optical remote-sensing techniques for the assessment of...

Optical remote-sensing techniquesfor the assessment of forestinventory and biophysicalparametersMike WulderPaci®c Forestry Centre, Canadian Forest Service, Natural Resources Canada,506 West Burnside Road, Victoria, BC V8Z 1M5, Canada1

Abstract: Forests are the most widely distributed ecosystem on the earth, affecting the lives ofmost humans daily, either as an economic good or an environmental regulator. As forests are acomplex and widely distributed ecosystem, remote sensing provides a valuable means ofmonitoring them. Remote-sensing instruments allow for the collection of digital data through arange of scales in a synoptic and timely manner. Accordingly, a variety of image-processingtechniques have been developed for the estimation of forest inventory and biophysical parametersfrom remotely sensed images. The use of remotely sensed images allows for the mapping of largeareas efficiently and in a digital manner that allows for accuracy assessment and integration withgeographic information systems. This article provides a summary of the image-processingmethods which may be applied to remotely sensed data for the estimation of forest structuralparameters while also acknowledging the various limitations that are presented. Currentadvancements in remote-sensor technology are increasing the information content of remotelysensed data and resulting in a need for new analysis techniques. These advances in sensortechnology are occurring concurrently with changes in forest management practices, requiringdetailed measurements intended to enable ecosystem-level management in a sustainable manner.

This review of remote-sensing image analysis techniques, with reference to forest structuralparameters, illustrates the dependence between spatial resolution to the level of detail of theparameters which may be extracted from remotely sensed imagery. As a result, the scope of aparticular investigation will influence the type of imagery required and the limits to the detail ofthe parameters that may be estimated. The complexity of parameters that may be extracted can beincreased through combinations of image-processing techniques. For example, multitemporalanalysis of image radiance values or multispectral image classification maps may be analysed toundertake the assessment of such forest characteristics as area of forest disturbances, forestsuccession and development, or sustainability of forest management practices. Further, thecombination of spectral and spatial information extraction techniques shows promise forincreasing the accuracy of estimates of forest inventory and biophysical parameters.

Key words: forest structure; forest inventory; forest biophysical parameters; optical remotesensing; image processing

*c Arnold 1998 0309±1333(98)PP200RA

Progress in Physical Geography 22,4 (1998) pp. 449±476

I Introduction

Forests and woodlands are the most widely distributed vegetation ecosystem on theplanet, covering approximately 40% of the global land surface (Westoby, 1989). Theeconomic importance of forests is clear, as through either consumption or utilization ofsome product or service forests affect the everyday life of most humans (van Martin,1984). Less clear is the impact of forests upon the global environment through processessuch as the regulation of the global climate, storage of carbon, conversion of carbondioxide to oxygen and energy exchange with the atmosphere (Gates, 1990). For example,it is known that deforestation and the combustion of fossil fuel contribute more than7 billion metric tonnes of carbon to the atmosphere each year above the natural flux,mostly in the form of CO2 (Jarvis and Dewar, 1993). Forests annually produce 70% of thenet global terrestrial carbon accumulation (Peterson and Running, 1989) which results inthe uptake of carbon from the atmosphere and the conversion of the greenhouse gas CO2to O2 (Perry, 1994). Sustainable forest management policies have been initiated toreconcile the competing aims for the use of forests (Toman and Ashton, 1995). Inter-generational responsibility dictates a need for monitoring of forests, as anthropogenicforces are potentially responsible for significant changes in ozone levels, desertification,deforestation and loss of biodiversity.

The temperate northern forests found in Canada serve as an example of the role offorests both in economic and environmental terms. Canada contains approximately 10%,or 417.6 million hectares, of the global forest cover (Westoby, 1989) with Canadian forestproducts accounting for 18% of the world's forest products exports, which in 1993 werevalued at $27 billion (NRC, 1995). Canada's forests have been estimated to contain2.6� 1010 tonnes of biomass and 2.4� 1010 m3 of gross merchantable timber (Brand,1990). This importance of forests, both environmentally and economically, has neces-sitated a change in Canadian forest policy. The implementation of ecosystem manage-ment shifted the emphasis from maintaining the ability to harvest a known quantityyearly, based on annual allowable cut, to the maintenance of healthy, diverse ecosystems.This change in forest management policy in Canada demonstrates the shift in globalforest management priorities from stand management to ecosystem management (NRC,1995). A key result of this change in priorities is the monitoring of complete naturalecosystem areas ± not just the artificial boundaries of a forest management area.Inventories under traditional forest stand management generally consisted of measuresof age, species and timber volume; yet within an ecosystem management framework, thelevel of detail of measures has increased, requiring information on soils, productivity andhabitat requirements. Further, within an ecosystem management framework, the futuremanagement goals for a forest are considered in terms of age, composition, structure,distribution and aesthetics (Gillis and Leckie, 1993), as well as nontimber values, such aspotential for employment and recreation (CCFM, 1995).

Assessment of forests within an ecosystem management framework implies bothgeographic and economic advantages in applying remote-sensing methods to generatedata on forest extent and location. Yet, often remote-sensing methods fail to capture thediversity of forests necessary for management decisions (Peterson and Running, 1989).Techniques such as multisensor fusion (Franklin et al., 1994), multitemporal analysis (Qiet al., 1993), increased spectral resolution (Gong et al., 1992) and increased spatialresolution (Leckie et al., 1995) may assist in providing the level of detail desired for theformulation of forest management decisions. The utility of remotely sensed data for

450 Optical remote-sensing techniques of forests

forest management is indicated by the Canadian forestry community, which is the largestmarket for satellite data in Canada of any application field, accounting for approximately22% of the annual sales of satellite imagery (Leckie, 1990).

This article is chiefly concerned with optical satellite and airborne remote-sensingmethods and applications for the estimation of forest inventory and biophysical para-meters. The promise of high spatial resolution sensors in satellite orbit in the near future(Aplin et al., 1997) will increase the utility of satellite data in the forest remote-sensingcontext and will likely increase the amount of users of remotely sensed data for forestmanagement. This article is intended as a resource to current and potential users of highspatial resolution optical remotely sensed data to extract digitally forest-related informa-tion. To provide a spatial resolution context for this review, the 30� 30 m pixel of theLandsat Thematic Mapper (TM) is the image spatial resolution range which is consideredupper bound, with aerial photographs as the lower bound. A survey of currently avail-able methods and applications is provided, as well as research issues and techniqueswhich are pertinent to the high variance imagery which is collected by high spatialresolution imagery. High spatial resolution airborne remotely sensed data have providedfor the development of methods and applications for the assessment of forests resultingin a rich information source for potential users of high spatial resolution satellite data.

The role and progress of radar remote sensing in forestry are acknowledged, yet thefocus of this article is upon analysis of the visible and near-infrared wavelengths, alsoconsidered as optical imagery. The behaviour of the spectral response in the microwaveregion differs from that which occurs in the optical wavelength region, requiringspecialized techniques for image preprocessing and analysis. A review of microwaveremote sensing in the assessment of forest structure may be found in Leckie (1998).

The ability to use remotely sensed spectral data for the assessment of forest structure isrelated to the changes in incoming solar radiation that is absorbed by the forest in thevisible and near-infrared wavelengths. For example, the amount of light available to thelower portions of trees is restricted by the level of the canopy closure; mutual shading oftrees results in differing light intensity available for interception by tree canopies; and,commonly, the amount of light penetrating through the canopy is inversly proportionalto the number of trees per unit area. As a result, measures have been developed whichrelate light transmission in individual stands to such stand density measures as leaf areaindex, trees per unit area, crown closure, basal area and stem density (Curran, 1980).

1 Forest structure through inventory and biophysical parameters

Forest structure is the above-ground organization of plant materials (Spurr and Barnes,1980), with the structure of a given forest being the result of competition for light, waterand nutrients at a particular location (Kozlowski et al., 1991). Accordingly, the ability toassess the structure of a forest permits insights into the environmental factors, such ashydrology, albedo, productivity and soils. Understanding of the forest structure allowsfor the ability to monitor, model and predict important biophysical processes, such as theinteraction between the forest and the atmosphere, based upon the input of a foreststructural measure to a forest productivity model (Running et al., 1994). Changes in foreststructure may also provide for forest inventory information related to forest vigour,harvests, burns, stocking level, disease and insect infestations (Gillis and Leckie, 1996).As suggested, forests may be characterized in terms of inventory measures or biophysicalparameters. Inventory parameters provide detailed data on the location and extent of

Mike Wulder 451

forest resources (see Table 1 for list of typical inventory parameters). Forest biophysicalparameters provide data on the productivity, structure and amount of forest resources(Table 2 presents and defines the most common forest biophysical parameters). Thesemeasures are most commonly used as they are often correlated measures and can also beapplied to any plant canopy and may be integrated into regional-scale models (Runningand Hunt, 1994). Forest biophysical parameters are an attempt to simplify themeasurement of forest structure into a single measure, such as leaf area index (LAI).LAI is an important structural attribute of forest ecosystems because of its potential to bea measure of energy, gas and water exchanges. Maximum canopy leaf area is correlatedto mean annual temperature, length of growing season, mean annual minimum airtemperature and water availability (Gholz, 1982). Further, physiological processes suchas photosynthesis, transpiration and evapotranspiration are related to LAI (Pierce andRunning, 1988). The collection of the detailed measures that characterize a forestinventory has previously been limited by the technical capabilities of remote-sensinginstruments. Current technological developments are enabling greater spectral andspatial resolution on a variety of platforms enabling the remote measurement ofinventory parameters (Leckie, 1990; Leckie et al., 1995). Forest assessment approacheswhich incorporate data from a variety of spectral and spatial resolutions are necessary toaddress the complexity of sustainable forest management with remotely sensed data.

Table 1 Typical provincial forest inventory parameters

Parameter Detail

Scale Normally 1 : 12 500 to 1 : 20 000Species Abbreviated species or tolerance levelDevelopment stage Development description, including elements such as cut, burn, regenerating

and matureCrown closure Percentage classes from 10 to 30% to greater than 90%Volume Volume of timber per unit area and by speciesBasal area Total cross-sectional area per unit area measured at 1.37 m above groundHeight Mean tree height per unit area and by speciesStand indicators Brief site descriptionNonforest conditions Nonforest characteristics, such as agricultural, mining, gravel and cut blocksOwnership Land ownership characteristicsMap symbol legend Explanation of the symbols used

Source: After Gillis and Leckie (1993).

Table 2 Typical forest biophysical parameters

Parameter Detail

LAI Leaf area index ± is a measure of area of foliage per unit area of groundBiomass Biomass ± is the total of absolute amount of vegetation present (often considered in

terms of the above-ground biomass)NPP Net primary productivity ± is similar to biomass, but has a temporal component as it is

related to the amount of biomass accumulated over a given time period

Source: De®nitions after Bonham (1989).


2 Background for the remote sensing of forests

Optical remote-sensing instrumentation for spatial data collection of forests ranges fromanalogue aerial photographs to complex spaceborne digital multispectral instruments. Acomplete discussion of remote-sensing instruments is beyond the scope of this review ±for a description of instruments, see Jensen (1996); for a problem analysis of the use ofremote sensing for vegetation management, see Pitt et al. (1997). In general, opticalremotely sensed data are a result of a complex series of interactions between theelectromagnetic radiation emitted by the sun reflected off the earth's surface and receivedby a sensor. In the forestry context, this complex series of interactions encompassesfactors such as the optical properties of the stand, spatial resolution (scale), stand objectrelationship to scale and spatial aggregation.

The interpretation of remotely sensed data of forest canopies requires knowledge of thefactors affecting their optical properties, which may be internal or external to the foreststand (Guyot et al., 1989). External factors that have an affect on the reflectance of forestsare the size of the viewed area, orientation and inclination of the view axis between thesurface and sensor, sun elevation, nebulosity and wind speed. Factors internal to thestand that have an effect on reflectance are canopy geometry, optical properties of thebackground and row orientation (Guyot et al., 1989). Terrain also has an effect uponstand reflectance (Craig, 1981; Civco, 1989) resulting in variations in reflectance basedupon the sun/surface/sensor geometry (Schaaf et al., 1994). The spatial resolution of aremotely sensed measurement is determined by the sensor's instantaneous field of view(IFOV), the area of the target which is viewed by a sensor in an instant of time. Withimaging sensors, this quantity is normally expressed as a pixel size. Using this definition,spatial resolution is analogous to scale (Woodcock and Strahler, 1987).

Scale is a fundamental concept in remote sensing and plays an important role indetermining the type and quality of information that can be extracted from an image.Marceau et al. (1994a) present the concept of decreasing variance with lower spatialresolution, the scale and spatial aggregation problem in a forestry context, and the effectscale changes have upon classification accuracy. The need for the sampling grid tocorrespond to cover type is suggested, based upon computation of minimum spectralvariance to assess the scale and aggregation characteristics of the cover type (Marceauet al., 1994b). The sampling grid, as dictated by the sensor, dictates the spatial variability,and has an effect on upon classification accuracy and potential information extraction(Irons et al., 1985; Chavez, 1992). Woodcock and Strahler (1987) present semivariogramsas a means to investigate optimal resolution for a particular investigation.

The relationship between the spatial resolution and the forest objects of interest influ-ences the information content as the elements which comprise an image are representedin detail as a function of the scale. Strahler et al. (1986) propose the concepts ofH-resolution and L-resolution to characterize scene models based on informationcontent. The H-resolution case occurs if the objects of interest in the scene are larger thanthe image resolution. The L-resolution case occurs when the resolution cells are largerthan the objects, resulting in an inability to resolve individual elements. The termshigh and low resolution often carry an absolute connotation of a pixel size, not ofinformation content. Differing abilities for analysis exist depending on the initial imagedata content. As a result, there is an implicit limit to the information content associatedwith a particular sensor through knowledge of the spatial resolution of the sensor. Forexample, with Landsat TM data the 30 m pixel results in a combined spectral response

Mike Wulder 453

representing an area, such as trees, understory and leaf litter. The L-resolution pixelresults in a low-variance environment with marked spatial association between neigh-bouring pixels. While an H-resolution situation, as sensed with an airborne multispectralscanner or digital frame camera, may collect image data at a spatial resolution ofless than 1 m, results in a high-variance environment often with great variabilitybetween neighbouring pixels. In an H-resolution situation, a group of pixels maycombine to represent an individual tree rather than the characteristics of an entire stand.As a result, it is the relationship between the ground objects of interest and the sensorresolution that dictates the information content. For example, 1 m spatial resolution datamay allow for an H-resolution image spatial structure when sensing large mature trees,yet not for smaller trees which are not large enough to comprise a number of pixels.

Remotely sensed instruments typically discretize a continuous natural surface into auniform grid of equally sized and shaped pixels (Jupp et al., 1988; Fisher, 1997). As aresult, there is no intrinsic geographical meaning to the spectral measures recorded byremote-sensing systems, with each image that is collected being a unique sample of thesurface features. This problem is demonstrated in the difficulty in extracting accurate,reproducible information from images of varying resolutions, and in this respect issimilar to a phenomenon understood to human geographers as the modifiable areal unitproblem (MAUP) (Openshaw, 1984). The MAUP is two sets of interacting problemswhich are related to the spatial scale of the data and any need for aggregation of thespatial data. There are two primary problems within the MAUP:

. A variety of different results may be computed for the same data as they areincreasingly aggregated (scale problem).

. The data may also be aggregated in a variety of ways (aggregation problem).

The scale problem refers to the variation of results that can be obtained when the sameareal data are combined into progressively larger units of analysis, and indicates a failureto discriminate the objects of geographical inquiry. The aggregation problem arises fromthe large number of ways in which these areal units can be combined, and reflects thecomplexity to account for the processes at work between scales. Knowledge of thepotential factors affecting the spectral response of forest canopies assists in the applica-tion of the most appropriate image analysis technique. In a remote-sensing context, theMAUP is related in varying image spatial resolution, combined with arbitrary gridpartitioning of the natural surface. Further, the MAUP must also be considered whenrelating data collected on the ground to image data; the ground data must represent anareal extent similar to that collected by the remote-sensing instrument.

The information content of remotely sensed imagery, as previously discussed, isdirectly related to the relationships between the image spatial resolution and the size ofthe objects of interest on the ground. The spatial resolution ranges commonly found for aseries of instruments are presented in Table 3, from aerial photographs, to airbornemultispectral, to satellite-borne instruments.

a Aerial photography: The human assessment of aerial photographs continues to bethe remote-sensing inventory method of choice. Conventional methods of interpretationare both time consuming and costly and results may vary between analysts. Thedevelopment of inventory techniques based upon the image processing of multispectralsatellite imagery increases the potential for reduced costs and automation. Leckie (1990)


presents a summary of the costs per unit area of remotely sensed data, which relates thetrade-off between costs and detail, with aerial photographs being the most costly whileproviding the greatest amount of information on the forest, with satellite data being theleast expensive per unit area with the least amount of detail. Yet, the high level ofinformation afforded by aerial photography is not always necessary; the lesser detailedand lower-costing satellite imagery may be appropriate for particular applications.

Aerial photography is the oldest and most frequently utilized form of remote sensing,originating with the merging of a photographic camera and a balloon platform in 1859.The spatial resolution of aerial photographs is very high and is limited mainly by filmproperties. The spectral resolution is normally panchromatic ( from 0.25 to 0.7 mm) withthe potential for collection of other spectral ranges (Avery and Berlin, 1992). The qualityof airphotos is influenced by factors such as atmospheric scattering, aircraft motion andvibration, and the resolving power of the film. Aerial photography is the most frequentlyutilized remote-sensing tool for the assessment of forests (Gillis and Leckie, 1996). Theuse of aerial photography in the estimation of biophysical parameters is infrequent in theliterature in comparison to usage for forest inventory.

The interpretation of air photos is generally an analogue procedure requiring anexperienced specialist to extract forest inventory information (Watts, 1983). As may be

Table 3 Example instrument-related spatial resolution ranges and levels of plantrecognition in forestry to be expected at selected image scales

Type or photo scaleApproximate range ofspatial resolution (m) General level of plant discrimination

Common satelliteimages

30 (Landsat)20 (SPOT multispectral)10 (SPOT panchromatic)

Separation of extensive masses of evergreenversus deciduous forests (stand-levelcharacteristics)

Small satellites >1 (panchromatic)>3 (multispectral)

Recognition of large individual trees and of broadvegetative types

Airborne multispectralscanners

>0.3 Initial identi®cation of large individual trees andstand-level characteristics

Airborne video >0.04 Identi®cation of individual trees and large shrubs

Digital frame camera >0.04 Identi®cation of individual trees and large shrubs

1 : 25 000 to 1 : 100 000 0.31 to 1.24* Recognition of large individual trees and of broadvegetative types

1 : 10 000 to 1 : 25 000 0.12 to 0.31 Direct identi®cation of major cover types andspecies occurring in pure stands

1 : 2500 to 1 : 10 000 0.026 to 0.12 Identi®cation of individual trees and large shrubs

1 : 500 to 1 : 2500 0.001 to 0.026 Identi®cation of individual range plants andgrassland types

Notes:`Small satellites' are the new generation of commercial high spatial resolution satellites (Fritz, 1996;Aplin et al., 1997).*Based upon a typical aerial ®lm and camera con®guration utilizing a 150 mm lens.Source: After Avery and Berlin (1992); Pitt et al. (1997).

Mike Wulder 455

expected, the results of the interpretation of air photos by photo interpreters areexpensive, time consuming and inconsistent. Digital methods have been developed in anattempt to streamline the photo interpretation process (Leckie, 1990; Dymond, 1992). Theability to produce digitally a complete forest inventory from airphotos is not yet matureenough for complete automation, with the need for the judgements of an experiencedphoto interpreter still required to account for unusual or unforeseen image characteristics(Leckie et al., 1998).

b Airborne remote sensing: Airborne remote sensing provides a flexible operationaland experimental tool. The ability to sense the surface of interest remotely at a desiredtime with the chosen technical specifications is the key feature of airborne remotesensing. Yet, airborne remote sensing has limitations due to such factors, as theinstability of the platform and limited ground coverage. Due to low flying altitudes,high-spatial resolution and sensor engineering, airborne multispectral remote-sensinginstruments often have a narrow image width often resulting in the need for mosaicking.The requirements of a particular project dictate the spectral characteristics, the desiredresolution, flight line azimuth and location, desired illumination conditions andacceptable flight conditions (Wulder et al., 1996a). In applications requiring lessspectral coverage, digital frame cameras provide a usually lower cost alternative tomultispectral scanners (see King, 1995, for a comprehensive review of digital framecamera technology). The high-resolution data collected from airborne platforms oftencreate an H-resolution environment with the associated increase in variance over presentsatellite systems. The increase in variance creates problems in the application oftraditional remote-sensing techniques, such as a multispectral classification (Gougeon,1995a).

c Satellite remote sensing: With the launch of the first Landsat satellite in 1972systematic, synoptic, repetitive, 80 m resolution multispectral images of the earth'ssurface were available from space. Data capture, transmission, storage and analysistechnologies have increased substantially since the first Landsat, with a wide variety ofspectral, spatial and temporal options available to users of satellite image technology.The use of satellite data in the estimation of forest structure has been directed at theestimation of biophysical parameters and the accurate estimation of forest inventoryparameters. The L-resolution sensors which have been available up to this point have notpovided the high accuracy inventory information required by forest managers (Gillis andLeckie, 1996).

The knowledge acquired from the analysis of airborne high-resolution data will soonbe applicable to a new suite of proposed high spatial resolution satellites (Aplin et al.,1997). In the near future, high spatial resolution satellites will provide spatial data at aresolution of approximately 4 m multispectral and 1 m panchromatic. The ability toestimate some elements of a forest inventory and to perform structural analysis from astable satellite platform with an image footprint of greater than 10 km may soon bepossible (Fritz, 1996).

d Image data-preprocessing considerations: Prior to the analysis of optical remotelysensed data a variety of considerations related to the integrity of the image data must beaddressed to account for image geometric and radiometric characteristics. To allow forthe locating of ground features on imagery, or the comparison between a series of


images, a geometric correction procedure is undertaken to register each pixel to real-world co-ordinates (Jensen, 1996). The concept of image radiometry is related to thespectral characteristics of the imagery. Radiometric corrections are applied to convert thesensor-specific digital numbers to radiance values. Image radiometry is also affected bythe atmosphere intervening between the ground surface and the sensor. The magnitudeof the scattering, absorption and emission effects upon the signal received by the satellitevaries by wavelength, resulting in a requirement for wavelength range-specific correctionprocedures (Avery and Berlin, 1992). In the visible wavelengths the main consideration isthe scattering caused by atmospheric haze, while in the infrared wavelengths it is thewater vapour resulting in atmospheric absorption of radiation, with different techniquesappropriate for satellite (de Haan et al., 1991; Teillet and Fedosejevs, 1995) and airborne(van Stokkom and Guzzi, 1984) collected data.

II Image-processing techniques for the assessment of forest structure

Studies involving remotely sensed data normally have data collection, processing andanalysis stages which relate the remote measure to ground-based validation measures(Steven, 1987). The data collected by multispectral remote-sensing instruments are in adigital form, allowing for mathematical analyses and manipulations. The informationcontent of remotely sensed data is enhanced by the ability to apply image analysistechniques to extract subtle structural information. Image processing provides the meansto assess the inter-relationships between pixel location and values. The following is anoverview of the image analysis techniques that have been used, or are in development,for the extraction of forest structural information from digital imagery. The suite ofimage-processing techniques which has been developed to generate forest biophysicalinformation from spectral information may be integrated with the newly developedobject-based approaches which may be applied to H-resolution data.

1 Spectral relationships

Relative changes in spectral response indicate variability in the groundcover. Spectralresponse is often related to the ground validation data allowing for the extrapolation ofthe calibration relationship to the entire image. Yet, due to the complexity of theinteraction between the downwelling radiation with the forest canopy and how thisrelationship is recorded by remote-sensing instruments, the relationship between theimage spectral response and the ground validation data is often poor (Danson andCurran, 1993; Baulies and Pons, 1995). Remotely sensed data are collected from above theforest, resulting in an inability to account for any vegetation characteristics not visiblefrom above. As a result, the degree of canopy closure often has an effect on standreflectance, especially in the near-infrared channels (Spanner et al., 1990). Furtherdifficulties in relating field validation data to image data are based upon such factors asthe dynamic range of the ground data in relation to the image data and the areal extentof the ground validation data compared to the image data. The use of spectral responseto estimate forest structure is likely the most common data extraction technique due tothe simplicity of the approach.

A test case for operational mapping with the low-resolution Landsat MSS demon-strated good agreement between the image estimates and ground measured parameters

Mike Wulder 457

when considering the scale of analysis (Bryant et al., 1980). A comprehensive estimationof forest stand parameters, including volume, from Landsat TM reflectance data foundresults comparable to conventional timber inventory methods (Franklin et al., 1986). Ingeneral a poor relationship is found between spectral response and volume (Gemmel,1995); yet Ripple et al. (1991), utilizing both SPOT and TM data, found throughcorrelation and regression analyses a significant relationship between softwood volumeand the log of reflectance data. Correlation methods are a common approach to theestimation of the inventory parameters of diameter at breast height, canopy closure,height, density and age (Danson, 1987; de Wulf et al., 1990). Correlation studies are alsoundertaken for the estimation of the forest biophysical parameter of leaf area index.When estimating leaf area index with spectral values, caveats to consider are the presenceof understory vegetation, within-stand shadowing, bidirectional effects and standdensity, as well as the canopy closure. Common to studies which estimate LAI fromremotely sensed data is a difficulty in estimating LAI once the foliage within the standsbegins to overlap (Baret and Guyot, 1991).

2 Spectral vegetation indices

Characteristically, green plants strongly absorb visible electromagnetic radiation andstrongly scatter near-infrared radiation (Curran, 1980). This is as a result of pigments,especially chlorophyll, which absorb visible wavelengths, while the air±water interfacesbetween the intercellular spaces and cell walls cause multiple refraction, resulting in highnet reflectance values in the near-infrared wavelengths (Gausman, 1977). Vegetationindices have been developed to emphasize the difference between the absorption in thevisible and reflectance in the infrared through mathematical processing of multispectralbands, such as ratioing and differencing. The normalized difference vegetation index(NDVI) is a commonly used vegetation index, calculated from the red (R) portion of thevisible and near-infrared (NIR) radiation in the form of:

NDVI � �NIR ÿ R�=�NIR � R� �1�This was initially developed as a measure of green leaf biomass (Tucker, 1979). NDVI hasalso been demonstrated to assist in compensation for changing illumination conditions,surface slopes and viewing aspects (Avery and Berlin, 1992). Canopy interceptiondictates the upper limit for vegetation to utilize available sunlight for photosynthesiswhich further drives production (Law and Waring, 1994). NDVI captures informationrelating to the amount of radiation absorbed in the visible (red) and reflected in theinfrared (NIR) by vegetation. NDVI has a range of 71 to 1, where values whichapproach 1 are the result of a high NIR value and low R value, indicating densevegetation, whereas low and negative values indicate such situations as low vegetationdensity or nonvegetated surfaces.

Vegetation indices, such as NDVI, may be viewed as a surrogate for scene vegetationcontent and are applied in an attempt to relate to physical measures of vegetation, suchas LAI. Other vegetation indices may be utilized with some success ± such as RVI,demonstrated by Spanner et al. (1994) to be well related to LAI. Nemani et al. (1993)utilized the middle infrared of the Landsat TM sensor to apply a correction for the effectsof changes in the level of canopy closure on the computation of NDVI. There are manydifferent vegetation indices which may related to particular characteristics in a given case


(Heute, 1988; Myneni et al., 1995; Chen, 1996). Chen and Guilbeault (1996) found that inrelationships between a vegetation index and LAI, the indices based upon simple ratiosof two bands were best correlated with ground measurements of LAI and fraction ofphotosynthetically active radiation (FPAR). Limiting the effectiveness of vegetationindices in the estimation of forest structure is that relationships are frequently nonlinearand reach an asymptote at an LAI value of approximately 3 (Asrar et al., 1984; Franklin,1986; Running et al., 1986; Baret et al., 1988; Baret and Guyot, 1991; Spanner et al., 1990;1994; Wulder et al., 1996b). Further, NDVI may even decrease as LAI increases due toincreased mutual shadowing in a mature stand.

Based upon the aforementioned considerations the application of NDVI in the estima-tion of vegetation characteristics should be undertaken with caution. Additionalinformation, such as texture, may be used to provide a context for impoved estimatesof forest structural parameters with NDVI.

3 Texture

Image spectral values alone are of limited utility in the estimation of forest structuralparameters as the spatial element of the imagery is not considered. In the context ofimage analysis, texture has many definitions, with the theme to the varied descriptionsof texture generally indicating the spatial variation in neighbouring pixel values. Theaddition of texture (the spatial variation in tones) may add structural information toestimates of forest stand parameters. The textural values act as surrogates for the actualphysical canopy composition through representing the distribution of the vegetation.Texture has been demonstrated to add structural information to the spectrally derivedvegetation indices and to improve estimates of LAI (Wulder et al., 1996b; Wulder et al.,1998). Texture has been utilized in remote sensing as ancillary information inmultispectral classifications of L-resolution imagery of vegetation cover (Frank, 1984;Franklin and Peddle, 1990; Kushwaha, 1994) and H-resolution estimation of foreststructure (Hay and Niemann, 1994; Wulder et al., 1996b; Wulder et al., 1998).The variation in texture is related to changes in the spatial distribution of terrestrialvegetation. Texture derivatives are supplemental to the image data that provide anadditional information source.

When assessing the textural characteristics of imagery, the spatial resolution ofthe imagery is an important consideration as L-resolution data have less potential fortexture than do H-resolution data. The suppressed variance environment captured byL-resolution data incorporates much of the variation in vegetation within a single pixel,thus reducing the texture. H-resolution data, which comprise pixels that are smaller thanthe objects of interest, are often highly textured.

The section on spectral vegetation indices summarized the shortcoming of NDVI inthe estimation LAI due to an asymptotic relationship for LAIs greater than � 3. Thevegetation index values saturate as a function of being derived from a remote platformwhich may only view the horizontal expression of a stand as seen from above. As LAIvalues increase, the vertical complexity of the stand also increases, proving difficult tomeasure from a nadir-viewing remote platform. For example, stands with differingvertical levels of complexity may appear the same to a horizontal viewing sensor. Thevertical structure refers to tree height distribution and the horizontal distribution pertainsto stand density and spatial distribution (St-Onge and Cavayas, 1995). Stand stratificationis primarily the result of variations in time and method of establishment and growth rate

Mike Wulder 459

of species. Accordingly, stands with varying vegetation composition and structure mayhave similar vegetation index values due to a similar vertical expression. The introduc-tion of spatially sensitive image texture variables in the estimation of LAI increases theaccuracy of LAI values obtained remotely in relation to field-collected data, especially forLAI values greater than 3 (Wulder et al., 1996b).

4 Multispectral image classi®cation

Multispectral image classifications are based upon finding patterns in the spectralresponse in relation to land-cover groups known to be present. The creation of a thematicmap of forest characteristics is often the goal of an image classification in a forestrycontext. Image classification procedures group pixels into classes, or categories, basedupon distinctive patterns of digital numbers. Image classification procedures arenormally categorized as either supervised or unsupervised. A supervised land-coverclassification uses training data input by an analyst which are based upon a prioriknolwedge of land cover at a given location. The spectral values found at the trainingsites are then applied to a multispectral classification procedure, such as maximumlikelihood, which enables the portions of the image not directly surveyed to be put intoclasses based upon similar spectral characteristics to the training areas (Jensen, 1996).Unsupervised classifications do not depend on the input of a training data set, andgenerate classes based upon clustering of the spectral values present into groups basedupon similarity (Lillesand and Kiefer, 1987). Once the spatial clusters are generated ananalyst often attempts to determine the nature of the clusters and to provide labels.

The most commonly applied classification procedure is the maximum likelihoodprocedure. Maximum likelihood is a statistical decision rule that examines the probabilityof a pixel in relation to each class with assignment of the pixel to the class with thehighest probability (Lillesand and Kiefer, 1987). The maximum likelihood procedure hasan underlying assumption of a normal distribution to the data within each class and maybe biased by unequally sized training classes. Visual image enhancements may beutilized to aid in the interpretation of imagery or for assistance in creation of trainingdata (Beaubien, 1994). The maximum likelihood procedure may be computed uponspectral values alone, or with the inclusion of ancillary data, such as a digital elevationmodel. Knowledge of the environmental characteristics of plant growth, such as anelevation gradient to tree species presence, indicates the utility of the incorporation of adigital elevation model to increase the accuracy of a multispectral image classification(Strahler et al., 1978; Franklin et al., 1986).

Contextual classifiers also allow for classification decisions to be based upon moreinformation than spectral values alone. Contextual classifiers utilize both the spectral andspatial characteristics of a pixel. In contextual classification methods the classification ofan individual pixel is influenced by the characteristics of the surrounding pixels (Gongand Howarth, 1992). Sharma and Sarkar (1998) have demonstrated a contextualclassification technique that is modifiable for either high or low-resolution imagery.Evidential reasoning is similar to contextual classification methods in its attempt toprovide an increase in information to the classifier. In a classification based uponevidential reasoning, evidential support for a given pixel is generated based upon theobserved frequency of occurrence within the training data set to create decision ruleswhich enable an image classification (Peddle et al., 1994). Evidential reasoning has beendeveloped to account for the complexity of environmental data as a tool that is robust to


large numbers of input layers from diverse sources at a variety of spatial scales and withdiffering measurement scales (Peddle, 1995).

Peddle et al. (1994) undertook a comparison of artificial neural networks, evidentialreasoning and maximum likelihood classifiers upon an alpine forest environment, wherethe neural network results were found to be superior. Neural networks are an artificialintelligence modelling technique which attempts to emulate the computational abilitiesof the human brain. In the remote-sensing context, neural networks have been appliedbecause neural networks have no underlying assumptions on the distribution of the dataand also allow for the inclusion of ancillary data in the analysis (Hepner et al., 1989).Hepner et al. (1989) demonstrated the robust nature of neural networks for a land-cover classification. In a forestry classification context, ArdoÈ et al. (1997) found resultswhen classifying forest damages to be comparable to those of a traditional classifier.Fuzzy classification methodologies have been developed to account for the tendency ofclassification algorithms to produce artificially abrupt boundaries between classes (Gathand Geva, 1989).

The spatial resolution of the imagery processed for classification will influence thelevel of detail which may be categorized through the classification procedure. Theclassification system developed by Anderson et al. (1976) acknowledges the variability ininformation content at different scales, based upon the ability to extract greater classifi-cation detail from higher-resolution imagery (or larger-scale photos). The Andersonclassification levels have become a common reference to the level of detail of groundcover captured by an image classification. Accuracy assessment of the classificationresults allows for investigation of the integrity of the resultant image classification.Congalton (1991) provides, in a forestry context, a review of accuracy assessmentprotocols allowing for a procedure for transparent appraisal of classification results.Further, an overview of accuracy assessment procedures and general guidelines may befound in Stehman and Czaplewski (1998).

The availability of H-resolution data has resulted in the need to reassess the commonapproaches to the classification of remotely sensed imagery. Traditional classificationmethods are generally based upon the lower variance image structure found withL-resolution imagery, yet with increases in spatial resolution per pixel classifiers do notnecessarily produce higher classification accuracies (Martin and Howarth, 1989). Theincrease in variance with H-resolution imagery precludes a bitmapping approach totraining a supervised classifier. Trees of the same class will often have variable spectralresponse because of such characteristics as position in the stand, age, height and health.Each tree crown which is distinguishable on H-resolution imagery will have a sunlit andshaded portion with the resultant variable spectral response. The stand areas which arein shadow or have understory species growth will also have variable spectral response.Strategies for H-resolution image classification generally require the separation of treecrowns from the background (Gougeon, 1995b). Once trees have been delineated throughthe application of digital methods, further information is available based upon the digitalnumbers found within each crown. Spectral information may be extracted from eachindividual tree crown to enable classification to species types based upon the observedspectral response. Traditional classification tools, successfully applied to L-resolutiondata (Robinove, 1981; Congalton, 1991), assume pixels to be independent and normallydistributed, which is not the case with the spectral values found within an individualtree crown. Gougeon (1995a) presents a comparison of possible multispectral classifi-cation schemes for individually delineated tree crowns.

Mike Wulder 461

5 Change detection

Change detection, or multitemporal image assessment, allows for the monitoring of thetemporal aspect of forests, through the comparison of a given location at differing pointsin time. Analyses of change between images often require accurate geometric registrationand attention to the radiometric characteristics present in the images (Varjo, 1997).Change detection analysis may be unertaken upon either classified thematic maps, imagespectral values or transformations of the image spectral values. Change detection uponclassified thematic maps requires a similar classification strategy between images and hasthe potential for problems based upon varying training data and placement of bound-aries. Lark (1995) found that the combination of multiple classifications may result in anincrease in error classes with the number of classes in each classification. Spectral methodsfor assessing the change between pixels of corrected and coregistered images may beundertaken through subtraction of pixel values for a temporal difference image, ordivision for a temporal ratio image (Muchoney and Haack, 1994). For example, Olsson(1994) demonstrates a spectral differencing technique to assess changes in reflectancecaused by forest thinning, where the measured reflectance, in reference to a benchmarkimage, decreased due to the removal of deciduous trees and increased as time progressedafter thinning. Leckie et al. (1992) assessed a time series of 13 Landsat MSS images todemonstrate the potential of predicting future stand conditions based upon an actualchange trend. Other methods for assessing change between images include analysis ofprincipal components of single bands from multiple date imagery (Fung and LeDrew,1987), change vector analysis (Lambin and Strahler, 1994) and fuzzy sets (Gong, 1993).

6 Data fusion

Data fusion may be undertaken between differing wavelengths of remotely sensedimagery or between remotely sensed data and meta-data. Electromagnetic data fusion isan attempt to exploit differing wavelengths to provide unique information to theextraction of forest structural parameters. The fusion of optical and SAR data marriesthe optical data which are sensitive to the vegetation reflective characteristics with theSAR data which are the result of the forest structural characteristics (Franklin et al., 1994).Meta-data fusion is the incorporation of remotely sensed data with geographic dataintended to provide unique locational information, such as soils or climate data (Daviset al., 1991). The meta-data are often stored in a geographical information system (GIS),which is a geographically referenced database. The integration between remotely sensedand GIS data allows for powerful modelling and analysis (Ehlers et al., 1989). The datastored in a GIS are often detailed and extensive yet static, while remotely sensed data maybe current. As a result, remotely sensed data may be used to update the forest stand datastored in a GIS. In a similar fashion, remotely sensed data may be used to break down, ordecompose, forest stand polygons stored in a GIS based upon remotely sensed data of ahigher resolution (Wulder, 1998a). Remotely sensed data also may provide currentinformation on vegetation extent and structure to computer models of forest productivity(Running and Hunt, 1994).

7 Linear mixture modelling

Linear mixture modelling, also known as spectral mixture analysis, is the reconstructionof pixel digital numbers based upon the spectral characteristics of the objects found


within the L-resolution pixel. In linear mixture modelling, an assumption is made that asmall number of materials can reproduce the observed spectra when mixed together invarious proportions (Adams et al., 1993). The small number of materials is referred to asendmembers, or components (Gong et al., 1994). Normally a spectroradiometer is utilizedto collect spectral endmembers in situ. Milton (1987) provides a summary of the steps forin situ spectral data collection. The general procedure for spectral mixture analysisrequires the collection of pure spectral signatures for all the vegetation which may befound within the study area. The endmembers, as collected with a spectroradiometer inthe field or lab, may then be used to reconstruct the pixel. The form of models utilizedfor spectral mixture analysis may be either linear (Adams et al., 1993) or nonlinear(Shimabukuro and Smith, 1994). The ability to reconstruct pixels based upon endmemberinformation is not a trivial task, often suffering from an oversimplification of the pixelspectral contents.

The requirement for mixed pixels to compute spectral fractions requires L-resolutiondata as a portion of the analysis. The potential of H-resolution data in spectralmixture analysis is in the collection of image-based endmembers. The ability to utilizeH-resolution airborne or satellite data will increase the ability to collect spectral end-members which characterize the region that is to be unmixed, providing a link betweenlocal and regional data sources. The ability to decompose the L-resolution pixeleffectively decreases the scale of the analysis. Hall et al. (1995) demonstrate the ability todiscern the spectral contents of a pixel, in conjunction with geometrical optical model-ling, to estimate forest biophysical parameters.

8 Bidirectional re¯ectance methods

The fundamental physical property governing the reflectance behaviour of any surface isthe bidirectional reflectance distribution function (BRDF). The BRDF relates irradiancefrom a given direction of incidence on a surface to the reflected radiance in the viewingdirection. The specific geometric and mathematic definition refers to directional radiationin infinitesimally small elements of solid angle which precludes its direct measurement(Deering, 1989). The BRDF is commonly thought of as being integrated into largerquantities of the solid angle, yielding the bidirectional reflectance factor (BRF) measuresactually acquired by field instruments (Milton et al., 1994). Walthall et al. (1985: 383)define BRF in simpler terms as the `reflectance at a multitude of possible view angles at agiven time or solar position'. Accordingly, the ability to collect BRF measurementsallows for spectral analyses that are not altered by sun/instrument geometry. Yet, tocharacterize the BRF of a surface, multiple spectral measurements are required from aseries of azimuths and incidence angles creating the need for directable instruments(Ranson et al., 1994) and generating large data volumes (Walthall et al., 1985). Theproblem implied by the presence of the BRDF is that the reflectance of the same forestvaries based upon the location of the sensor in relation to the sun and such factors as theunderlying topography. The potential benefits of collecting BRF measurements arenormally out-weighed by the measurement difficulties encountered, such as the need forspecialized instruments. The understanding of BRDF characteristics of forest canopies isimportant for relating remote-sensing observations of biomass, species, stand structureand albedo (Kimes et al., 1987). Ranson et al. (1994) demonstrated a relationship betweenforest structure, PAR albedo and BRF values.

Mike Wulder 463

9 Geometrical-optical models

A number of geometric-optical models has been developed to assess canopy bi-directional characteristics (see Goel, 1987, for a review), with the Li and Strahler (1985)model, developed for conifer environments, the most commonly used. The Li±Strahlerseries of models treats vegetation canopies as an assemblage of discrete, three-dimensional objects illuminated at a specific angle and casting shadows on a contrastingbackground. The areal proportions of following basic components of sunlit and shadedtree crowns, sunlit and shaded background, are determined by geometric optics andBoolean set theory. The directional reflectance of the forest is estimated by the proportionof each component weighted by independently determined spectral signatures, normallyacquired in the field, effectively performing a three-dimensional pixel decomposition.Ongoing improvements of the basic model have incorporated the effects of mutualshadowing at high solar zenith angles (Li and Strahler, 1992), topography (Schaaf et al.,1994), and the use of radiative transfer principles to describe multiple scattering effectswithin tree crowns (Li et al., 1995). Woodcock et al. (1997) have developed a technique forinversion of the Li±Strahler model, which allows for estimation of reliable estimates offorest cover, yet improvement is required for estimates of tree size. A mapping projectbased upon Landsat TM imagery and inversion of the Li±Strahler model is described inWoodcock et al. (1994).

10 Hyperspectral image analysis

Other information which may be remotely sensed to provide insights into the state offorest vegetation is hyperspectral image data. Imaging spectrometers provide spectralinformation for a large number of narrow spectral bands for each pixel, resulting in a fullspectral curve for each pixel (Vane and Goetz, 1993). Research with spectroradiometerdata has demonstrated the potential for detailed analysis of vegetation spectralcharacteristics based upon absorption features (Curran, 1989), spectral characteristics(Bracher and Grant, 1994), or derivative spectroscopy (Demetriades-Shah et al., 1990).Distinct absorption features relate to chemicals that are found at discrete locations on theelectromagnetic spectrum. Of interest in the assessment of forest structure are thelocations of chlorophyll a and b, which are found at 0.64 and 0.66 mm (Curran, 1989) andthat provide an indication of tree health and vigour. The ability to collect full spectralinformation allows for the detailed assessment of the spectral characteristics of plants,such as the `red edge' which relates to plant vigour and chlorophyll (Bracher and Grant,1994). Derivative spectroscopy is an analysis technique borrowed from analyticalchemistry used for the elimination of unwanted background signal and for resolving andenhancing overlapping spectral features (Demetriades-Shah et al., 1990).

At present there are more than a dozen airborne and spaceborne imaging spectro-meters (Vane and Goetz, 1993). Johnson et al. (1994) assessed the utility of AirborneVisible/Infrared Imaging Spectrometer (AVIRIS) in the prediction of canopy biochem-istry as a component of the OTTER project. The red-edge spectral region was found to bestrongly related to LAI, canopy total nitrogen, and canopy chlorophyll content, whileinsignificant results were found for starch concentration. As a component of the OTTERproject, Gong et al. (1992) found a strong relationship between casi hyperspectral dataand a LAI from 0 to 3. A limitation that has hampered existing instruments for detailedforest structural analysis is L-resolution pixel size, which is often also irregularly shaped


due to longer along-track integration time in comparison to across track. Niemann (1995)investigated the relationship between stand age with L-resolution hyperspectral casi data.The analysis found a poor relationship between red edge, derivative spectra, and bandratios with age classes.

11 Individual tree isolation

The ability to delineate tree crowns of a forest digitally allows for improved estimation ofinventory elements such as density, volume and canopy closure. Forest stand densityestimation has been attempted on L-resolution data with an accuracy too low forinventory usage, but high enough for regional density characterization (Franklin, 1994;Wu and Strahler, 1994). Methods are being developed to estimate forest structure directlythrough the delineation of the actual objects of interest. Once delineated, trees may beclassified, separated from the understory or spectrally assessed. An international forumon the `Automated interpretation of high spatial resolution imagery for forestry' demon-strated several current approaches to the delineation of tree crowns, such as valleyfollowing (Gougeon, 1998), radiance peak filtering (Niemann and Adams, 1998), edgefinding (Pinz, 1998), template matching (Pollock, 1998), morphology (Barbezat and Jacot,1998), and clustering (Culvenor et al., 1998; Wulder, 1998b).

An example of the valley-following approach to stem counting and classificationhas been demonstrated upon H-resolution MEIS (Multi-detector Electro-optical ImagingSensor) multispectral scanner data of conifer forests. Image-processing techniques areapplied for the isolation of conifer tree species (Gougeon, 1995b), and subsequentclassification schemes for the crowns delineated on the H-resolution imagery (Gougeon,1995a). The tree delineation approach outlined by Gougeon (1995b) considers the darkunderstory areas between trees as `valleys' and attempts to join together the com-paratively low valley digital numbers. An iterative approach to the joining of valley pixelsresults in nearly delineated trees requiring further sharpening with a rule-based strategy.The rule-based classification is based upon spatial discrimination of the trees consideredas objects, and applying techniques borrowed from manual aerial photo interpretation.

A stem isolation approach based on radiance peak filtering has been presented by Hayet al. (1996). Radiance peak filtering is based upon the premise that each tree in anH-resolution scene has a bright pixel which represents the apex of the tree. Passing afilter over the image and seeking the highest digital number within the filter kernelprovide an estimate of tree stem locations (Dralle and Rudemo, 1997). Normally a kernelof a fixed size is passed over the image, which does not account for the presence of treesof a variety of sizes. Hay et al. (1996) used semivariance to customized the kernel size forthe extraction of the radiance peaks. Daley et al. (1998) dynamically allocated the kernelsize for each pixel location based upon the local scene structure. The problem common toradiance peak filtering is a large number of pixels indicated as trees that are not. Theproblem of false positives has been investigated by Burnett et al. (1998) through theintroduction of spatial autocorrelation and spectral directionality.

Edge-finding techniques for the isolation of tree crowns often work in conjunctionwith a radiance peak filter. Once the radiance peak is isolated, the edges to the imageobject may be delineated based upon an iterative search approach (Pinz, 1998). Template-matching approaches are based upon the spectral characteristics found within an imagescene compared to synthetic image templates. The synthetic image templates aregenerated based upon known stand characteristics which may be matched to image

Mike Wulder 465

scenes to indicate the crown distribution present in the imagery (Larsen, 1998; Pollock,1998). The morphology of digitized airphotos has also been presented as a means toextract tree crowns (Barbezat and Jacot, 1998).

The overlap of foliage common to deciduous species results in difficulty in thedelineation of individual stems. The foliage of codominant species is found to overlap,while suppressed growth beneath the more dominant trees provides confusionspectrally. Wulder (1998b) has approached the problem through generation of clustersbased upon spatially dependent groups of pixels. On 1 m spatial resolution imagery of adeciduous forest the clusters represent either a large single tree or a cluster of spectrallysimilar trees. The clusters of trees may be spectrally classified, provide clues to crownclosure and also indicate the stand-level structure. Culvenor et al. (1998) demonstrate ahybrid approach to clustering assisted by valley following to isolate the clusters.

The interest by the forestry and ecological communities to possess an accurate account-ing of all trees within a management unit or an ecosystem is illustrated by the variety oftree crown delineation approaches. Factors which affect the remote sensing of forestcanopies also have an effect on the efficacy of the tree crown delineation approaches. Therelationship between the image resolution and the tree size is important, as the ability toreconstruct trees with a number of individual pixels is often required. Cover type is alsoimportant as conifers are generally more readily delineated than deciduous species. Thedifficulty in finding an automated approach to tree crown delineation that is robust toforest conditions indicates the utility of a system which pairs the automated crowndetection approaches with a photointerpretation approach (Leckie et al., 1998).

12 Spatial discriminators

Shape information has been shown to increase the accuracy of image classification resultswith L-resolution imagery (Xia, 1996). The utility of shape information has been seen as ameans to increase the ability to discriminate spatially between objects which have beenisolated on H-resolution imagery. The ability to discern individual conifer trees has beendemonstrated in the previous section, based upon spectral differences between theoverstory and understory and the implementation of rule-based procedures. The rule-based procedures which are implemented attempt to introduce techniques utilized in thevisual interpretation of aerial photographs based upon shape, size and distribution(Gougeon, 1995b). Fournier et al. (1995) present a catalogue of spatial discriminators forpotential implementation of a rule-based delineation procedure based upon crownoutline, bright areas, crown radiometric profiles and tree shadows. Visual discriminationof stems was undertaken on 40 cm MEIS data to assess the utility and validity of thecatalogue of spatial discriminators. The accuracies found from visual interpretation of theMEIS imagery based upon the spatial discrimination classification rules were found to bebelow the accuracies obtained from spectral information in comparable studies (Leckieand Dombrowski, 1984; Leckie et al., 1995). The image spatial information showspotential for future integration to rule-based crown identification procedures as higherspatial resolution imagery becomes available.

13 Image semivariance

A variogram describes the magnitude, spatial scale and general form of the variation in agiven set of data (Matheron, 1963). Semivariance is the variance per site when sites areconsidered as profiles or areas of pixels, and is developed from the theory of regionalized


variables (Curran, 1988; Woodcock et al., 1988a). A thorough review of semivariance andgeostatistics in remote sensing is presented in Curran and Atkinson (1998). Variogramshave been the tool used to link models of ground scenes to spatial variation in images(Woodcock et al., 1988b). Semivariograms are a graphical representation of the spatialvariability, and provide a means of measuring the spatial dependency of continuouslyvarying phenomena. Figure 1 illustrates the general form of a semivariogram, where thenugget is a measure of the intrapixel variation, the sill is a the point at which variancepeaks and the range indicates the number of lags, or distance, to reach the sill.Accordingly, the range represents the point at which the similarity of the pixels to theinitial pixel is no more marked than to other pixels along the calculation transect. Cohenet al. (1990) describe the semivariance response in terms of forest canopy structure.

Image semivariance has been used extensively in the assessment of L-resolution foreststructure (see Cohen et al., 1990; Bowers et al., 1994, for examples; Ramstein and Raffy,1989, for the theory). In analysis of H-resolution forest imagery each pixel represents nearpure spectral characteristics of an object. Image semivariance has been demonstrated torepresent image structural information (Franklin and McDermid, 1993). St-Onge andCavayas have utilized the information inherent in the directional variogram as a methodto estimate the stocking and height of forest stands on simulated H-resolution data (1995)and MEIS data (1997), which are also both strongly related to LAI. Semivariance responsehas also been exploited in the multivariate estimation of LAI as an image texturalindicator based upon extraction of discrete points from the semivariance response curve(Wulder et al., 1998), or similarly as a textural classifier (Carr and Myers, 1984).

14 Spatial autocorrelation

Spatial autocorrelation is present when a particular value, or attribute, is found to exhibitsome degree of dependence to location. Thus, the measurement of spatial autocorrelationinvolves the simultaneous consideration of both locational and attribute information

Figure 1 Sample semivariogram showing nugget, sill, range and curve form

Mike Wulder 467

(Goodchild, 1986). In the case of remotely sensed images, the locations are pixels and theattribute data are the digital numbers. As remote-sensing methods regularize continuouslandscapes into a grid of equally sized and regularly spaced data in the form of pixels(Fisher, 1997), it is anticipated that there will be some degree of dependency betweenpixels. Further, this dependency is likely to take the form of positive spatial auto-correlation, with similar values found in close association. In general, the level of spatialautocorrelation is not dependent of the scale at which these data are analysed, withnegative spatial autocorrelation being more sensitive to scale changes (Chou, 1991).

The autocorrelated nature of imagery was noted early in the assessment of remotelysensed data (Craig, 1979) and the effect that image autocorrelation would have upontraditional classification techniques was noted by Campbell (1981). Later, studies ofimage spatial autocorrelation examined pixel inter-relationships using a scanline tech-nique (Labovitz and Masuoka, 1984). This approach, however, only allows spatialdependency to be assessed in a limited number of directions. In another approach,semivariance is often applied in the assessment of autocorrelation, with the observationsfound within the range considered as the limit of autocorrelation (Jupp et al., 1989;Franklin et al., 1996). Wulder and Boots (1998) summarize the traditional methods thathave been applied to assess autocorrelation in remotely sensed imagery and present acomplementary approach based upon the spatial dependence exhibited between pixels.

Spatial autocorrelation may be measured using either global or local statistics. Globalindicators provide a single measure summarizing all the spatially referenced inter-relationships. These measures may prove unreliable if the nature and extent of spatialautocorrelation vary significantly over the study area (Haining, 1990). To remedy theseconcerns, local indicators of spatial association (LISA) have been developed (Anselin,1995). Getis (1994) investigated the potential of the spatial dependency relationships ofremotely sensed imagery. Wulder and Boots (1998) present a statistic and procedure forthe extraction of spatial dependence information from remotely sensed imagery. In onestudy, investigation of the spatial dependence characteristics of Landsat TM image of amanaged forest region indicates a strong Landsat TM channel and cover-type depend-ence to local spatial autocorrelation. Analyses of H-resolution forest imagery demon-strate a local spatial dependency of image objects, which allow for the extraction of treesand clusters of trees (Wulder, 1998b).

III Summary

The future composition of forests due to the effects of climate variability can currentlyonly be hypothesized upon. Monitoring changes in forest structure will provide theability to assess forests from a measurement baseline and assess changes againstthis baseline. The techniques made possible through the collection of remotely senseddata and image processing provide the ability to assess changes in forest structure.The incorporation of techniques and methods from forest science and inventory withH-resolution imagery is providing for unique solutions to persistent problems. Successfulassessment of forest structure with high spatial resolution instruments may borrow fromthe established L-resolution satellite techniques and develop new techniques suited to theproperties of the new technology. Yet, in many applications the potential increase inaccuracy of estimates from higher-resolution imagery may be offset by the regionalcoverage which is possible with lower-resolution sensors.


The image-processing techniques presented illustrate the availability of a variety ofpotential information extraction options dependent upon the goals of a particularapplication. Also demonstrated by the presentation of image-processing techniques is thegreater maturity of spectral techniques in relation to spatial techniques. Spatial informa-tion extraction techniques, especially for H-resolution imagery, provide for strongerrelationships to be created between the image and ground data. The relationshipsdeveloped between image spectral response and forest inventory parameters may nowbe augmented with detailed spatial information. Parameters may be estimated empiric-ally, such as with regression-based models, or in a more deterministic fashion, such asderiving stem counts from an image of isolated tree crowns. The combination ofempirical and deterministic approaches may provide for an increase in the ability toestimate forest inventory and biophysical parameters. Spatial statistical techniques mayallow for an improved linkage between image spectral and spatial information.

Due to the environmental and economic importance of forests, the demand forinformation related to the structure and composition of forests will only increase withtime. The increasing sophistication of techniques for analysis has resulted in increasedaccuracy in the prediction of forest inventory and biophysical parameters from remotelysensed imagery. The addition of high-resolution satellites to the currently available suiteof aerial photographs, airborne multispectral imagery and existing satellite instrumentswill provide for continued growth in the use of remote-sensing technologies in forestry.

Acknowledgements

The author is grateful to Dr Ellsworth LeDrew (University of Waterloo), Dr Barry Boots(Wilfrid Laurier University) and Dr Philip Howarth (University of Waterloo) for adviceand encouragement throughout his graduate career at the University of Waterloo.Thanks are also extended to Dr Don Leckie (Canadian Forest Service) for offeringthoughtful suggestions which improved this manuscript.

Note

1. The author was formerly of Waterloo Laboratory for Earth Observations, Department ofGeography, University of Waterloo, Ontario, Canada.

References

Adams, J., Smith, M. and Gillespie, A. 1993:

Imaging spectroscopy: interpretation based upon

spectral mixture analysis. In Pieters, C. and

Englert, P., editors, Topics in remote sensing andremote geochemical analysis: elemental and miner-alogical composition, Cambridge: Cambridge Uni-

versity Press, 145±66.

Anderson, J., Hardy, E., Roach, J. and Witmer, R.

1976: A land use land cover classification

system for use with remote sensor data. Geological

Survey Professional Paper 964. Washington, DC:US Government Printing Office.

Anselin, L. 1995: Local indicators of spatial associa-tion ± LISA. Geographical Analysis 27, 93±115.

Aplin, P., Atkinson, P. and Curran, P. 1997: Finespatial resolution satellite sensors for the nextdecade. International Journal of Remote Sensing 18,3873±81.

ArdoÈ , J., PilesjoÈ , P. and Skidmore, A. 1997: Neuralnetworks, multitemporal Landsat Thematic

Mike Wulder 469

Mapper data and topographic data to classifyforest damages in the Czeck Republic. CanadianJournal of Remote Sensing 23, 217±29.

Asrar, G., Fuchs, M., Kanemasu, E. and Hatfield,J. 1984: Estimating absorbed photosyntheticradiation and leaf area index from spectralreflectance in wheat. Agronomy Journal 76,300±306.

Avery, T. and Berlin, G. 1992: Fundamentals ofremote sensing and airphoto interpretation (5th edn).Toronto: Maxwell Macmillan.

Barbezat, V. and Jacot, J. 1998: The CLAPA project:automated classification of forest with aerialphotographs. In Proceedings of the InternationalForum on Automated Interpretation of High SpatialResolution Digital Imagery for Forestry, Victoria,BC: Natural Resources Canada, ForestryCanada, Pacific Forestry Centre, 10±12 February,in press.

Baret, F. and Guyot, G. 1991: Potentials and limitsof vegetation indices for LAI and APAR assess-ment. Remote Sensing of Environment 35, 161±73.

Baret, F., Guyot, G., Begue, A., Maurel, P. andPodaire, A. 1988: Complementarity of middle-infrared with visible and near-infrared reflectancefor monitoring wheat canopies. Remote Sensing ofEnvironment 26, 213±25.

Baulies, X. and Pons, X. 1995: Approach toforest inventory and mapping by means ofmultispectral airborne data. International Journalof Remote Sensing 16, 61±80.

Beaubien, J. 1994: Landsat TM satellite images offorests: from enhancement to classification.Canadian Journal of Remote Sensing 20, 17±26.

Bonham, C. 1989: Measurements for terrestrial vegeta-tion. New York: Wiley.

Bowers, W., Franklin, S., Hudak, J. andMcDermid, G. 1994: Forest structural damageanalysis using image semivariance. CanadianJournal of Remote Sensing 20, 28±36.

Bracher, G. and Grant, P. 1994: Estimation of foliarmacro-nutrients and chlorophyll in douglas-firseedlings by leaf reflectance. Canadian Journal ofRemote Sensing 20, 102±15.

Brand, D. 1990: Advances in Canadian forestresearch: an introduction. Canadian Journal ofForest Research 20, 373±74.

Bryant, E., Dodge, A. and Warren, S. 1980: Landsatfor practical forest type mapping: a test case.Photogrammetric Engineering and Remote Sensing46, 1575±84.

Burnett, C., Daley, N., Wulder, M., Niemann, K.and Goodenough, D. 1998: Directional varia-bility in 1 m casi imagery: an analysis usingsemi-variance range and slope shape. InIGARSS'98, Seattle, WA, 6±10 July, 2342±44.

Campbell, J. 1981: Spatial correlation effectsupon accuracy of supervised classification ofland cover. Photogrammetric Engineering andRemote Sensing 47, 355±63.

Canadian Council of Forest Ministers (CCFM)1995: Defining sustainable forest management:a Candian approach to criteria and indicators.Ottawa: Natural Resources Canada.

Carr, J. and Myers, D. 1984: Application of thetheory of regionalized variables to the spatialanalysis of Landsat data. In Proceedings, Pecora 9Spatial Information Technologies for Remote SensingToday and Tomorrow, Sioux Falls, SD: IEEEComputer Society Press, 55±61.

Chavez, P. 1992: Comparison of spatial variabilityin visible and near-infrared spectral images.Photogrammetric Engineering and Remote Sensing58, 957±64.

Chen, J. 1996: Evaluation of vegetation indicesand a modified simple ratio for boreal applica-tions. Canadian Journal of Remote Sensing 22,229±42.

Chen, J. and Guilbeault, M. 1996: Examination ofvegetation indices for retrieving biophysicalparameters of boreal conifer forests. In Proceed-ings of the 26th International Symposium on RemoteSensing of the Environment and 18th AnnualSymposium of the Canadian Remote Sensing Society,Vancouver, 168±71.

Chou, Y.H. 1991: Map resolution and spatialautocorrelation. Geographical Analysis 23, 228±46.

Civco, D. 1989: Topographic normalization ofLandsat Thematic Mapper digital imagery. Photo-grammetric Engineering and Remote Sensing 55,1303±309.

Cohen, W., Spies, T. and Bradshaw, G. 1990:Semivariograms of digital imagery for analysis ofconifer canopy structure. Remote Sensing ofEnvironment 35, 167±78.

Congalton, R. 1991: A review of assessing theaccuracy of classifications of remotely senseddata. Remote Sensing of Environment 46, 35±46.

Craig, R. 1979: Autocorrelation in Landsat data. InProceedings of the Thirteenth International Sym-posium on Remote Sensing of Environment, Vol. 3,Ann Arbor, MI: Environmental Research Instituteof Michigan, 23±27.

ÐÐ 1981: Precision in the evaluation of Landsatautocorrelation: the terrain effect. In ProceedingsFifteenth International Symposium on RemoteSensing of Environment, Ann Arbor, MI, May,1305±14.

Culvenor, D., Coops, N., Preston, R. and Tolhurst,K. 1998: A spatial clustering approach to auto-mated tree crown delineation. In Proceedings of theInternational Forum on Automated Interpretation ofHighSpatial Resolution Digital Imagery for Forestry,


Victoria, BC: Natural Resources Canada, ForestryCanada, Pacific Forestry Centre, 10±12 February,in press.

Curran, P. 1980: Multispectral remote sensing ofvegetation amount. Progress in Physical Geography4, 315±41.

ÐÐ 1988: The semivariogram in remote sensing:an introduction. Remote Sensing of Environment24, 493±507.

ÐÐ 1989: Remote sensing of foliar chemistry.Remote Sensing of Environment 30, 271±78.

Curran, P. and Atkinson, P. (1998) Geostatisticsand remote sensing. Progress in Physical Geogra-phy 22, 61±78.

Daley, N., Burnett, C., Wulder, M., Niemann, K.and Goodenough, D. 1998: Comparison of fixed-size and variable-sized windows for the esti-mation of tree crown position. In: IGARSS'98,Seattle, WA, 6±10 July, 1323±25.

Danson, F. 1987: Preliminary evaluation of therelationship between SPOT-1 HRV data andforest stand parameters. International Journal ofRemote Sensing 8, 1571±75.

Danson, F. and Curran, P. 1993: Factors affectingthe remotely sensed response of coniferous forestplantations. Remote Sensing of Environment 43,55±65.

Davis, F., Quattrochi, D., Ridd, M., Lam, N.,Walsh, S., Michaeksen, J., Franklin, J., Stow, D.,Johannsen, C. and Johnston, C. 1991: Environ-mental analysis using integrated GIS and remo-tely sensed data: some research needs andpriorities. Photogrammetric Engineering andRemote Sensing 57, 689±97.

Deering, D. 1989: Field measurements of bidirec-tional reflectance. In Asrar, G., editor, Theory andapplications of optical remote sensing, Toronto:Wiley, 14±65.

de Haan, J., Hovenier, J., Kokke, J. and vanStokkom, H. 1991: Removal of atmosphericinfluences on satellite-borne imagery: a radiativetransfer approach. Remote Sensing of Environment46, 1±21.

Demetriades-Shah, T., Steven, M. and Clark, J.1990: High resolution derivative spectra in remotesensing. Remote Sensing of Environment 33, 55±64.

de Wulf, R., Goossens, R., de Roover, B. andBorry, F. 1990: Extraction of forest stand para-meters from panchromatic and multispectralSPOT-1 data. International Journal of RemoteSensing 11, 1571±88.

Dralle, K. and Rudemo, M. 1997: Stem numberestimation by kernel smoothing of aerial photos.Canadian Journal of Forest Research 26, 1228±36.

Dymond, J. 1992: Nonparametric modeling ofradiance in hill country for digital classification

of aerial photographs. Remote Sensing of Environ-ment 39, 95±102.

Ehlers, M., Edwards, G. and BeÂdard, Y. 1989:Integration of remote sensing with geographicinformation systems: a necessary evolution.Photogrammetric Engineering and Remote Sensing55, 1619±27.

Fisher, P. 1997: The pixel: a snare and adelusion. International Journal of Remote Sensing18, 679±85.

Fournier, R., Edwards, G. and Eldridge, N. 1995:A catalogue of potential spatial discriminatorsfor high resolution digital images of individualtree crowns. Canadian Journal of Remote Sensing21, 285±98.

Frank, T. 1984: The effect of change in vegetationcover and erosion patterns on albedo and textureof Landsat images in a semiarid environment.Annals of the Association of American Geographers74, 393±407.

Franklin, J. 1986: Thematic mapper analysis ofconiferous forest structure and composition.International Journal of Remote Sensing 7, 1287±301.

Franklin, J., Logan, T., Woodcock, C. and Strahler,A. 1986: Coniferous forest classification andinventory using Landsat and digital terrain data.IEEE Transactions on Geoscience and Remote Sensing24, 139±49.

Franklin, S. 1994: Discrimination of subalpineforest species and canopy density using digitalCASI, SPOT PLA, and Landsat TM data. Photo-grammetric Engineering and Remote Sensing 60,1233±41.

Franklin, S., Lavigne, M., Wilson, B. and Hunt, E.R.Jr 1994: Empirical relations between balsam fir(Abies balsamea) forest stand conditions and ERS-1SAR data in western Newfoundland. CanadianJournal of Remote Sensing 20, 124±30.

Franklin, S. and McDermid, G. 1993: Empiricalrelations between digital SPOT HRV and CASIspectral response and lodgepole pine (Pinuscontorta) forest stand parameters. InternationalJournal of Remote Sensing 14, 2331±48.

Franklin, S. and Peddle, D. 1990: Classification ofSPOT HRV imagery and texture features. Inter-national Journal of Remote Sensing 11, 551±56.

Franklin, S., Wulder, M. and Lavigne, M. 1996:Automated derivation of geographic windowsfor use in remote sensing digital image textureanalysis. Computers and Geosciences 22, 665±73.

Fritz, L. 1996: The era of commercial earth observa-tion satellites. Photogrammetric Engineering andRemote Sensing 62, 39±45.

Fung, T. and LeDrew, E. 1987: Applicationof principal components analysis to change

Mike Wulder 471

detection. Photogrammetric Engineering and RemoteSensing 53, 1649±58.

Gates, D. 1990: Climate change and the response offorests. International Journal of Remote Sensing 11,1095±107.

Gath, I. and Geva, A. 1989: Unsupervised optimalfuzzy clustering. IEEE Transactions on PatternAnalysis and Machine Intelligence II, 773±89.

Gausman, H. 1977: Reflectance of leaf components.Remote Sensing of Environment 6, 1±9.

Gemmel, F. 1995: Effects of forest cover, terrain,and scale on timber volume estimation withthematic mapper data in a rocky mountain site.Remote Sensing of Environment 51, 291±305.

Getis, A. 1994: Spatial dependence and hetero-geneity and proximal databases. In Fother-ingham, S. and Rogerson, P., editors, Spatialanalysis and GIS, London: Taylor & Francis,105±20.

Gholz, H. 1982: Environmental limits on aboveground net primary production, leaf area, andbiomass in vegetation zones of the Pacific North-west. Ecology 54, 152±59.

Gillis, M. and Leckie, D. 1993: Forest inventorymapping procedures across Canada. InformationReport PI-X-122. Petawawa: Natural ResourcesCanada, Petawawa National Forestry Institute.

ÐÐ 1996: Forest inventory update in Canada. TheForestry Chronicle 72, 138±56.

Goel, N.S. 1987: Models of vegetation canopyreflectance and their use in estimation of bio-physical parameters from reflectance data. RemoteSensing Reviews 3, 1±212.

Gong, P. 1993: Change detection using principalcomponents analysis and fuzzy set theory.Canadian Journal of Remote Sensing 19, 22±29.

Gong, P. and Howarth, P. 1992: Frequency-based contextual classification and gray-levelvector reduction for land-use identification.Photogrammetric Engineering and Remote Sensing58, 423±37.

Gong, P., Pu, R. and Miller, J. 1992: Correlatingleaf area index of ponderosa pine with hyper-spectral casi data. Canadian Journal of RemoteSensing 18, 275±82.

Gong, P., Miller, J. and Spanner, M. 1994: Forestcanopy closure from classification and spectralunmixing of scene components ± multisensorevaluation of an open canopy. IEEE Transactionson Geoscience and Remote Sensing 32, 1067±79.

Goodchild, M. 1986: Spatial autocorrelation.In Concepts and techniques in modern geography,Vol. 47, Norwich: Geo Books, 3±6.

Gougeon, F. 1995a: Comparison of possible multi-spectral classification schemes for tree crownsindividually delineated on high spatial resolution

MEIS images. Canadian Journal of Remote Sensing21, 1±9.

ÐÐ 1995b: A crown-following approach to theautomatic delineation of individual tree crownsin high spatial resolution aerial images. CanadianJournal of Remote Sensing 21, 274±84.

ÐÐ 1998: Automatic individual tree crowndelineation using a valley-following algorithmand rule-based system. In Proceedings of theInternational Forum on Automated Interpretation ofHigh Spatial Resolution Digital Imagery for Forestry,Victoria, BC: Natural Resources Canada, ForestryCanada, Pacific Forestry Centre, 10±12 February,in press.

Guyot, G., Guyon, D. and Riom, J. 1989: Factorsaffecting the spectral response of forest canopies:a review. Geocarto International 3, 3±18.

Haining, R. 1990: Spatial data analysis in thesocial and environmental sciences. Cambridge:Cambridge University Press.

Hall, F., Shimabukuro, Y. and Huemmrich, K.1995: Remote sensing of forest biophysicalstructure using mixture decomposition and geo-metric reflectance models. Ecological Applications5, 993±1013.

Hay, G. and Niemann, K. 1994: Visualizing 3-Dtexture: a three-dimensional approach to modelforest texture. Canadian Journal of Remote Sensing20, 89±101.

Hay, G., Niemann, K. and McLean, G. 1996: Anobject-specific image texture analysis of H-resolu-tion forest imagery. Remote Sensing of Environ-ment 55, 108±22.

Hepner, D., Logan, T., Rittner, N. and Bryant,N. 1989: Artificial neural network classificationusing a minimal training set: comparison toconventional supervised classification. Photo-grammetric Engineering and Remote Sensing 56,469±73.

Heute, A. 1988: A soil-adjusted vegetation index.Remote Sensing of Environment 25, 295±309.

Irons, J., Markham, B., Nelson, R., Toll, D.,Williams, D., Latty, R. and Stauffer, M. 1985:The effects of spatial resolution of the classifi-cation of thematic mapper data. InternationalJournal of Remote Sensing 6, 1385±403.

Jarvis, P. and Dewar, R. 1993: Forest in the globalcarbon balance: from stand to region. In Ehler-inger, J. and Field, C., editors, Scaling physiologicalprocesses: leaf to globe, Toronto: Academic Press,191±221.

Jensen, J. 1996: Introductory digital processing: aremote sensing perspective. Englewood Cliffs, NJ:Prentice-Hall.


Johnson, L., Hlavka, C. and Peterson, D. 1994:Multivariate analysis of AVIRIS data for canopybiochemical estimation along the Oregon trans-ect. Remote Sensing of Environment 47, 216±30.

Jupp, D., Strahler, A. and Woodcock, C. 1988:Autocorrelation and regularization in digitalimages. I. Basic theory. IEEE Transactions onGeoscience and Remote Sensing 26, 463±73.

ÐÐ 1989: Autocorrelation and regularization indigital images. II. Simple image models. IEEETransactions on Geoscience and Remote Sensing 27,463±73.

Kimes, D.S., Sellers, P.J. and Diner, D.J. 1987:Extraction of spectral hemispherical reflectance(albedo) of surfaces from nadir and directionalreflectance data. International Journal of RemoteSensing 8, 1727±46.

King, D. 1995: Airborne multispectral digitalcamera and video sensors: a critical review ofsystem designs and applications. Canadian Journalof Remote Sensing 21, 245±73.

Kozlowski, T., Kramer, P. and Pallardy, S. 1991:The physiological ecology of woody plants. Toronto:Academic Press.

Kushwaha, S., Kuntz, S. and Oesten, G. 1994:Applications of image texture in forest classifi-cation. International Journal of Remote Sensing 15,2273±84.

Labovitz, M. and Masuoka, E. 1984: The influenceof autocorrelation in signature extraction ±an example from a geobotanical investigation ofCotter Basin, Montana. International Journal ofRemote Sensing 5, 315±32.

Lambin, E. and Strahler, A. 1994: Change-vectoranalysis in multitemporal space: a tool to detectand categorize land-cover change processes usinghigh temporal-resolution satellite data. RemoteSensing of Environment 48, 231±44.

Lark, R. 1995: Components of the accuracy of mapswith special reference to discriminant analysis onremote sensor data. International Journal of RemoteSensing 16, 1461±80.

Larsen, M. 1998: Finding an optimal match windowfor spruce top detection based on an optical treemodel. In Proceedings of the International Forum onAutomated Interpretation of High Spatial ResolutionDigital Imagery for Forestry, Victoria, BC: NaturalResources Canada, Forestry Canada, PacificForestry Centre, 10±12 February, in press.

Law, B. and Waring, R. 1994: Remote sensing of leafarea index and radiation intercepted by under-story vegetation. Ecological Applications 4, 272±79.

Leckie, D. 1990: Advances in remote sensingtechnologies for forest survey and management.Canadian Journal of Forest Research 20, 464±83.

ÐÐ 1998: Forestry applications of imaging radar.In Manual of remote sensing: principles and applica-tions of imaging radar, Bethesda, MD: AmericanSociety of Photogrammetry and Remote Sensing,in press.

Leckie, D., Beaubien, J., Gibson, J., O'Niell, N.,Piekutowski, T. and Joyce, S. 1995: Data proces-sing and analysis for MIFUCAM: a trial for MEISimagery for forest inventory mapping. CanadianJournal of Remote Sensing 21, 337±56.

Leckie, D. and Dombrowski, A. 1984: Enhance-ment of high resolution MEIS II data for soft-wood species discrimination. In Proceedings, 9thCanadian Symposium on Remote Sensing, St John's,Newfoundland, 617±26.

Leckie, D., Gillis, M., Gougeon, F., Lodin, M.,Wakelin, J. and Yuan, X. 1998: Computer-assisted photointerpretation aids to forest invent-ory mapping: some possible approaches. InProceedings of the International Forum on AutomatedInterpretation of High Spatial Resolution Imagery forForestry, Victoria, BC: Canadian Forest Service,Pacific Forestry Centre, 10±12 February, in press.

Leckie, D., Gillis, M. and Joyce, S. 1992: A forestmonitoring system based on satellite imagery.In Proceedings of the 15th Canadian Symposium onRemote Sensing, Toronto, 1±4 June, 85±90.

Li, X. and Strahler, A. 1985: Geometric-opticalmodeling of a conifer forest canopy. IEEE Trans-actions on Geoscience and Remote Science GE-23,705±21.

ÐÐ 1992: Geometric-optical bidirectional reflec-tance modeling of the discrete crown vegetationcanopy: effect of crown shape and mutualshadowing. IEEE Transactions on Geoscience andRemote Sensing 30, 276±92.

Li, X., Strahler, A.H. and Woodcock, C.E. 1995: Ahybrid geometric optical-radiative transferapproach for modeling albedo and directionalreflectance of discontinuous canopies. IEEETransactions on Geoscience and Remote Sensing 33,466±80.

Lillesand, T. and Kiefer, R. 1987: Remote sensingand image interpretation. Toronto: Wiley.

Marceau, D., Howarth, P. and Gratton, D. 1994a:Remote sensing and the measurement of geo-graphical entities in a forested environment. 1.The scale and spatial aggregation problem.Remote Sensing of Environment 49, 93±104.

Marceau, D., Gratton, D., Fournier, R. and Fortin,J. 1994b: Remote sensing and the measurement ofgeographical entities in a forested environment.2. The optimal spatial resolution. Remote Sensingof Environment 49, 105±17.

Martin, L. and Howarth, P. 1989: Change-detectionaccuracy assessment using SPOT multispectral

Mike Wulder 473

imagery of the rural±urban fringe. Remote Sensingof Environment 30, 55±66.

Matheron, G. 1963: Principles of geostatistics.Economic Geology 58, 1246±66.

Milton, E. 1987: Principles of field spectroscopy.International Journal of Remote Sensing 8, 1807±27.

Milton, T., Blackburn, G., Rollin, E. and Danson,F. 1994: Measurement of the spectral directionalreflectance of forest canopies: a review ofmethods and a practical application. RemoteSensing Reviews 10, 285±308.

Muchoney, D. and Haack, B. 1994: Changedetection for monitoring forest defoliation. Photo-grammetric Engineering and Remote Sensing 60,1243±51.

Myneni, R., Hall, F., Sellers, P. and Marshak, A.1995: The interpretation of spectral vegetationindexes. IEEE Transactions on Geoscience andRemote Sensing 33, 481±86.

Natural Resources Canada (NRC) 1995: The state ofCanada's forests. Ottawa: NRC.

Nemani, R., Pierce, L., Running, S. and Band, L.1993: Forest ecosystem processes at the watershedscale: sensitivity to remotely-sensed leaf areaindex estimates. International Journal of RemoteSensing 14, 2519±34.

Niemann, K.O. 1995: Remote sensing of foreststand age using airborne spectrometer data.Photogrammetric Engineering and Remote Sensing61, 1119±27.

Niemann, K.O. and Adams, S. 1998: Automatedtree crown identification using digital ortho-photo mosaics. In Proceedings of the Inter-national Forum on Automated Interpretation ofHigh Spatial Resolution Digital Imagery for Forestry,Victoria, BC: Natural Resources Canada, ForestryCanada, Pacific Forestry Centre, 10±12 February,in press.

Olsson, H. 1994: Changes in satellite-measuredreflectances caused by thinning cuttings in borealforest. Remote Sensing of Environment 50, 221±30.

Openshaw, S. 1984: The modifiable areal unit problem,concepts and techniques in modern geography.CATMOG Series, Vol. 38. Norwich: GeoBooks.

Peddle, D. 1995: Knowledge formulation for super-vised evidential classification. PhotogrammetricEngineering and Remote Sensing 61, 409±17.

Peddle, D., Foody, G., Zhang, A., Franklin, S. andLeDrew, E. 1994: Multi-source image classifi-cation. II. An empirical comparison of evidentialreasoning and neural network approaches. Cana-dian Journal of Remote Sensing 20, 396±407.

Perry, D. 1994: Forest ecosystems. Baltimore, MD:Johns Hopkins University Press.

Peterson, D. and Running, S. 1989:Applicationsin forest science and management. in Asrar, G.,editor, Theory and applications of optical remotesensing, New York: Wiley, 429±73.

Pierce, L. and Running, S. 1988: Rapid estimationof coniferous forest leaf area index using aportable integrating radiometer. Ecology 69,1762±67.

Pinz, A. 1998: Tree isolation and species classifi-cation. In Proceedings of the International Forum onAutomated Interpretation of High Spatial ResolutionDigital Imagery for Forestry, Victoria, BC: NaturalResources Canada, Forestry Canada, PacificForestry Centre, 10±12 February, in press.

Pitt, D., Wagner, R., Hall, R., King, D., Leckie, D.and Runesson, U. 1997: Use of remote sensingfor forest vegetation management: a problemanalysis. The Forestry Chronicle 73, 459±77.

Pollock, R. 1998: Individual tree recognitionbased on a synthetic tree crown image model.In Proceedings of the International Forum onAutomated Interpretation of High Spatial ResolutionDigital Imagery for Forestry, Victoria, BC: NaturalResources Canada, Forestry Canada, PacificForestry Centre, 10±12 February, in press.

Qi, J., Huete, A., Moran, M., Chehbouni, A.and Jackson, R. 1993: Interpretation of vegetationindices derived from multi-temporal SPOTimages. Remote Sensing of Environment 44, 89±101.

Ramstein, G. and Raffy, M. 1989: Analysis ofthe structure of radiometric remotely sensedimages. International Journal of Remote Sensing10, 1049±73.

Ranson, K., Irons, J. and Williams, D. 1994:Multispectral bidirectional reflectance of northernforest canopies with the Advanced Solid-StateArray Spectroradiometer (ASAS). Remote Sensingof Environment 47, 276±89

Ripple, W., Wang, S., Isaacson, D. and Paine, D.1991: A preliminary comparison of LandsatThematic Mapper and SPOT-1 HRV multi-spectral data for estimating coniferous forestvolume. International Journal of Remote Sensing12, 1971±77.

Robinove, C. 1981: The logic of multispectralclassification and the mapping of land. RemoteSensing of Environment 11, 231±44.

Running, S. and Hunt, E.R. Jr 1994: Generalizationof a forest ecosystem process model for otherbiomes, BIOME-BGC, and an application forglobal scale models. In Ehleringer, J. and Field,C., editors, Scaling physiological processes: leaf toglobe, Toronto: Academic Press, 141±57.

Running, S., Loveland, T. and Pierce, L. 1994: Avegetation classification logic based on remote


sensing for use in global biogeochemical models.Ambio 23, 77±81.

Running, S., Peterson, D., Spanner, M. andTeuber, K. 1986: Remote sensing of coniferousforest leaf area. Ecology 67, 273±76.

Schaaf, C., Li, X. and Strahler, A. 1994: Topo-graphic effects on bidirectional and hemisphericalreflectances calculated with a geometric-opticalcanopy model. IEEE Transactions on Geoscience andRemote Sensing 32, 1186±93.

Sharma, K. and Sarkar, A. 1998: A modifiedcontextual classification technique for remotesensing data. Photogrammetric Engineering andRemote Sensing 64, 273±80.

Shimabukuro, Y. and Smith, J. 1994: The least-squares mixing models to generate fractionimages derived from remote sensing multi-spectral data. IEEE Transactions on Geoscienceand Remote Sensing 29, 16±20.

Spanner, M., Johnson, L., Miller, J., McCreight,R., Freemantle, J., Runyon, J. and Gong, P. 1994:Remote sensing of seasonal leaf area index acrossthe Oregon transect. Ecological Applications 4,258±71.

Spanner, M., Pierce, L., Peterson, D. and Running,S. 1990: Remote sensing of temperate coniferousforest leaf area index: the influence of canopyclosure, understory vegetation and backgroundreflectance. International Journal of Remote Sensing11, 95±111.

Spurr, S. and Barnes, B. 1980: Forest ecology.Toronto: Wiley.

Stehman, S. and Czaplewski, R. 1998: Design andanalysis for thematic map accuracy assessment:fundamental principles. Remote Sensing ofEnvironment 64, 331±44.

Steven, M. 1987: Ground truth: an underview.International Journal of Remote Sensing 8, 1033±38.

St-Onge, B. and Cavayas, F. 1995: Estimating foreststand structure from high resolution imageryusing the directional variogram. InternationalJournal of Remote Sensing 16, 1999±2021.

ÐÐ 1997: Automated forest structure mappingfrom high resolution imagery based on direc-tional semivariogram estimates. Remote Sensing ofEnvironment 61, 82±95.

Strahler, A., Logan, T. and Bryant, N. 1978:Improving forest cover classification accuracyfrom Landsat by incorporating topographicinformation. In Proceedings of the 12th InternationalSymposium of Remote Sensing of the Environment,Vol. 2, Ann Arbor, MI: Environmental ResearchInstitute of Michigan, 927±42.

Strahler, A., Woodcock, C. and Smith, J. 1986:On the nature of models in remote sensing.Remote Sensing of Environment 20, 121±39.

Teillet, P. and Fedosejevs, G. 1995: On the darktarget approach to atmospheric correction ofremotely sensed data. Canadian Journal of RemoteSensing 21, 374±87.

Toman, M. and Ashton, P. 1995: Sustainable forestecosystems and management: a review article.Forest Science 42, 366±77.

Tucker, C. 1979: Red and photographic infraredlinear combinations for monitoring vegetation.Remote Sensing of Environment 8, 127±50.

Vane, G. and Goetz, A. 1993: Terrestrial imagingspectrometry: current status, future trends.Remote Sensing of Environment 44, 117±26.

van Martin, A. 1984: Forests and forestry innational life. In Hummel, F., editor, Forest policy:a contribution to resource development, The Hague:Kluwer Academic, 1±19.

van Stokkom, H. and Guzzi, R. 1984: Atmosphericspectral attenuation of airborne remote sensingdata: comparison between experimental andtheoretical approach. International Journal ofRemote Sensing 5, 925±38.

Varjo, J. 1997: Change detection and controllingforest information using multi-temporal LandsatTM imagery. Acta Forestalia Fennica 258, 9±10.

Walthall, C., Norman, J., Welles, J., Campbell, G.and Blad, B. 1985: Simple equation to approxi-mate the bidirectional reflectance from vegetativecanopies and bare soil surfaces. Applied Optics 24,383±87.

Watts, S. 1983: Forestry handbook for British Colum-bia. Cloverdale, BC: D.W. Friesen & Sons.

Westoby, J. 1989: Introduction to world forestry.Oxford: Blackwell.

Woodcock, C., Collins, J., Gopal, S., Jakabhazy,V., Li, X., Macomber, S., Ryherd, S., Wu, Y.,Harward, V., Levitan, J. and Warbington, R.1994: Mapping forest vegetation using LandsatTM imagery and a canopy reflectance model.Remote Sensing of Environment 50, 240±54.

Woodcock, C., Collins, J., Jakabhazy, V., Li, X.,Macomber, S. and Wu, Y. 1997: Inversion of theLi-Strahler canopy reflectance model for map-ping forest structure. IEEE Transactions on Geo-science and Remote Sensing 35, 405±14.

Woodcock, C. and Strahler, A. 1987: The factor ofscale in remote sensing. Remote Sensing ofEnvironment 21, 311±22.

Woodcock, C., Strahler, A. and Jupp, D. 1988a:The use of variograms in remote sensing. I. Scenemodels and simulated images. Remote Sensing ofEnvironment 25, 323±48.

ÐÐ 1988b: The use of variograms in remotesensing. II. Real digital images. Remote Sensingof Environment 25, 323±48.

Mike Wulder 475

Wu, Y. and Strahler, A. 1994: Remote estimation ofcrown size, stand density, and biomass on theOregon transect. Ecological Applications 4,299±312.

Wulder, M. 1998a: The prediction of leaf area indexfrom forest polygons decomposed through theintegration of remote sensing, GIS, UNIX, andC. Computers and Geoscience 24, 151±57.

ÐÐ 1998b: Image spectral and spatial informa-tion in the assessment of forest structuraland biophysical data. In Proceedings of the Inter-national Forum on Automated Interpretation of HighSpatial Resolution Digital Imagery for Forestry,Victoria, BC: Natural Resources Canada, ForestryCanada, Pacific Forestry Centre, 10±12 February,in press.

Wulder, M. and Boots, B. 1998: Local spatialautocorrelation characteristics of remotelysensed imagery assessed with the Getis statistic.International Journal of Remote Sensing, 19,2223±31.

Wulder, M., Lavigne, M. and Franklin, S. 1996b:High spatial resolution optical image texture forimproved estimation of forest stand leaf areaindex. Canadian Journal of Remote Sensing 22,441±49.

Wulder, M., LeDrew, E., Franklin, S. and Lavigne,M. 1998: Aerial image texture information in theestimation of northern deciduous and mixedwood forest leaf area index (LAI). Remote Sensingof Environment 64, 64±76.

Wulder, M., Mah, S. and Trudeau, D. 1996a:Mission planning for operational data acquisitioncampaigns with the casi. In Proceedings, SecondInternational Airborne Remote Sensing Conferenceand Exhibition, Vol. III, San Francisco, CA,24±27 June, 53±62.

Xia, L. 1996: A method to improve classificationwith shape information. International Journal ofRemote Sensing 17, 1473±81.


optical remote-sensing techniques for the assessment of...

Documents