of the within-class heterogeneity to the effectsof atmosphere,soil, climate, and aspect.To problems, which have been the bane of many forest classifi- gather enough training statisticsto adequatelyaccount for cations in this region using TM data, were reduced by utiliz- these types of variability is a difficult task. By using multi- ing the spatial resolution of vss data. The 79-m, radiative temporal image ratioing and ratio differencing techniques, input of an MSSpixel sufficiently generalizesspatial and many of these effectsare normalized, and comparatively few spectral cover type characteristicssimilar to the way in training statisticsare necessary. which a photo-interpreterallows for some degreeof within- Although the classificationtechniques presented in this classheterogeneity when delineating cover-typeboundaries. paper generally worked well, there is potential for improve- Toll (1985)alluded to the spatial and spectralgeneralization ment and refinement.First, imagesfrom the same year or a properties of tr,tssdata when studying sensorparameters re- short span of years will work better when using the forest sponsible for differencesin TM and MSSclassification accura- phenology approach for forest cover-typeclassihcation. Con- cles. temporaneousimagery will minimize or eliminate problems Furthermore,other studies suggestthat classificationac- associatedwith forest harvestingoperations and eriors asso- curaciesare likely to degradeas a result of improved spatial ciated with forest succession.Second, incorporation of digi- resolution while other sensorparameters are kept constant tal National Wetlands Inventory information or digital soils (Townsendand Justice,1981; Toll, 1984;Latty ef d1.,19BS; information (soil series)would help resolve errors between Martin et o,1.,19BB; and Moore Bauer,1990). Because it was lpland and_lowland forest classes.These types of ancillary only necessaryin most instancesto detect leaf-on versus data, though not available to date, will be available for thii Ieaf-off vegetationstatus using the MSSdata, it is doubtful region in the near future. Third, variations in forest phenol- that the added spectral and radiometric resolution of lra data ogy within large regions remain somewhatproblemalic. The-

LL4O PE&RS PEER.REVIEWED ARIICTE

and Mapping,2-7 April (Balti- oretically, Iarge scale (full ru scene)forest classifications American Congresson Surveying 276-291'. employing digital climate data could be used to stratify these more, Maryland), PP. University of effects.Finalty, the use of the more expensivetM data in Curtis, J.T., 1959.The Vegetationof Wisconsin, 657 place of MSSdata could provide some improvement in classi- Wisconsin Press, P. Scanner fication results due to the potential for refinementsin image Downs, A.L., 1S81.A Comparisonof Landsat Multispectral Coior Infnred Photography Type Map- coregistrationaccuracy. However, besidesregistration im- Data and Digitized for ping EloquetForestry Center,PIan B' Paper,University of previous studies show that the 30-m spatial res- at the provements, Minnesota,St. Paul,47 p. olution of tv data is responsiblefor only slight increasesin Don't unless it's autumn, of Fotestry' accuracybetween MSSand TM data. Therefore, Eder, I.J.,1989. loutnal classification 87(6):50-51, from an economic standpoint, using multi-temporal TM data ERDASField Guide, Version 7'5, ERDAS,Inc' Atlanta, rather than MSSdata (using the same methods) may not pro- ERDAS, 7591. justify Georgia,394P. duce results that the added cost. tlnitcd States Eyre,' F.H. (editor), 1980.Forest Cover Typesof the and Canado,The Society of American Foresters,Washington' Acknowledgments D.C.,1aB p. has been funded by a U.S. Forest ServiceEcosys- This work Goodwin, T.W., 1S58.Studies in cartenogenesis.24' The changesin tem ManagementCooperative Agreement with D. ). Mladen- carotenoidan chlorophyll pigments in the-leaves of deciduous off and G. E, Host, NRzu-UMD,and T. R' Crow, USFS.We trees during autumn, Biochemistrylournal, 68:503-511' from David Verbyla, appreciatecomments on this manuscript Goward, S.N.,C.J. Tucker, and D'G' Dye, 1985.North American veg- fames Vogelmann,leff Eder, Phil Polzer, Mark White, and etation oatternsobserved with the NOAA-7 Advanced Very three anonymous reviewers.This is Contribution Number High ResolutionRadiometer, Vegetatio, 6413-14' Number 1,27of the Natural ResourcesResearch Institute, and Hall, F.G.,D.B. Botkin,D.E. Strebel, K.D. Woods,and S'J'Goetz,- . 28 of the Natural ResourcesGIS Laboratory, University of 1991..Large-scalepatterns of forest successionas determinedby Minnesota, Duluth. remote sensing, Ecology,7 2(2):628-64O. Hixson, M., D. Scholz,N. Fulls, and T' Akiyama, 1980' Evaluation of severalschemes for classificationof remotely senseddata, Pfto- -1553' Refelences togrammetric Engineering & Remote Sensing, 46(1'2):7547 P.F.,A.L. Maclean,and T.M. Lillesand,I'g8B' Assessment Ahlgren, C., 1.557.Phenological observations of nineteen native tree Hopkins,' Mapper imagery for forestry applications under species in norlhern Minnesota. Ecology. 38:622. of Thematic Lake States conditions, Photogrummetric Engineeting & Remote E.E. Hardy, Roach, and R'E. Witmer, 1'976'A Anderson, 1.R., J.T. Sensing,54(1 ):61-68. Land use and Land Cover Classification System fot Use with Re' and F.J.Ahern, 1986.Forestry information content of mote Sensor Dafa, U.S. Geological Survey Professional Paper Horler, D.N.H., Thematic Mapper data, International lournal of Remote Sensing, 964, Washington, D.C., 28 P. 7(3):4o5-428. Badhwar, G., and K. Henderson, 1s82. A comparative study of The- E.R.,and B.N. Rock, 1989.Detection of changesin leaf water matic Mapper and Landsat spectral bands from field measure- Hunt. content using near- and middle-infrared reflectance,Remote ment data,-Proceedings of 1982 Machine Processing of Remotely Sensingof Environmenf,30:43-54. Sensed Data Symposium, 7-9 fuly (West LaFayette' Indiana), pp. of leaf 266-272. Hunt, E.R.,B.N. Rock, and P.S.Nobel, 1987.Measurement relative water content by inlrared reflectance,Remote Sensing of Badhwar, G.D., R.B. MacDonald, and F.G. Hall, 1gB4' Spectral char- Environment,22:429-435. acterization of biophysical characteristics in a boreal forest: images,Pro- Relationship between Thematic Mapper band reflectance and Kalensky,2.,7974. ERTS thematic map from multidate '84 Sensingand Ieaf area index for aspen, Proceedings of IGARSS Sympo- of the 7s74 Symposiumon Remote -Photo """hing" 7-11 October sium,27-3O August (Strasbourg,France], pp. 111-115. lnterpietaiion, CanadianInstitute of Surveying, (Banff,Alberta), (7):767-7 8s. Beaubein, Forest type mapping from Landsat digital data, 1.,797s. forest mapping Photogrammetric E n ginee ri n g & Rem ote Sensing, 45 (B): 1 1 3 5- Kalensky,2., and R.R. Scherk, 1975.Accuracy of 't144, from Landsat computer compatible rapes,Proceedings of the 10th internationalsymposium on Remote Sensing of Environ- Benson, A.S., and S. DeGloria, 1985. Interpretation of Landsat-4 The- menf, 6-10October (Ann Arbor, Michigan)'pp' 1159-1167' matic Mapper and Multispectral Scanner data for forest surveys, study: Landsat Photo grammetric E n gi ne eri ng & Remote Sensing, 5 1 (9) :1 2 81 - Kan, E.P.,and F.P. Weber, 1978.The ten-ecosystem in the United States'Pro- 7289. ADP mapping of forest and rangeland ceedingi ;f the flth Symposium on Remote Sensing of Environ- Bolstad, P.V., and T.M. Lillesand, 1992. Improved classification of menf,Ann Arbor, Michigan,pp. 1809-1825' forest vegetation in northern Wisconsin through a rule-based Yegorov,1992- Spectral combination of soiis, terrain, and Landsat Thematic Mapper Kharuk, V.I., A.M. Alshansky,and V.V. factorsof variability, Intema- data, Foresf Science, 38[1):5-20. characteristicsof vegetationcover: tional of Remote Sensing,L3(77):3263-3272' Boyer, M., Miller, M. Belanger, and E.]lare, 1988. Senescence and lournal " I. Williams, D' Toll, and spectral reflectance in leaves of northern pin oak (Quercus pal- Lattv, R.S.,R. Nelson, B. Markham, D. I' between information ex- uitris Muench.), Rentote Sensing of Environment' 25:77-87 ' irons, 1985.Performance comparisons traction techniquesusing varilble spatial resolution data,Phofo- E., A.G. Dodge, and S.D' Warren, 1980. Landsat for Bryant, Jr., grammetric Engineering & Remote Sensing, 51 (9) : 1 1 59-1 1 70' practical forest type mapping: a test case, Photogrammetric Engi' Locey, C., 1990.Foresl CompartmentReconnaissance Handbook neering & Remote Sensing, 46(L2):'1'575-1584' - ialz,Wi""ontin Departmentof Natural Resources Bureau of Buchheim, M.P., A.L. Maclean, and T.M' Liilesand, 1985. Forest Forestry,Madison, Wisconsin, 40 P. Cover type mapping and spruce budworm defoliation detection D'O. Ohlen,and Brown, 1991' using simulated SPOT imagery, Photogrammetric Engineeting & Loveland,T.R., J.W. Merchant, J'F' characteristics databasefor the Remote Sensing,51tBJ:1115-1122. Develooment of a land-cover contetminous IJ.S., Photogtammetric Engineering & Remote 1989. Knowledge-based land use and land cov€r map- Civco, D.L., Sensing,57 (11) :L453 - 1463. pilrs, Proceedings of the 1989 Annual Convention of the Ameri- Hoffer, 1985.Evaluation of a layered Loi Society of Fho[ogrammetry and Remote Sensing and the Lozano-Garcia,D.F., and R.M.

PE&RS PEER REVIEWED ARIICTE

approach for classifying multi-temporal Landsat MSS,Proceed- visualizing forest change dynamics, International lournal of Re- ings- of the Pecora Ten Symposium: Remote Sensing in Forest mote Sensing, 13(16):305S-3067. and Range Resource Management,20-22 August (Fort Collins, Samson, S.A., 1993. Two indices to characterize temporal patterns Colorado), pp. 189-199. in the spectral response of vegetation, Photogramm-etric Engineer- Martin, L.R.G., P.I. Howarth, and G. Holder, 1988. Multispectral clas- ing & Remote Sensing, 59(4):511-b12. sification of land use at the rural-urban fringe using SpOf dutu, Sanger, J.8., LS71,.Quantitative investigations of leaf pigments from Canadian lournal of Remote Sensing,'J.4(2):7Z-7 g, their ince-ption in through autumn coloration to decompo- Mead, R.A., and M.P. Meyer, 1977. Landsat digital data application sition in falling leaves, Ecology, S2:1OTS-1.O89. to forest vegetation and land use classification in Minnesota, Sayn-Wittgenstein, L., l,g6l Phenological Aids to Tree Species Proceedings of the 1977 Symposium on Machine Processing of Identification on Air Photograpis, Technical Note No. 104, For- Remotely Sensed Data, 21.-23 (West LaFayette, Indiana), pp. June est Research Branch, 270-279. Canada Department of Forestry, Ottawa, 26 p. Miller, J.R.,J. Wu, M.G. Boyer, M. Berlanger, and E.W. Hare, 1991. Schwaller, M.R., and S.|. Tkach, 198b. Premature leaf senescence as Seasonal patterns in leaf reflectance red-edge characteristics, In- an indicator in geobotanical prospecting with remote sensing ternational Journal of Remote Sensing, I2(7)IISO9-1523. techniques, Economic Geology, 80:250-2b5. Mladenoff, D.J., M.A. White, J. Pastor, and T.R. Crow, 1993. Shen, S.S., G.D, Badhwar, and Carnes, 1985. Separability of bo- Comparing spatial patterns in unaltered old-growth and dis- I.G. real forest species in the Lake area, Minnesota,- turbed forest landscapes, Ecological Applications, 3(2):254-306. Jennette Photogrammetric Engineering & Remote Sensrng, 51,(I'1,11,77 S- Mladenoff, D.f., and l. Pastor, 1993. Sustainable forest ecosystems in 7783. northern hardwood and conifer region: Concepts and manage- Schriever, J.R., and R.G. Congalton, 1993. Mapping forest cover-tvpes ment, Defining Sustainable Forestry (G.H. Aplet, I.T. Olson, N. in New Hampshire using multi-temporal Landsat Thematic Map- Johnson, and V.A. Sample, editors), Island Press, Washington, per data, ASPRS/ACSM Annual D.C.,pp. 145-180. Convention & Exposition, TS-79 February (New Orleans, Louisiana), j:JJ3-342. Mladenoff, D.f., and G.E. Host, 1994. Ecological applications of re- Singh, A., 1986. Change detection in the tropical forest environment mote sensing and GIS for ecosystem management in the of northeastern India using Landsat, Remote Sensing and Tropi- northern Lake States, Forcst Ecosystem Management at the cal Land Management (M.J. Eden and Parry, editors), Landscope Level: The Role of Remote Sensing and GIS in Re- l.T. fohn Wiley and Sons, New York, 36b p. source Management Planning, Analysis and Decision Making (V.A. Sample, editor), Island Press, Washington, D.C., in preis. Skidmore, A.K., and B.|. Turner, 1988. Forest mapping accuracies are improved using a supervised nonparametric classiffer with Mladenoff, D.J., and F. Stearns, 1993. Eastern hemlock regeneration SPOT data, Photogrammetric Engineering & nemote Sensing, and deer browsing in the northern Great Lakes region: A re-ex- 54(1, 0) 11,41,5 - 1 421.. amination and model simulation, Consewation Biology, 7(a):BBg- 900. Stenback, f.M., and R.G. Congalton, 19S0. Using Thematic Mapper imagery to examine forest understory, Photogrammetric Engi- Moore, K.G., 1965. Senescence in leaves of Acer pseudoplatanusL. neering & Remote Sensing, 56(9):1285-1290. and Parthenocissus fricuspidafa Planch. I. Changes in some leaf constituents during maturity and senescence, Anatomical Bot- ToIl, D.L., 1984. An evaluation of simulated TM data and Landsat any,29i433-444. MSS data for determining suburban and regional land use and land cover, Photogrammetric Engineering & Remote Sensing, Moore, M.M., and M.E. Bauer, 1990. Classification of forest 5O(72):7773-1.724. vegetation in north-central Minnesota using Landsat Multispec- tral Scanner and Thematic Mapper data, Foresf Science,36(2): 1385. Effect of Landsat Thematic Mapper sensor parameters 330-342. on land cover classification, fiemofe Sensing of Enviionment, 1,7: 129-\40. Nelson, R.F., R,S. Latty, and G. Mott, 1984. Classifying northern for- ests using Thematic Mapper simulator daIa, Photogrammetric Townsend, I., and C. fustice, 1981. Information extraction from Engineering & Remote Sensing, 50(5):607-617. remotely sensed data, a user's view, International lournal of Re- mote Sensing, 2(a)SI3-329. Pastor, |., and D.f. Mladenoff, 1992. The southern boreal - northern hardwood forest border, A Svstems Analvsis of the Global Bo- Tucker, C.1., 1980. Remote sensinq of Ieaf water content in the near real Forest (H.H. Shugart. R. Leemans, atia C.n. Bonan, editors), infrared, Remote Sensing of Environment. 1O:23-32. Cambridge University Press, Cambridge, U.K., pp. 21.6-240, Tucker, C.1., J.R.G. Townsend, and T.E. Goff, 1985. African land- Polzer, P., 1992. Assessment of Classification Accuracy Improvement cover classification using satellite data, Science, 227(a69s): Using Multi-Temporcl Satellite Data: Case Study in the Glacial 369-375. Hobitat Restoration Area in East Central Wisconsin, Master's Vogelmann, J.E., 1SSS.Detection of forest change in the Green Thesis, University of Wisconsin, Madison, Wisconsin. Mountains of Vermont using Multispectral-Scanner data, Inter- Potter, 1., and M. Shelton, 1974. Effects of atmospheric haze and sun national lournal of Remote Sensing, S(Z):11,87-I2OO. angle on automatic classification of ERTS-1 daIa, Proceedings of Vogelmann, J.E., and B.N. Rock, 1988. Assessing forest damage in the 9th Symposium on Remote Sensing of Environmenl Ann Ar- high-elevation coniferous forests in Vermont and New Himp- bor, Michigan, pp. 865-374. shire using Thematic Mapper data, Eemofe Sensing of Enviion- Price, J.C., 1987. Calibration of satellite radiometers and the ment, 24:227-246. comparison of vegetation indices, Remote Sensl'ng of Environ- 1989. Use of Thematic Mapper data for the detection of for- ment, 21:1.5-27. est damage caused by the Pear Thrips, Remote Sensing of Envi- -225. Ripple, W.]., 1986. Spectral reflectance relationships to leaf water ronment, 30'.277 stress, Piofogrammetric Engineering & Remote Senslng, 52(10): Vogelmann, J.E., and D.M. Moss, 1993. Spectral reflectance measure- 1669-1675. ments in the genus Sphagnum, Remote Sensing of Environment, Roller, N.E.G., and L. Visser, 1980. Accuracy of Landsat forest 45:273-279. covertype mapping in the Lake State region of the U.S., Proceed- Walsh, S.J., 1980. Coniferous tree species mapping using Landsat ings of the 14th International Symposium on Remote Sensing of data, -Remofe Sensing of Environment, S:1t-26. Environment, Ann Arbor, Michigan, pp. 1b11-1520. Weismiller, R.A., S.J. Kristof, D.K. Scholz, P.E. Anuta, and S.A. Sader, S.A., and J.C.Winne, 1992. RGB-NDVrcolour composites for Momin, 1977. Change detection in coastal zone envrronmenrs,

LL42 PE&RS PEER.REVIETYED ARIICTE

Photogra mmetri c Engine ering & Rem ote Sens-ing,43 (1 2) : 1 5 3 3- Wolter, P.T., 1990.Detection of Moisture Stressin Eastern White 1539. Pine Using a Hand Held Ratioing Radiometer with Landsat TM of New Hamp- Williams, D.L., 1975.Computer analysisand mapping of Gypsy EquivalentBonds, Master'sThesis, University Moth defoliation levels in Pennsylvaniausing Landsaf l digital shire, Durham, New Hampshire. d.ala,Proceedings of the 1975 NASA Earth ResourcesSumey Woodwell,G.M., J.E. Hobbie, R.A. Houghton,J.M. Mellilo, B.J.Peter- Symposium on the Practical Application of Earth ResourcesSur- son,G.R. Shaver, T.A. Stone,B. Moore,and A.B. Park,1983. De- vey Data,9-12lune (Houston,Texas), pp.767-177. forestation Measured by Landsat: Steps Toward a Method,DOE EV10468-1,National Technical Information System,Springfield, Williams, Irons, B.L. Markham, R.F. Nelson, D.L. Toll, R.S. I.R., I.R. Virginia. Latty, and M.L. Stauffer,1s84. A statisticalevaluation of the ad- vantagesof LandsatThematic Mapper data in comparisonto (Received2 June 1993;revised and accepted4 April 1994;revised Multispectral Scanner daIa, IEEE Trans. Geosci. Remote Sensing, 21 April 19ea) GE-22(3):294-3O2. SAOEEP'96 Caf l for Papers

KeystoneResoil, Colorado Aprif28 - May1,1996

Theninth AnnualSymposium on theApplication of Geophysicsto Engineeringand Environmental Problems (SAGEEP), sponsoredby the Environmentaland Engineering Gmphysical Society (EEGS), will beheld at KeystoneResot Colo- rado,28 April - I May,1996. SACEEP is dedicatedto sharingnew applications of geophysicswith thoseworking in the geotechical,hydrogeological environmental and regulatory as well asthe geophysicalprofessions. Leadingoff on Sunday,28 April, will bea shortcourse on EnvironmentalGeophysics and Groundwater Model- ing. Plannedsessions on Monday-Wednesday,29April -l Mayinclude:

Applications: RslatedTopics: Evaluationof Exhistinf Structures PositionSystems (GPS) Predictionof In-SiteConditions DataIntegntion (GIS) location of BuriedObjects RelatingGeochemistry to Geophysical Forensics Parameters Site Chancteristics lxo ContaminantDetection Legal/ProfessionalIssues: Liability TechnologyAdvanceme nls: Certifi cation/Registntion/Licensin g DirectDetection ASTMStandards DataProcessing and Presentation Contracting NewEquipment Software

Technicalpapers on researchon researchand applicationof geophysicalmethods in geotechnicaland environmen- tal problemsare requested for both onl and posterpresentations. One-page abstncts are due by 1 October1995. Extendabstncts of all onl and posterpapers will be requiredand due by 15 January1996 for includsionin the Proceedinglsvolume. '96, Abstncts shouldbe directedto ProgramChairman: Linda Hadley,SAGEEP Geophysical222IEast Street, GoldenCO 80401; phondfax: 303-278-1488.

PE&RS LLt