An Introduction to Digital Methods in Remote Sensing of Forested Ecosystems…myweb.facstaff.wwu.edu/~wallin/envr435/pdf_files/cohen_etal_1996.pdf · An Introduction to Digital Methods

An Introduction to Digital Methods in Remote Sensingof Forested Ecosystems: Focus on the PacificNorthwest, USAWARREN B. COHEN’USDA Forest ServicePacific Northwest Research StationCorvallis, Oregon 97331, USA

JOHN D. KUSHLAWILLIAM J. RIPPLEEnvironmental Remote Sensing Applications LaboratoryDepartment of Forest ResourcesOregon State UniversityCorvallis. Oregon 97331, USA

STEVEN L. GARMANDepartment of Forest ScienceOregon State UniversltyCorvallis, Oregon 97331, USA

ABSTRACT / Aerial photography has been routinely used for

several decades by natural resource scientists and managersto map and monitor the condition of forested landscapes.Recently, along with the emergence of concepts in managingforests as ecosystems, has come a significant shift in

emphasis from smaller to larger spatial scales and thewidespread use of geographic information systems. Thesedevelopments have precipitated an increasing need for

vegetation information derived from other remote senslngimagery. especially digital data acquired from high-elevationaircraft and satellite platforms. This paper Introducesfundamental concepts in digItal remote sensing and

describes numerous applications of the technology. Theintent is to provide a balanced, nontechnical view, discussing

the shortcomings, successes, and future potential for digitalremote sensing of forested ecosystems.

During the past decade, digital remote sensing hasbecome an increasingly important tool for mapping andmonitoring forest resources around the globe. This isdue, in part. to an increasing visibility and understand-ing of remote sensing data, in general, and to the greatlyexpanded use of geographic information systems (GIS).Resource scientists and managers now require spatiallyexplicit vegetation data over extensive geographic areas,which means that traditional field survey techniques,even when coupled with aerial photography, are of lim-ited use. Another important factor is an increased un-derstanding that large-scale monitoring of forest condi-tions is practical only if digital remote sensing isincluded in sampling and mapping schema.

In the past several years, the authors have had numer-ous conversations with forest managers and scientistsconcerning some fundamental issues associated with theuse and understanding of digital remote sensing data.Although there are several texts on the subject (e.g.,Jensen 1986, *Mather 1987, Richards 1993, Lillesand andKiefer 1994), and a rich body of technical literature,

KEY WORDS: Remote senslng; Geographic information systems; For-est management: Ecosystem management; Forest in-ventory

*Author to whom correspondence should be addressed.

Environmental Management Vol. 20. No. 3, pp. 421-435

there is need for a current summary of fundamentalconcepts in digital remote sensing from a nontechnicalperspective. In addition to providing such a perspective.this paper reviews some important research and applica-tions of digital remote sensing in both forest manage-ment and science. For this we focus on the Pacific North-west region of the United States. a region of the globewhere remote sensing has been widely used. Finally, wediscuss several important current and emerging issuesin remote sensing.

Aerial photographs lairphotos) have been com-monly used for decades to assist in the mapping of forestresources (Barrett and Curtis 1992, p. 12). Thus, thefocus here is on other remote sensing data, such asdigital aircraft and satellite images. and nonimagingradiometer measurements. To ease the transition, webegin by comparing digital imagery to airphotos, withthe intent of establishing a baseline for common under-standing. Throughout. we strive to present a balancedperspective, one useful in understanding both the capa-bilities and limitations of t-emote sensing.

General Considerations

Like airphotos. digital images record energy proper-ties at a point in time for a portion of the Earth’s surface,Using different combinations of film sensitivity and fil-

Q 1996 Spnnger-Verlag New York Inc.

422 W. B. Cohen and others

Table 1. Regions of electromagnetic spectrum mostcommonly used in remote sensing and approximatewavelength boundaries and sensors used in detectionof energy in each region

Region Wavelength Sensor

Visible 0.4-0.7 km Reflected solar energy detectedby the human eye, black andwhite panchromatic film,color film, and electroopticalsensors.

Reflected o.;-3.0 pm Reflected solar energy detectedinfrared by infrared-sensitive film (up

to 0.9 pm) and electro-optlcal sensors.

Thermal 3- i and Emitted surface energyinfrared B-14 km detected by electrooptical

thermal sensors.S,licrowave 0.1 mm-l m Emitted surface energy and

reflected ener<gy from“active” microwave trans-mitters detected bymicrowave sensors.

ters. airphotos can selectively record certain wavelengthranges of the electromagnetic spectrum. Digital sensorsalso use filters, but in lieu of halide crystals in a a filmemulsion. they use energy detectors that are similar inconcept to voltmeters. Energy incident upon a detectoris converted to a digital number, commonly 8-bit, butoften 9-. lO-, 12-, or 16-bit. Normally, one detector isdedicated to a single wavelength range, and multipleranges are sensed using multiple detectors. Whereasphotographic film is limited in sensitivity to a narrowrange of the electromagnetic spectrum, digital sensorscan operate in a much wider range of the spectrum(Table 1).

Airphotos have an inherent spatial scale that is afunction of camera focal length and aircraft flyingheight. Although photo scale can be thought of as re-lated to the unit area of the Earth’s surface that can beresolved, resolution of airphotos is also a function ofthe film‘s halide crystal grain size (or film speed). Digitalsensors also have inherent spatial properties, but ratherthan referring to scale, the term spatial resolution or“pixel size” is most commonly used. Digital image spatialresolution refers to the size of the individual physicalsample unit on the ground that is sensed by a givendetector at any instant in time. For example, a resolutionof 10 m means a single digital cell contains integratedspectral information from a nominal 10-m X 10-m unitof the Earth’s surface. Likewise, a l-km resolutionmeans that an integrated signal from a l-km X l-km

area of the Earth’s surface was detected and recordedin a single digital cell.

Atmospheric effects are an important problem inremote sensing. Clouds, haze, and the like contaminateenergy signals from the Earth’s surface. Sensing geome-try is another important confounding factor. Sun angle,topographic variation, and the position of the sensorrelative to these all have the potential to strongly influ-ence the energy sensed. Although atmospheric effectsand sensing geometry are important problems in air-photo interpretation. they are more important prob-lems in digital imagery. The primary reason for this isthat in the former case, as photointerpreters we canbring multiple corroborative sources of information tobear on our interpretations, such as size. shape, shadow,location, and convergence of evidence (Paine 1981). Inthe latter case. we are only now beginning to sufficientlyunderstand the phenomena so that we can developmodels and write computer codes that minimize theireffects.

Image Processing Fundamentals

Basic processing considerations for digital images in-clude geometric correction, radiometric correction, im-age enhancement. thematic classification and relatedprocedures. and change detection. Not all of these pro-cedures are applied for every project, hut an under-standing of these fundamental principles is essential forintelligent use and interpretation of digital images andmaps derived from them.

Geometric Correction

Geometric corrections include compensations fordistortions that prevent images from being used directlyas maps. Sources of these distortions include variationsin sensor altitude, attitude, and velocity, Earth curva-ture, and relief displacement (Lillesand and Kiefer1994). Some distortions are systematic and well under-stood. As such, corrections for these are relativelystraightforward to apply (EOSAT 1994). Nonsystematicand uncompensated systematic distortions are cor-rected by a process that uses a set of ground controlpoints (GCPs) that are located in the imagery (Heardand others 1992). Using the GCPs, a geometric transfor-mation is derived that projects the image into a selectedmap projection. During this process, the image is “re-sampled” by one of several techniques, whereby datavalues from pixel locations in the original image areused to assign values to pixels in the output, “rectified”image (Chiesa and Tyler 1994). Distortions caused bytopographic relief remain generally unaffected unlessa digital elevation model (DEM) data set is used during

rectification. Resampling requires that an output imagepixel size be declared. Commonly, the pixel size chosenis near that of the original spatial resolution, but forsome applications a much different pixel size mightbe declared.

Geometric rectification is inexact. Error tolerancefor geometric rectification is commonly given in termsof a root mean square error, which describes how wellthe transformation fits the GCPs (Jensen (1986). Evenunder the most exacting of circumstances, one can ex-pect to find that a given ground resolution cell (pixel)is displaced at least one cell from its “true” groundlocation, and it often can be displaced several cells. Theimportance of this error is amplified when attemptingto register two or more images together or an image toother map data sets. For example. the greatest problemwe have encountered has been the use of digitized poly-gons derived from airphotos and residing in a federalagency data base in conjunction with georeferencedsatellite images in our data base. With few exceptions.the polygons are shifted and/or stretched so that twoor more completely different cover types are included ina single polygon. Depending on the spectral differencesamong the cover types, the use of these polygons intraining or testing of classification algorithms can heseriously compromised.

RadiometrIc Correction

Digital images are a set of twodimensional rasters ofdigital numbers (DN). The two dimensions. x and y,represent geographic space, and each member of theset consists of recorded electromagnetic energy in agiven wavelength range or band (Figure 1). The corol-lary in airphotos is the three separate, superimposedlayers of color or color-infrared film. Some digital datasets contain only one hand (e.g., panchromatic visible),whereas others may contain over 200 narrow wavelengthbands over the full spectrum of reflected energy fromvisible to short-wave infrared. Energy recorded in a digi-tal image is more than just a function of the covertypes sensed. Factors such as topography, illuminationconditions, atmospheric haze, and sensor characteristicsinfluence the quality of the imagery. Radiometric cor-rections involve algorithms that attempt to remove thesesources of “noise” from the image, such that the databest represent the Earth’s surface features of interest.

Radiometric corrections to digital image data includecalibration to known energy sources, calibration amongsensors, and conversions from radiance to reflectanceand temperature. For satellite images, some radiometriccorrections are routinely applied before they are deliv-ered to the user. For some applications, such as linkingremote sensing to energy balance models, calculations

Remote Sensing of Forested Ecosystems 423

green band

red band

infrared band

/ /764x-axis

Figure 1. Structure of a digital image. The x and y axes repre-sent geographic space and the different data lavers (bands)represent multispectral space. Each X-y cell consists of digitalnumbers (DN), one for each hand. When d-hit data are viewedon a computer monitor, a DN of zero appears black and aDN of 255 appears white. All other DN are linearly scaledbetween black and white. Three hands can be simultaneouslyviewed on the monitor, one through each of the three “guns,”red, green, and blue. This enables “true-color” and “false-color” viewing.

of albedo, or change detection, additional radiometriccorrection efforts are crucial. The most common radio-metric corrections applied by users of digital remotesensing data involve algorithms for minimizing atmo-spheric and illumination angle effects (Teillet and oth-ers 1982 Ahern and others 1987. Hall and others199la). The amount of literature on these two subjectsis phenomenal, which is indicative of both our lack ofa thorough understanding of the phenomena and ofthe intractability of the problems. The important thingto realize is that these corrections are only approximate.and the corrected image may still contain noise, someof which is new noise introduced by the correction al-gorithm.

I m a g e E n h a n c e m e n t

Enhancement techniques are performed on the im-agery to aid visual interpretation and to transform im-ages into more meaningful data sets for specific digitalanalyses. The central purpose of image enhancementis to improve contrast among features of interest. Thesetechniques include contrast stretches, spatial filtering,and derivation of spectral vegetation index (SVI)images.

Contrast stretching uses a transfer function to maporiginal image intensities into a transformed image hav-ing improved contrast (Cracknell and Hayes 1991).With such stretches the pixel intensities are manipulatedin a nonspatial context, i.e., irrespective of intensitiesof neighboring pixels. Examples include linear contraststretching and histogram equalization, each of whichalters the frequency distribution of pixel intensity values


Figure 2. Different contrast stretches applied to 8-bit digital imagery left. no stretch: middle, linear stretch: and right. histogramequalization stretch. The associated frequency histograms for each stretch are shown under the images are shown under the images.

and thereby changes the appearance of the image (Fig-ure 2).

Spatial filtering is used to enhance spatial featuresin images and thus relies on analyses in specified pixelneighborhoods (Richards 1993). Low-pass spatial filterssuppress high spatial frequency detail, whereas high-pass filters enhance high-frequency detail. A techniquethat is used to sharpen an image is known as edgeenhancement, which can have the effect of delineatingobjects in the scene (Jensen (1986). Texture algorithmsprovide a twodimensional statistical measure of an im-age, which relies on a moving window of some specifiedsize (Hord 1986) and can he used to assist in segmentinga scene into different objects (Woodcock and H a r -ward 1992).

Spectral vegetation indices (SVIs) are multispectraltransformations of image data that generate new setsof image components, or bands, and thus representalternative descriptions of the original data (Richards1993). All SVIs ar nonspatial in nature, operating onthe multispectral digital values of individual pixels. Thesimplest SVIs are ratios, in which one image band isdivided by another or one band is subtracted from an-other and the result is divided by the sum of the values

in the two bands (Figure 3). Excellent descriptions ofmany of the common SVIs are given by Tucker (1979)and Perry and Lautenschlager ( 1984). Principal compo-nents analysis (PCA) and the tasseled cap transforma-tion are two widely used sets of SVIs that neither ofthe two references above discuss. PCA is a standardmultivariate statistical procedure, described in any mul-tivariate statistical text. The tasseled cap was specificallydesigned for Landsat data, having a multispectral scan-ner (MSS) variate (Kauth and Thomas 1976) and athematic mapper (TM) variate (Crist and Cicone 1984).All of the SVIs are primarily designed to enhance vegeta-tion components, generally by contrasting the vegeta-tion against soil and background components in thescene.

Thematic Classification and Related Procedures

Thematic classification of multispectral images in-volves the assignment to pixels of labels containing real-world descriptions of ground features (Figure 3). Stan-dard classification algorithms involve supervised andunsupervised methods (Mather 1987, Lillesand andKiefer 1994). Unsupervised classification commonly re-lies on statistical clustering to separate pixels into groups


Figure 3. Raw digital imagery (upper left is the red hand, upper right is the near-infrared band), a spectral vegetation indexcreated by dividing the near-infrared hand by the red hand of the imagery (lower left), and a thematic classification of the imagery.


Table 2. Typical error matrix used in remote sensing classification accuracy assessment (adapted fromCongalton 1991)

Reference data

Classified data Conifer forest Hardwood forest Water Agriculture Row total

Conifer forest 25 6 2 1 34Hardwood forest 8 15 0 4 27Water 1 1 0 7Agriculture 2 3 0 13 18Column total 36 25 7 18 X6

based solely on the likeness of their multispectral values.Subsequent to definition of statistical clusters. labelscan be applied to the clusters based on knowledge ofthe scene from ground data, field visits, or airphotos.Supervised classification requires the use of “trainingsets,” which are groups of’ pixels of a known type orlabel. The training sets are used to statistically define theknown classes in spectral terms. Using some statisticaldecision rule, such as maximum likelihood, or nearestneighbor. the multispectral values of each pixel in theimage to be classified are compared to the training datato determine which class the pixel is most like, and thepixel is labeled accordingly.

Many other options for classification exist. There arenumerous examples where supervised and unsupervisedclassification methods were combined (e.g., Nelson1981. Chuvieco and Congalton 1988). Texture imagescan he used for classification, with or without spectralbands (Peddle and Franklin 1991 I. Regression analysisis commonly used to derive relationships betweenground data and spectral data for specific numericalattributes within a given class (Butera 1986, Petersonand others 1986). and predictions from regression equa-tions may then be collapsed to classes (Cohen and others1995). Ancillary data, such as digital elevation models(DEM), are often used to provide additional informa-tion during image classification (Strahler 1981, Franklinand Wilson 1992). Spectral mixture analysis has beenused to map proportions of basic scene components(e.g., green vegetation, nonphotosynthetic vegetation,and shade), that were then collapsed into classes (Smithand others 1990a.b).

Thematic classification normally is followed by anassessment of classification accuracy. Reference datafrom field plots, aerial photography, and the like areused to array predicted versus observed observations ina table known as an error matrix (Table 2). Two typesof’ error are possible for any given thematic class X:

commission, in which pixels from classes other thanclass X are classified as class X. and omission. in whichpixels of class X are classified as another class. Thereare numerous problems associated with accuracy asses-ment (Congalton 1991), especially those concerningviolation of underlying statistical assumptions, and theprocess is commonly subjective.

Change DetectionChange detection involves the comparison of images

from a given location at two or more points in time.One can simply compare summaries of classificationsfor a given area at different points in rime or conducta spatially explicit analysis involving direct comparisonson a pixel-by-pixel basis (Figure 1). In the latter, andmore usual case. accurate spatial registration of‘ two ormore images is required. Commonly used algorithmsfor conducting change detection include image differ-encing and image ratioing (Singh 1986. Muchoney andHaack 1994). In the former, digital numbers of’ a singleimage band from one date are subtracted from digitalnumbers of a single band from a different date, and inthe latter, a single image band from a given date isdivided by a single image band from another date (Jen-sen 1986). Other methods involve calculation of princi-pal components (PCA) of single hands of multi&ateimagery (Richards 1984, Fung and LeDrew 1987), useof fuzzy set theory (Gong 1993), color additive displayof a single band from three different points in time(Sader and Winne 1992), Gramm-Schmidt orthogonal-ization (Collins and Woodcock 1994), and calculationof postclassification transition matrices (Hall and others1991 b). Change vector analysis (Malila 1980) describesthe vector and magnitude of change in multispectral,multidate imagery. This particular approach is currentlybecoming popular (Michalek and others 1993, Lambinand Strahler 1994) and is likely to see increased at-tention.


Figure 4. A three-band, false-color composite rendition of a 1972 image (upper left), the same for a 1991 image of the samearea (upper right), and a change detection map created by subtracting one date of imagery from the other (lower left).

With ail change-detection algorithms, either raw ortransformed images can be used. Determining the accu-racy of change detection is a difficult problem unlessgood reference data exist for multiple dates. The greaterthe number of change features desired (e.g., clear-cut,insect damage, succession) and the higher the spatialfrequency of these features in the imagery, the greaterthe chance for error. One of the greatest problemsis associated with spatial misregistration of multidateimages, which can cause high rates of error aroundedges of scene features (Townshend and others 1992).

Remote Sensing Research and Applications

This section illustrates the multifaceted utility of re-mote sensing data. The intent here is not to conduct acomprehensive review but to provide context to theprevious sections by summarizing some varied applica-tion and research studies using remote sensing data.We concentrate on a region of the United States inwhich remote sensing has been widely used, the PacificNorthwest (PNW). Although limited in geographicscope, recent use of remote sensing among the land-


management agencies, forest scientists, and other inter-ested groups of this region has blossomed, and with ithas come an array of very large-scale projects. Whereasproduction of vegetation cover and change maps is themost common goal in applications of digital remotesensing pertaining to forest ecosystems of the region,scientific aspects of remote sensing are more directedat understanding relationships between physical or eco-logical properties and spectral properties of these sys-terns and at developing algorithms to process digitalimagery into accurate vegetation maps. Much of theremote sensing work in the PNW region has been acombination of research and application, driven by afew simple goals like mapping old-growth and generalforest cover, understanding relationships between im-age data and stand structural, compositional, and func-tional attributes. and mapping changes in landscapepatterns.

Mapping Forest Cover

Mapping with digital remote sensing data in the PNWregion has involved general land cover mapping, theseparation of structural and successional classes, andmapping of wildlife habitat. In one of the earliest stud-ies. Walsh (1980) used Landsat MSS data to map 12land cover types in Crater Lake National Park, Oregon.USA. with an 88.8% accuracy. In addition to cover type,topographic slope and aspect had strong effects on im-age spectral properties, with tree size and density havinglesser effects. That study was later repeated (Walsh1987), with a more in-depth analysis, but similar results.Isaacson and others (1982) mapped elk habitat in theBlue Mountains of northeastern Oregon using MSS dataand large-scale aerial photography. Mapped attributesincluded vegetation type, crown cover, vertical struc-ture, and disturbance. No accuracy statistics were re-ported. Cibula and Nyquist (1987) used MSS data tomap vegetation cover in Olympic National Park. Wash-ington, USA. Combining topographic data and climato-logical models with the MSS imagery, they achieved a91.7% map accuracy for 21 land cover classes.

The largest set of mapping efforts in the PNW regioninvolved locating remaining old-growth forests on thewest side of the Cascade Range. This effort was under-taken independently by the USDA Forest Service, theWilderness Society, and the Washington Departmentof Wildlife. The earliest work was by Eby (1987)) whomodeled the relationship between near-infrared re-flectance, stand age, and solar incidence angle at thetime of acquisition of Landsat MSS imagery. This workwas based on the fact that older forests exhibit lowernear-infrared reflectance than younger forests and thatillumination angle and near-infrared reflectance are

highly correlated in older forests due to shadows associ-ated with complex canopy structure. Using regressionanalysis, predicted ranges of values for different forestage classes at different incidence angles were calculated,and then these were used to develop a deterministicclassification model. The experience from this researchwas extended by Eby and Snyder (1990) to map old-growth forests on 11.3 million acres of western Washing-ton. They reported 80% accuracy for the Cascades ofWashington and 85% accuracy for the Olympic pen-insula.

Morrison and others ( 1991) used a variety of data setsand methods, from airphoto interpretation to relativelysophisticated digital techniques using multiple imagesources [Landsat MSS, Landsat TM, and panchromaticSPOT high resolution visible (HRV)] and DEM data,to map old-growth forest on the national forests of thewest side of the Oregon and Washington CascadeRange. Detailed documentation of actual methodologyis not published, nor are map accuracies. Congaltonand others ( 1993) mapped old-growth on much of thesame terrain as Morrison and others (1991). The map-ping was done with Landsat TM, airphotos, DEM, andfield measurements. Although detailed methods are notpublished, the analysis appears to have involved exten-sive testing of relationships between ground and photodata and derived image variables, such as band ratios,band textures, and principal components. Accuraciesbetween 80% and 91% were reported for nine nationalforests. An interesting and important observation canbe made by comparing the results of Morrison andothers (1991) and Congalton and others (1993). Al-though these independent estimates of forest condi-tions may have narrowed the uncertainty of the amountand location of old-growth forest on nine national for-ests in the PNW region, their acreage estimates for anygiven national forest differed as much as 100%.

Fiorella and Ripple (1993a) used unsupervised classi-fication of TM imagery with an ERDAS (1993) topo-graphic relief image calculated from a DEM to classifysuccessional stages from clearcut to old growth in Doug-las-fir forests with an overall accuracy of 78.3%. Useof the topographic relief image improved classificationaccuracy for younger stands, but not for later succes-sional stages. They also found that the ratio TM 4/5and the tasseled cap wetness were strongly correlatedwith each other and with stand age, except on poorlyregenerated sites (Fiorella and Ripple 1993b). Ripple(1994) mapped percent conifer cover on 10.9 millionha of forest in Oregon using Advanced Very High Reso-lution Radiometer (AVHRR) imagery. The analysis wasbased on a regression relationship between Landsat MSSand coregistered AVHRR band values. The map de-


picted percent closed canopy conifer cover in l-km cellsand was presented as an analysis of forest fragmentationin Oregon. Correlation (r) between the AVHRR conifercover map and observations from U-2 airphotos was0.90.

Cohen and others (1995) used TM data to map forestcover over a 1.24 million ha multiownership landscapein western Oregon. Unsupervised classification was usedto separate four forest cover classes: open (<30%),semiopen (30-85%), closed canopy mixed conifer-hardwood (>85%), and closed canopy conifer (>85%).Of primary interest was distinguishing among succes-sional stages within the closed canopy conifer class.Thus. for this class, regression analysis was used to ex-plore relationships between the tasseled cap SVIs(brightness. greenness, and wetness), topography, andstand age. Topography strongly influenced the re-sponses of brightness and greenness. but not ofwetness.A regression model for predicting forest age from wet-ness was developed and applied. Forest age predictionswere collapsed to three classes: young (<8O years), ma-ture (80-200 years), and old growth (>200 years) Accu-racy of predictions for the three age classes was 75%.Overall. for the full land cover map, an accuracy of 82%was achieved.

Structure, Composition. and Function of Vegetation

Nonmappinq remote sensing studies focusing on bio-physical and ecological properties of PNW forests arerelatively numerous and have primarily been research-oriented. Relevant studies that have concentrated onvegetation structure. composition, and function are de-scribed below.

The Oregon Transect Ecological Research (OTTER)project was a major NASA-funded effort to evaluate theutility of a variety of sensors to provide input to ecosys-tem models for predicting forest growth and nutrientallocation (Peterson and Waring 1994). One major fo-cus was on estimating leaf area index (LAI). In earlystudies across the transect, Spanner and others (1984)developed regression relationships between LAI ofclosed-canopy conifer stands and the simple ratio (SR).The SR is a spectral vegetation index (SVI) derived bydividing near-infrared reflectance by red reflectance.The SR was highly responsive up to an LAI of about 3,at which time it began to level off with increased LAI.Beyond an LAI of about 5, there appeared to be littlesensitivity of the SR. Using the same data, Running andothers (1986) found chat correcting the imagery foratmospheric effects enhanced the regression relation-ship and showed that the SR was not asymptotic untilan LAI of about 10. With additional data, Peterson andothers (1987) confirmed the value of the SR for estimat-

ing LAI across the OTTER transect and determinedthat it was better related to LAI than a number of otherSVIs. Spanner and others (1990) showed that the SRwas greatly influenced by canopy closure, understoryvegetation, and background reflectance. In that study,the SR was asymptotic at an LAI of about 4 or 5. Usingspectral reflectance data from a variety of scene compo-nents (e.g., crown foliage. understory vegetation, treebark), Goward and others (1994) confirmed that SVIsare a function of not just LAI. but of canopy closureand background reflectance. as well as canopy opticalproperties. Using reflectance data of two understoryvegetation species collected with a field spectrometer,Law and Waring ( 1994) demonstrated that SVIS leveledoff at an LAI value of about 6.

Li and Strahler (198.5) developed a geometric-opticalcanopy reflectance model that can be inverted to pro-vide estimates of tree size and density. The model hasbeen used across the OTTER transect, with an observedversus predicted correlation coefficient of at least 0.90for both crown radius and tree density (Strahler andothers 1988). On a site-specific basis, however, or withina given forest cover type, the model provides less accu-rate predictions (Wu and Strahler 1994). Using a moreadvanced configuration of the model, Abuelgasim andStrahler (1994) demonstrated the potential for rstimat-ing tree size. shape, and density using angular radiancemeasurements from newer experimental sensors. Mogh-addam and others ( 1994) demonstrated that across theOTTER transect radar backscatter saturates at low levelsof biomass and that at low levels of biomass the backscat-ter signal was only weakly related to biomass amount.Johnson and others (1994) and Matson and others( 1994) used imaging spectrometer data (a sensor havingover 200 narrow spectral wavebands) to estimate canopvbiochemistry across the OTTER transect and found thatthe spectral region from red to near-infrared (the red-edge) was strongly related to canopy total nitrogen andcanopy chlorophyll content.

Cohen and others (1990) used semivariograms tocharacterize the spatial domain of l-m-resolution aerialvideography in relation to stand structural complexity ofDouglas-fir ‘forests. Image spatial patterns were stronglyrelated to canopy size and vertical layering. In a subse-quent study, Cohen and Spies (1992) found that textureof SPOT HRV 10 m was strongly correlated (>0.83)with several stand structural attributes (e.g., tree size,density, basal area). Cohen and others (1995) exploredrelationships between stand structure and TM tasseledcap SVIs. Models to predict structural attributes frombrightness and greenness were significantly improvedwhen the image data were stratified by topographic/solar incidence angle classes; however, these models


were not as strong as nonstratified models based onwetness. As a part of this study, the effect of the definednumber of classes on percent accuracy in mapping wasevaluated. For two or three classes of any given structuralattribute, acceptable accuracies (75% or greater) wereobserved, but for five or more classes accuracies de-clined to below 50%.

Ripple and others (1991) found that relatively strongrelationships exist (correlation coefficients of -0.89and -0.83, respectively) between near-infrared re-flectance of SPOT HRV 20 m and TM 30 m data andforest volume in stands 25 years old and greater on theMacDonald-Dunn Forest along the Willamette Valleyfringe of the Oregon Coast Range. Fiorella and Ripple(1993b) developed regression models to predict age offorest stands from 0 to 35 years old using a variety ofSVIs and found that TM band ratio 4/5 gave the bestresults (r = 0.96). They also found that conifer regener-ation success could be determined at approximately 12years after planting.

Thermal imagery was used by Holbo and Luvall(1989) to detect cover types on the H.J. Andrews Experi-mental Forest (HJA) in the western Cascades of Oregon.They compared frequency distributions from two setsof diurnal multispectral thermal data to develop modelsfor specific cover types. With additional analysis, Luvalland Holbo (1989, 1991) again used thermal data tomodel the radiation balance for specific cover typesand develop models of short-term thermal responses todiscriminate these different surfaces. They found thatbarren surfaces had the lowest response while forestedsurfaces had the highest, indicating that forest covermoderated incident radiation and was more efficient atdissipating heat. Sader (1986) found that slope andaspect had a greater effect on thermal emission of youngregeneration than on older stands in the HJA. However,mean surface temperature decreased as age increasedregardless of topographic position (Sader 1986).

Change Detection

Although several change detection projects usingdigital imagery are underway by various groups in thePNW region, few results are currently available. Thusfar only coarse changes associated with harvesting andother major disturbances have been evaluated.

Spies and others (1994) evaluated the effects of foresthavesting and regrowth between 1972 and 1988 on for-est fragmentation over 258.000 ha of the west-centralOregon Cascades. They used raw MSS data from 1972,1976, 1981, 1984, and 1988. After the images were coreg-istered, each was independently classified into threebroad cover types: closed canopy conifer forest, water,and other forest and nonforest types. Using a GIS, the

classified images were registered to an elevation classmap derived from a DEM and a digitized land ownershipdata layer. Maps for multiple dates of forest edge andinterior were then derived. From these, edge lengthand amount of interior habitat were quantified, andvarious landscape-level statistics calculated by owner-ship class.

Current and Emerging Issues

There is great potential for use of remote sensingto derive detailed information about forest conditions..Much has already been done with long-standing datasets like MSS, TM. SPOT HRV. and AVHRR, and newersensors having finer spatial, spectral. and radiometricresolution are becoming more readilv available. In thissection, we summarize some of the most important cur-rent and emerging issues that must be addressed tobetter integrate developing remote sensing technolo-gies with resource management needs and objectives.

Users of remote sensing data need a common frameof reference for efficient and effective communication.An excellent place to start is with the taxonomic struc-ture for remote sensing models developed by Strahlerand others (1986). This taxonomy distinguishes ‘be-tween a ground scene and an image of that scene. thecontinuous versus discrete nature of a scene, imagespatial resolution and scene object resolution, and de-terministic and empirical models of a scene. Conceptsassociated with scale and spatial resolution in relationto image-processing models are further developed byWoodcock and Strahler (1987) This paper is requiredreading for anyone faced with a choice of image dataand processing schemes for a specific set of mappingobjectives. What Woodcock and Strahler (1987) demon-strate is that the spatial structure of a scene in combina-tion with the type of information desired from associ-ated imagery tend to limit the choice of appropriateimage processing models for classification (e.g., spectralclassifiers. spatial classifiers, mixture models, and tex-ture models). Together, these two seminal papers pro-vide a foundation from which to build a solid under-standing of remote sensing.

Forest scientists and resource managers routinely de-fine forest stands visually by drawing polygons on airpho-tos. No two people will define stands in exactly the sameway, and this problem is one that will be prevalent forthe foreseeable future. When the focus shifts to stand/polygon definition in digital imagery using digital tech-niques, the problem is greatly exacerbated. Not onlydo we still have the interpreter-specific stand-definitionproblem, but we now have the additional difficulty ofdeveloping a computer algorithm with an appropriate


set of rules. There are numerous examples of polygondefinition algorithms (e.g., Kauth and others 1977,Hong and Rosenfeld 1984, Woodcock and Harward1992). but additional research is needed to develop arule-based system (Nazif and Levine 1984, Corr andothers 1989) that is flexible for different purposes. Forexample, a wildlife biologist should be able to work withthe same digital data set as a forest manager but shouldbe able to have the computer program define a set ofpolygons that is different from those of the forestmanager.

An area in need of substantial development is accu-racy assessment. Widely applied techniques for accuracyassessment technology are relatively old (Congalton1991). New mathematical and statistical techniques havebeen developed. and there have been some efforts toincorporate these into new ways of conducting assess-ments of accuracy (Craplewski and Catts 1992, Ma andRedmond 1995). Fuzzy set theory (or logic) is one partic-ular approach that has the potential to revolutionizethe field of accuracy assessment (Gopal and Woodcock1994). The basic premise of fuzzy logic is that we maynever be certain of a given label’s correctness, but weare often quite confident. Fuzzy logic enables us to haverelative degrees of certainty about the correctness of alabel and what other possible labels may be correct.

The spectral resolution of most current operationalremote sensing systems is quite limited. Landsat MSShas four spectral bands in the reflective portion of theelectromagnetic spectrum. and TM has six there andone in the thermal-infrared region. SPOT HRV multi-spectral imagery consists of only three spectral bands.On the horizon is imaging spectrometer data (e.g.,>200 narrow spectral bands), already available on anexperimental basis (Vane and Goetz 1993). These dataprovide detailed spectral signatures that enable finespectral absorption features to be evaluated. Much re-search has already been done using such data (e.g.,Kruse and others 1993, Mustard 1993, Roberts and oth-ers 1993), but not for the extraction of detailed for-est information.

One of the most difficult challenges in remote sens-ing of forests has been tree species identification. Thereare a multitude of factors influencing the spectral re-sponse of digital imagery, and species is only a minorinfluence relative to forest structure and topography(Colwell 1974). Life forms and functional groups, likehardwood, conifer. brush, etc., can be differentiatedwithout too much difficulty based on spectral propertiesalone. By incorporating other factors such as climate,elevation, topographic aspect, soil properties, and thelike, a more refined species differentiation is possible,as demonstrated by Woodcock and others (1994). How-

ever, few such models exist for the numerous forestregions around the globe. Imaging spectrometer datamay also provide improved species identification, if nar-row-band species-specific absorption features can beidentified. Additionally, we need to explore temporaldata sets to capture phenological events associated dif-ferent tree species.

Detection of changes in forested environment-s isan increasing emphasis in the use of remote sensing.Progress has been made in detecting forest clear-cutactivity (Skole and Tucker 1993) insect damage (Collinsand Woodcock 1994), forest succession (Hail and others(199lb), pollution damage (Vogelmann and Rock1986), and other important forest changes. However.the techniques used are not well developed or widelyapplied. Problems associated with spatial misregistrationand radiometric differences among images in a tempo-ral series are potentially large obstacles to detectionof subtle forest changes. Algorithm development forchange detection in forest environments are not welltested in a variety of forest types.

Remote sensing can play an important role in initial-izing, parameterizing, and testing landscape models ofvegetative dynamics that are used to project successionalchanges under natural and anthropogenic distur-bances. The types of modeling approaches used vary asa function of the intended objectives. Methods rangefrom applying individual-based or stand-level simulatorsto polygonal units. each of which delineate similar vege-tative conditions. to simplified cell-based state-transitionmodels where predetermined states are advancedthrough time based on deterministic or multiple path-ways of transition. Regardless of the modeling approachused, it is essential to define initial vegetative conditionsof the landscape at a grain size and at a level of detailcommensurate with the vegetative model. For detailedmodeling applications, remote sensing is essential forcharacterizing the structural and compositional distri-bution of the overstory. State-transition models requirerepresentation of general seral stage, which is easilydetectable with remote sensing, and sometimes an ex-plicit estimate of age. Broader consideration of the dy-namics of understory species and dead wood is an inte-gral part of the ecosystem management emphasis.Simulation methods for handling these features are be-ginning to come on-line. Delineating understory andcoarse woody debris levels are problematic, however,with remote sensing. Nevertheless. given generaloverstory characteristics, observed data for nonoverstoryfeatures from similar areas can be extrapolated to thosedelineated only by remote sensing. Parameterization ofmodels relies on a host of procedures and data sources.Analysis of historical changes in structural and composi-


tional changes or general development of seral stageswith remote sensing offers an efficient means to aid incalibrating the dynamic attributes of vegetative models.Dividing the time series into model calibration and cor-roboration data sets additionally provides the abilitv toindependently test model behavior, at least over thetemporal span represented by available data.

Sustainable forest management requires consistentvegetation data for large geographic areas. While thereis a definite role for remote sensing in providing thesedata, current remote sensing technology cannot providethe level of detail required for all purposes. Further-more, a great amount of research is required to keepahead of applications needs. As new data sets becomeavailable, time is required to explore those data andto develop algorithms for processing them into usefulvegetation maps. There is ample reason to expect that:-emote sensing, when properly understood and applied.will be of increasing utility for the foreseeable future.This future will need to include adoption of remotesensing technology by agencies willing to make an oper-ational commitment to applications. It will be importantto use remote sensing in concert with GIS. as part ofan ongoing decision support system to set policy basedon both historical trends and future simulations of land-scapes with spatial data. Specific future applications in-clude topics in landscape ecology, forest fire analysis,biodiversity, habitat models for rare and endangeredspecies, ecosystem management based on natural distur-bance regimes, analysis of riparian zones, forest health.forest inventory, and harvest scheduling. With remotesensing and GIS, we will have more functional and inte-grated systems for spatial analysis. The repetitive andsynoptic coverage provided by these technologies willhelp give us a better understanding of forest systems.how they function, and how to manage them with aholistic view.

Acknowledgments

This research was funded in part by the Forest andRangeland Ecosystem Science Center, USDI NationalBiological Survey (PNW 94-0489) ; the Ecology, Biology,and Atmospheric Chemistry Branch, Terrestrial EcologyProgram, of NASA (W-18,020 and W-18,437); the Na-tional Science Foundation-sponsored H.J. Andrews For-est LTER Program (BSR 90-I lti63); and the GlobalChange Research Program and the Inventory and Eco-nomics Program of the PNW Research Station, USDAForest Service. We gratefully acknowledge reviews ofthis manuscript by Steven Franklin, Janet Franklin, andCurtis Woodcock.

Literature Cited

Xbuelgasim. .t X., and A. H. Strahler. 1994. SIotlelinq bidirec-tional radiance measurements collected hv rhe advancedsolid-state array spectroradiometer I.Y%S) over Oregontransect conifer forests. Remote Sensing of Enuzronmfnt47261~275.

Ahern. F. J.. R. ,J. Brown, J. Cihlar. R Gauthier, J. .Murphv.R. .\. Neville, and I’. M. T’eillet. 1987. Radiomerrir correctionof visible and infrared remote sensing data .u the CanadaCrntre For Remote Sensing. International Jcwxnl r)j. Remote.smng 8: 1349-1376.

Barrett. E. C.. and L. F. Curtis. 1992. Introduction to environ-mental remote senslng, 3rd cd. Chapman SC Hall. I.ondon.426 pp.

B u r e r a . M. K. 1986. ;\ correlation and regression ,malvsis ofpercent canopy closure versus TMS hpecrral response forselected forest sine in the Sal1 Juan National For-est. (:oitrrado. IMX 7’rtm~nct~ons o n Gosrience d ibwie Smsnq“4: I‘)‘~-1 ‘19-- -.

Chiesa. (:. (:., .utd W. A. Tyler. 1994. Beyond cubic co~~~oluuon:ERIM restoration <or remotelv-sensed imagen. I:nrlh Ohctn~-lion .\/q~:/nr :1:-&O-44.

Cibula. \V. G., a n d 11. 0 . Syqulst. l!Mi. U\e of IopotyaphIcand climatological models in A geographic data base toi m p r o v e Laudsat MISS classiiicauon lor Olvmpic Xauon,liP;lrk. /‘/to/p-ammelnc hzgzneenn,q i;” &mole .S~n.c~n,~ 3X:67-75.

(~ohrn, Lt’. IL ;md T. ‘1. S p i e s . 1992. Estunututg \rructuralattrlbutes of’Dorlglas-fir/western hemlock forest stands tiomLaudsat a n d S P O T irnagcrv. /time/P Gnrzrq 01 f<mvv~nvwn~Al:I-17.

Cohen. IV. B., T. A. Spies. and G. A. Bradshaw. 1990. Semlclrlo-grams of digital imagerv for analysis of conifer canup\ \truc-ture. Remote .‘itmsmg 01 I<m~wrmmml 3-L: 167-l 78.

Cohen. LV. 13.. T. A. Spies, and M. Fioreila. 3995. Estimating theage aud structure of forests in a mulriownership landscape otwestern Oregon. USA. Intematzonal Journal 01 Remoie Spns-,n,g 16:721-746.

Collins, j. B., and Waodcock, C. E. 1994. Change derecnonusing the <;ramm-Schmidt transformatiou applied t o mapp i n g f o r e s t m o r t a l i t y . R e m o t e %nsinq fjj f:‘nvwonmenf50:267-279.

Colwell. J. E. 1974. Vegetation canopy reflectance. Remote Sens-ing o/‘En7fironmat 3: 175-I 83.

Congaltou, R. G. 1991. A review of assessing the accuracyof classifications of remotely sensed data. Remote Sennsmg ufLnumnmat 37:35-46.

Congalton, R. C., K Green, and J. Teply. 1993. Mapping old,rrowth forests on national forest and park lands III rile,,Pacific Northwest from remotely sensed data. Photqrum-metric Eqzneting tj Remote .%-ruing 59529435.

Corr, 0. G., Taylor, A. M., Cross, A., Hoggs, D. C., Lawrence,D. H.:Mason. D. C., and Petrou, M. 1989. Progress in auto-matic analvsis of multi-temporal remorelv-sensed data. Inlet-natzonal Jokmal of Remote Sensuq 10:1175-l 195.

Remote Sensing of Forested Ecosystems

Cracknell, A., and L. Haves. 1991. Introduction to remotesensing. Tavior and Francis, London, Eng. 293 pp.

Crist, E. P., and R. C. Cicone. 1984. .I phvsicallv-based transfor-mation of thematic mapper data-the TM tasseled cap. LEEETransactions on Geoscience & Remote Sensing L’2:256-263.

Czaplewski, R. L.. and G. P. Catts. 1992. Calibration of remotelysensed proportion or area estimates for misclassificationerror. Raote Spnszn~ 0j~Environm~nt 39:2943.

Eby. J. R. 1987. The use of sun incidence angle and infraredreflectance levels in mapping old-growrh c,ontfer-otts I‘orests.Proceedings, .I.SPRS-ACSM fall convenuon: Prospecting newhorizons, ASPRS Technical Paper. Reno. Neclda. +-9 Octo-ber, American Society of Photoqammetric k Kemore Srns-ing, Falls (:hurch Virgmia. pp. M--&O.

Ehv. J. K.. .md 51. (1. Snvder. 1990. The \tatus of old growthin western Washington: .I I.andsat perspective. Report.\Vashington Department ol~~Viltililt. Olvmpia, \Vashington.34 pp.

EOSAT. 1994. ‘1 process of compensation. I:O.%l7 MJ/P.~ 9:s.

ERDAS. 1993. Earth resource data .tnalvs~s svstem. vets. X.0 1.ERDAS Inc.. .~tlanta. (;corgta.

Fiorella. M., and W.,J. Ripple. 1993h. .\nalvs~s of conifer forestregenerntinn ttaing l.andsat t h e m a t i c Inapper d a t a . 1’12orrkpmmrtnc +prfnn,q LY Remote Smsznq 59: ISHS-1 S-M.

Fratlklin. S. E.. and R. .A. \Vilson. 1992. .I three-stage classtfier1or t emote sensing f~f mountainous environments. /‘holo-,gmmmrl7ir I:‘n,pwnny d I~%tnot~ Yrnsrny 3: 449-4.X.

Fttng, T.. and E. l.eDrrw. 1987. .Ipplicauon of principal com-j>onents analysis to c-ttange detecuon. ~~oto,qmmmelnr E:‘nq-z-wri77~ ti RmlOlP .SP7Lwq 53: 1649-I GU.

(ion?, P. 1993. Change detectton usmg princtpal componentnnalvsis and fuzzy \et theorv. (Irrnnriiun ]rntmc~l ofRemote .S’FTWInq 19:?2-29.

Gopal. S.. and C. Woodcock. 1994. Theorv and methods foraccuracv assessment of thematic maps using tuzrv sets. P!to&,;pmmminc I:‘n~qnrPnn,q d &mole .Senc~ng 60: 1 Ul-‘I 88.

Goward. S. N., D. G. Dve, S. Turner. andJ. Yang. 1994. Visible-near infrared spectral reflectance of landscape componentsin vvestern Oregon. I~mote.Cn.c~n,~o~I:‘nctronmenl Ai: 190-203.

Hall, F. C., D. E. Strebel, J. E. Nickeson, and S. J. Goetz. 1991a.Radiometric rectification: Toward a common radiometnct-esponse among multidate. multisensor images. l&mole Seru-lng of Lnvimnmt 35: 1 l-27.

IHall. F. G., D. B. Botkin, D. E. Streble, Ii. D. \Voods, and S. J.Goetz. 1991 b. Large-scale patterns of forest buccession .tsdetermined by remote sensing. Ecolopy 72:628-640.

Heard, M. I., P. M. blather, and C. Higgins. 1992. CERES: .Iprototype expert system for the geometric rectification ofremotely-sensed images. Internuttonul Journal of Remote Serwlng 13:3381-3385.

Holbo, H. R., and J. C. Luvall. 1989. ,Modeling surface tempera-ture distribuuons in forest landscapes. Rao~eSasingo/Envr-ronmmt 27: 11-24.

Hong, T. H., and A. Rosenfeld. 1984. Compact region extrac-

433

tion using weighted pixel linking in a pvramid. IEEE Tmnsnc-l i o n s o n P&rn ,-inaLysu and .\~larhznr Int&gence PAMI-6:222-229.

Hord, R. .M. 1986. Remote sensing: Methods and applications.John Wiley & Sons, Sew York, 362 pp.

Isaacson, D. L., D. A. Leckenby, and C. J. .Vexander. 1982.The use of large-scale aerial photographv for interpretingLandsat digital data m an elk habitat-analysts project.Jountalof .A#lied Photo,yrammetric ICngineting 85 l-57.

Jensen.J. R. 1986. Introducton digital image processing. Pren-tice-Hall, Englewood Cliffs. Sew Jersev, 379 pp.

Johnson. L. F., C. A. tilavka. and D. L. Peterson. 1994. .\lultivar-iate analvsis of‘ AVIRIS data for canopv biochemical estima-t ion .tlonq the Oregon transect. /ir~mo/p \1?1,szn,g o/ I:?tzjirnn-tnmt 472 16-230.

Kauth. K. J., and G. S. Thomas. 1976. The tasselled cap-agraphic description of the spectral-temporal developmentofagricultural crops as seen by l.andsat. Proceedings, secondannual svmposium on machine processing of remotelv<ensed d a t a P u r d u e Universin L.ahomton ot .Ipplied Ke-mote Sensing, West I.afavette. Indrana. ti June-2 .Julv, pp.qB/-ll-.‘,I.

Lltrh, K. 1,. .\. I’. I’rntlantl. and G. 5. Thomas. 1977. Blob: ,111unsupervised clustering approach to spatial preprocessingof\ISS IIII:~CI~. I’roceedinqs, eleventh :tlrernauonal wmpo-\tum on remote scnstng of environment. vol. II. Environ-mental Research Insti tute of Slichtqm. .&in .irhor. Ilichi-;gd”. ‘X3--.‘)O .-\prii.

Ejuse, F. .I., .A. B. Lefkoff, and,J. B. Dietz. 1993. Expert svstem-based mineral mapping in northern Death \‘alley, Califor-ttta, Se~ttla, ~tstng auborne visible. itltrared Imaging spec-cromerer. (<emore Smstnq II/ Ertr~rrtJtr mvnl I-&:.%%336.

Lambin. E. F.. .md A. H. Strahler. 1994. Change-vector analysistn multi temporal space: .\ tool to cletecr .tnd ca t egor i zeland-cover change processes ttsing htqh temporal-resolunonratellite data. Remote .Gnw7g I,/ f~nvwonment %231-244.

Law. B. E.. ,md R. H. Warmg. 1994. Remote sensing of leaf.areain&x .it~tl raciiatioti intercepted bv imderstorv vegetation.Ecnlogiral ,4pplirations 4272-279.

Li. S.. and :I. 1-l. Srrahler. 1985. Geometnc-c)pucal modeling

L

L

L

of a conifer forest canopy. i&IX Tmnsactz0n.s on Geosaence tiRemote Sensing GE-233705-721

,illesand, T. M., and R. \V. Kiefer. 1994. Remote <ensing andimage interpretation, 3rd ed. John M’ilev & Sons. New York,7 5 0 p p .

,uvall, J. C., a n d H . R . H o l b o . 1 9 8 9 . \leasurements o fshort term thermal responses of coniferous forest canopiesusing thermal scanner data. Remote .Senszng of Environment27: l-10.

.uvall,J. C., and H. K. Holbo. 1991. Thermal remote sensingmethods in landscape ecology. In Ecofo,@cnl Studies Vol. 82,pp. 127-152.

la, Z., and R. L. Redmond. 1995. Tau coefficient for accuracyassessment of classification of remote sensing data. Photo-grammettic Enginfwng ti Remote Senszng 61:439-439.

Malila. W. A. 1980. Change vector analysis: An app roachfor detecting forest changes with Landsat. Proceedings .sixth aunual sympos ium on mach ine p rocess ing o f r e -motely sensed data. Pages 326-335 in P. G. Burroff and


D. B. Morrison (eds.), Soil information systems and re-mote sensing and soil survey. Purdue University Laboratoryof Applied Remote Sensing, West Lafayette. Indiana, 3-6June.

Mather. P. .M. 1987. Computer processing of remote sensedimages: An introduction. John Wiley SC Sons, Chichester.England. 352 pp.

Watson, P.. L. Johnson, C. Billow, J. Miller, and R. Pu. 1994.Seasonal patterns and remote spectral estimation of canopychemisttv .tcross t h e O r e g o n t r a n s e c t . i?r~lo,gi~nl ,\#>,llhca-lions 4:280-298.

Michalek, J. L.. T. \V. Wagner, J. ,J. Luczkovich, and R. W.Stoffle. 1993. .Cluitispectral change vector analysis for moni-toring coastal marine environments. !‘holo,~ammvlnr I:n,e-nwrin,g 2 Remote Sensing 59:38 l-384.

Moghaddam. W.. S. Durden, and H. Zebker. 1994. Radar mca-surement of forested areas during OTTER. Remote SenszngujEnvironm.ent -17:15&166.

Morrison, P. H.. D. Klopfer, D. A. Leversee. C. M. Socha. 2ndD. L. Ferber. 1991. .Xncient forests in the Pacific .Uorthwrst..Anahsis and maps of hvelve national forests. The 1t’ildernessSocietv, Washington, DC, 22 pp.

,Muchonev, D. M.. and B. A’. Haack. 1994. (zhange detection formomtoring forest detoliation. I-‘ho&gr~~mmrmc f:‘nk~nvmlt~: dRemote Senszng 60: 1243- 125 1.

Mustard. J, F. 1993. Relationships crf \oil, grass. and bedrockover the Kaweah Serpentine AMelange through mixture anal-vsisof.AVIRIS data. ~oteStnszn,~ofFnuzronfnunt~14:L’t)3-~~OX.

Nazif,;l. 11.. and !vl. D. Levine. 1984. Low level imagesegmenta-non: an expert system. Ii%? Trczn.sartzons on Inlflactzons t,ff’flttem In&i5 fznri Marhinr Intelligence PAMI-6:3.55-577.

Nelson, R. F. 198 I. h comparison of two methods tor classifvingforestland. Inkmatzonal ~uurnnl of Remote Sensznq 249-60.

Paine, D. P. 1981. Aerial photography and image interpreta-tion for resource management. ,John Wiley SC Sons. NewSork. 371 p p .

Peddle, D. R., and S. E. Franklin. 1991. Image texture pro-cessing and data integration for surface pattern discrimina-t i o n . Photo,pmmetric En~peezing 3 liemote Sensznp 37:413-420.

Perry, C. R., and 1~. F. Lautenschlager. 1984. Functional equiva-lence of spectral vegetation indices. f&ma/e Senting o/f+ntiron-m-nt 14583-597.

Peterson, D. L., and R. H. Waring. 1994. Overview of theOregon transect ecosystem research project. f~cologwaf A#@-cations 4:211-225.

Peterson. D. L.. W. E. Westman, N. J. Stephenson, V. G.Ambros ia , J. A. Brass, and M. A. Spanner. 1986. Analvsisof forest structure using thematic mapper simulator data.I,!22 Transactzons on Geoscience @ Remote Sensing GE-24:113-121.

Peterson, D. L., .M. A. Spanner, S. W. Running, and K B.Teuber. 1987. Relationship of thematic mapper simulatordata to leaf area index of temperate coniferous forests. Re-mote .Senszng of Environment 22:323-341.

Richards, J. .-\. 1984. Thematic mapping from multitemporalimage data using principal components transformation. fi-mote Senszng of Environrmznt 16:35-46.

Richards, J. .I. 1993. Remote sensing digital image analysis:An introduction, 2nd ed. Springer-Verlag, Berlin, 340 pp.

Ripple, W. J. 1994. Determining coniferous forest cover andforest fragmentation with NOAA-9 advanced very high reso-lutlon radiometer data. PhotogrammetricEqineeting ti RemokSasinq 60533-140.

Ripple, W. J., S. Wang, D. L. Isaacson, and D. P. Paine. 1991..\ preliminatv comparison of Landsat thematic mapper andSPOT-l HRV multispectral data for estimating coniferoust’orrst volume. Intemcztionnl ,Jmtmal of R e m o t e .+nszng IL’:1971-1977.

Robe r t s , D . A., M. 0. Smith. and ,J. B. .Adams. 1993. G r e e nvegetation, nonphotosvnthetic vegetation. and soils in.vIRIS. Remote .Smsznq ojEnvironment 44:255-269.

R u n n i n g , S. IV., D. I . . Petersot,, M. .\. S p a n n e r . .md 1;. !%.Teuber. 1986. Remote sensing ofcomferous forest leaf area.I?olog 67:273-976.

%ader, 5. .\. 1986 . Analvsis of etfective r a d i a n t t e m p e r a t u r e sin LL Pnctfic Sorthwest torest usmg thermal infrared inulti-spectral <canner data. Remote .Sensing 0/ I:nvzronmxnt 19:10.5-l I.‘,.

Satler, bi. .4 . . .mtI,j. Cl. \t’intie. l!t92. RGB-ND\‘1 wlwtr c~,mpw-ites forvisualizmg forest c-harigr tivnamics. /nkmt~lioncriJour-!IN/ t,, /holr .sm.szrt~ 13::3055-3067.

Singh, .\. 1986. Change detection in the tropical forest enviroii-ment of t1ortheastern lntiia nsin,g I.msdt. hqes ?J~-L’% i n

\l.,J. Etlcri .mtt,J. T. Pat-iv I rds. I, Remote setising atid troptcalImtl management. John \\‘ilet. S: Sons. Sew York.

Sk&. I).. .md rucker. C. 1 9 9 3 . T r o p i c a l deforestanon .mdhaliitat rm~mcntaiion tii the .\ma~o~i: S a t e l l i t e rlata from1978 IO 1988. .Sczmce 1’60: 1905- I9 IO.

S m i t h . \I. 0 . . 5. L . Ustin, ). 1~. .\tlams. .md .\. R. (;illesple.199Oa. \‘rgetation in deserts: 1. A regional measure of abun-dance ii0m multispectral images. lirmotr .&nszng (I/ Envtron-mmt 31:1-‘6.

‘imith. SI. 0. . S. I,. Ustin, J. B . .\dams, a n d .\. R. (Gillespie.199Oh. \‘egetation in cleseru: II . Environmental inlluenceso n r e g i o n a l a b u n d a n c e . f&mote .Sen.wrq i,l f,‘nvzmnmrnt:i 1:L’7-i‘2

Spanne r , 11. A.. W. .kevedo, K W. Teuber. S. LV. Kunninq,D. L. Peterson, D. H. Card. and D. A. .Mouat. 1984. Remotesensmg of the leaf area index of temperate coniferous for-ests. Proceedings. tenth international symposium on ma-chine processmg of remotelv sensed data. Pages 362-369 i nX4. !vl. Klepfer and D. B. Morrison (eds.), Thematic mapperdata and geographic informatton systems. Purdue UniversityLaboratolv of Applied Remote Sensing, West LafJsette, lndi-ana. 12-14 June, pp. 362-369.

Spanner, .\I. .I., L. L. Pierce, S. W. Running, and D. L. Peterson.1990. The seasonality of AVHRR data of temperate conifer-ous forests: relationship with leaf area index. Remok Sensingof Enuzronment 33:97-l 12.

Spies, T. A.. W. J. Ripple, and G. A. Bradshaw. 1994. Dynamicsand pattern of a managed coniferous forest landscape. Eco-@Cal Applications 4:555-568.

Strahler, A. H. 1981. Stratification of natural vegetation forforest and rangelandinventoryusing Ldndsat digital imageryaird collateral data. Internattonai Journal I,/’ Remote S’enszng2:15-41.


Strahler. A. H.. C. E. Woodcock, and J. A. Smith. 1986. Onthe nature of models in remote sensing. Rerno& Sensing ofEnvironment 20:121-l 39.

Strahler, A. H., Y. Wu, andJ. Franklin. 1988. Remote estimationof tree size and density from satellite imagery by inversionof a geometricoptical canopy model. Proceedings, twenty-second international symposium of remote sensing of envi-ronment. hbidjan. Cote d’Ivoire. 20-26 October.

Teillet. P. M., B. Guindon, and D. G. Goodenough. 1982. Onrhe slope-aspect correction of- multispectral scanner data.(~unndian,Jotcnal (fl Remote Sensin,q 8:84-106.

Townshend, J. R., C. 0. Justrce. C. Gurnev, and J. McManus.1992. The impact of misregistration on change detection.L!CEE Transactions on Geasnace & Remote .Senensing 30:1054-l 060.

Tucker. CJ. 1979. Red and photographic infrared linear com-binations for monitoring vegetation. lGfnole Sensmg of Envi-ron7nenl H:l?7-150.

Vane, G., and A. F. Goetz. 1993. Terrestrial imagingspectrome-~IK (:urrent status, future trends. /&rnole Semrng of I:nvrron-tmt 44:117-126.

?‘ogelrnann. J. E., .md B. N. Rock. 1986. Assessing forest decline

in coniferous forests of Vermont using N-001 thematicmapper simulator data. International Journal of Remote Sens-ing 7:1301-1321.

Walsh, S. J. 1980. Coniferous tree species mapping using Land-sat data. Remote Sensing of Environment 9:l l-26.

Walsh, S. J. 1987. Variability of Landsat MSS spectral responseof forests in relation to stand and site characteristics. Infema-&ma1 Journal of Remote Sensmg 8: 1289- 1299.

Woodcock, C., J. Collins, S. Gopal, V. Jakabhazy, X. Li, S.Macomber, S. Ryherd, Y. Wu, V. J. Harward, J. Levitan.and R. Warbington. 1994. Mapping forest vegetation usingLandsat TM imagery and a canopy reflectance model. Remote.Sensing of I:‘nvw.mment 50:240-254.

Woodcock, C. E., and V. J. Harward. 1992. Nested-hierarchicalxene models and image segmentation. International Journalo/Kaote Sensing 13:3167-3187.

Woodcock, C. E., and A. H. Strahler. 1987. The factor of scalein remote sensing. RemoteS~in~o/~n:nvl~onmenf21:31 l-332.

Wu. Y., and A. H. Strahler. 1994. Remote estimation of crownsize, stand density, and biomass on the Oregon transect.I:rolop.r~l :~pjG,licnlions 1:?99-312.

An Introduction to Digital Methods in Remote Sensing of Forested Ecosystems…myweb.facstaff.wwu.edu/~wallin/envr435/pdf_files/cohen_etal_1996.pdf · An Introduction to Digital Methods

Documents