425

Potential of Woody Carbon Stock Estimation Using HighSpatial Resolution Imagery: A Case Study of Spruce Stands

Yoshio AWAYA1, Satoshi TSUYUKI2, Eiji KODANI3 and Gen TAKAO4

1Forestry and Forest Products Research Institute,1, Matsunosato, Tsukuba, Ibaraki 305-8687, Japan

2Graduate School of Agricultural and Life Sciences, The University of Tokyo,Tokyo 113-8657, Japan

3Shikoku Research Center, Forestry and Forest Products Research Institute,Kochi 780-8064, Japan

4Hokkaido Research Center, Forestry and Forest Products Research Institute,Sapporo 062-0045, Japan

Abstract. The performance of visible and near-infrared radiometers wasexamined for forest stand parameter estimation using airborne images. Theimages were obtained in September 1993, October 1994 and June 1995 with 4-meter spatial resolution using the Compact Airborne Spectrographic Imager(CASI), which had a similar specification as the advanced spaceborne sensors,such as IKONOS. The relationships between tree height, stem volume anddigital number (DN) were analyzed for evergreen coniferous stands usingstatistical methods. The minimum DN of stands, which showed the darkness ofa canopy shadow, had a higher correlation with top layer height (H1) than theaverage and maximum DN. While the average DN had a higher correlation withstem volume per hectare (Vha), which composed most of the carbon stock inforests. There was a linear relationship between the measured and the estimatedH1. On the other hand, the estimates were saturated at higher Vha than 250 m3

per hectare, as has been pointed out in the literature. It is necessary to overcomethe saturation problem to give better estimation of carbon stocks in forests.Visible channels between 537 and 687 nm showed high correlation coefficientswith H1 and Vha for the three images, and the highest correlation appeared forthe case of H1 using the September image. Late summer would be the bestperiod for stand parameter estimation. The specification of high resolutionsatellite sensors seemed to be useful for stand parameter estimation. If seasonand observation angle of images are selected properly, stable estimation of H1and Vha can be achievable.

Keywords: CASI, conifer, remote sensing, stem volume, tree height

1. INTRODUCTION

Standing volume is not only one of the most important parameters in forestmanagement, but also one of the major political interests, after the 3rd Session ofthe Conference of the Parties to the United Nations Framework Convention onClimate Change (COP3) in 1997. If stem volume and other stand parameters are

Global Environmental Change in the Ocean and on Land, Eds., M. Shiyomi et al., pp. 425–440.© by TERRAPUB, 2004.

426 Y. AWAYA et al.

estimated in large areas with sufficient accuracy, they can be used for forestmanagement to replace or supplement the current inventory system. Such spatialinformation may also be important in political decision making about carbonemission controls. Remotely sensed data have potential for accurate biomassestimation.

The Thematic Mapper (TM) of Landsat has been the most popular spacebornesensor during the last decade. The TM records the brightness of ground objectsin 256 grey levels. Although the TM digitizes objects in 256 grey levels, the dataranges of dense evergreen conifer stands are usually less than 10 in digital number(DN) in visible channels. Therefore, it is hard to identify small differences inreflectance caused by canopy structures using sensors with a narrow dynamicrange, such as TM and other traditional spaceborne sensors. On the other hand,with advances in technology, airborne optical sensors have been improveddrastically during the last two decades. Hyper-spectral channels with narrowerspectral ranges have been achieved, and wider dynamic ranges and better signal-to-noise ratio make it possible to identify slight radiation differences in targets.The Compact Airborne Spectrographic Imager (CASI) (Babey and Anger, 1989)is one of these airborne sensors. The observation parameters of the CASI can beprogrammed by the operator and it observes in the visible and near infrared (NIR)regions with a minimum of 1.8-nm spectral resolution. The images used in thisstudy have 15 or 16 channels with 30-nm spectral resolution, 4-meter spatialresolution and a 12-bit dynamic range, namely 4096 gray levels. Thesespecifications would not be the limiting factor for forest analysis.

The advanced technology is being applied to spaceborne sensors such asIKONOS or QuickBird (Fritz, 1996). The ground resolution of these sensors isabout 4 or 2.5 meters in multispectral mode. The sensors digitize objects in 2,048grey levels. Thus, there is a great possibility to improve the accuracy of standparameter estimation in large areas without the limitation of sensor performance.

With agricultural crops there is a curvilinear relationship between the leafarea index (LAI) and radiance or reflectance factors (Holben et al., 1980; Curran,1983; Huete, 1987). As in crops, the relationship between LAI of forests andsatellite sensor signals, namely radiance, has been studied previously (Petersonet al., 1987; Curran et al., 1990; Spanner et al., 1990; Baulies and Pons, 1995;Chen and Cihlar, 1996). On the other hand, some models show that standstructures such as canopy coverage, canopy size, and canopy shadow haveinfluences on reflectance or radiance (Strahler et al., 1988; Nilson and Peterson,1994).

Thus, it is reasonable to study stand parameters such as diameter at breastheight (DBH), tree height, and stem volume using remote sensing data. Actually,several studies have estimated these parameters using high spatial resolutiondata. The stand parameters are estimated well by using radiance data for theboreal and temperate coniferous forests (Franklin, 1986; Cohen and Spies, 1992;Franklin and McDermid, 1993; Baulies and Pons, 1995; Cohen et al., 1995;Trotter et al., 1997, Awaya et al., 2000). Seasonal variations of vegetation spectra

Potential of Woody Carbon Stock Estimation 427

(Blackburn and Milton, 1995; Awaya and Tanaka, 1999; Kodani et al., 2002) anddifferences of spectra among trees probably reduce the accuracy of stand parameterestimation.

The relationships between LAI and DN in the NIR and the intensity of avegetation index are also changed significantly by canopy closure, and affectedby the undergrowth in open canopy forests (Spanner et al., 1990). When thecanopy closure is small, the radiance changes greatly with the change of theclosure due to the influence of forest floor (Awaya et al., 1992). As a result, thespectra of forest floors interfere with the biomass analysis when the canopy is notclosed sufficiently. Since this paper focuses on the evaluation of relationshipsbetween DN and the stand parameters and their seasonal differences, it wasnecessary to avoid the influences of topography and the forest floor. Therefore,closed forests on flat terrain were selected to examine relationships that are freefrom the effects of topography and forest floor.

This study evaluates the performance of advanced visible and near-infraredradiometers in stand parameter estimation, and discusses (1) the effects of canopylayers on CASI signals, (2) the better stand parameters to be estimated usingCASI images, (3) the best wavelength channel for estimation of the standparameters, and (4) the stability of parameter estimation among results using thethree images in three seasons, with an accuracy assessment.

2. STUDY SITE AND DATA

2.1 Study site

The study site was situated in the national forest near Tomakomai, Hokkaido,Japan (Fig. 1, 42°45′ N, 141°32′ E). It has a very flat topography with an elevationof approximately 100 meters above sea level, and is located in the transition zonebetween the cool temperate forest and boreal forest. Most of the forests in thisarea are artificial plantations of Ezo spruce (Picea Jezoenthis Carr.), Akaezospruce (Picea glehnii Mast.), Todo fir (Abies sachalinensis var. mayrianaMiyabe et Kudo), and Japanese larch (Larix kaempferi Carr.). Since two sprucespecies have the longest plantation history, spruce-dominated stands were mainlyselected for the study.

2.2 Field surveys

Field surveys were conducted four times (July 19 to 23, 1993, September 9to 13, 1994, June 6 to 10, 1995, and July 22 to 26, 1997). The plot area wasbetween 0.01 and 0.09 ha and was chosen according to tree height. The DBH wasmeasured for all trees greater than or equal to 4 cm, and tree height was measuredfor three trees in the top layer, one tree in the middle layer and one tree in the lowerlayer in each plot. The plot location was measured using a compass and tapemeasure from any clear point on maps or aerial photos. Only the plots with clearlocation documentation were used in the analysis.

428 Y. AWAYA et al.

2.3 Airborne data

A Canadian airborne sensor, CASI of Iters Research Ltd., was used to collectremote sensing data. The CASI was able to measure spectral radiation from 420to 920 nm. The spectral positions of the channels could be set freely to 18channels for the spectral mode on condition that these channels did not overlap.Several flights were executed by the National Space Development Agency ofJapan (NASDA) in three different seasons: September 1993 (CASI9309), October1994 (CASI9410) and June 1995 (CASI9506) to study seasonal differences. Thespecifications of CASI and the images that were used in the analysis are shownin Table 1. Part of the images, covering 8 km by 2 km approximately, was usedin this study. Geometric skews of the images, which were caused by unstableaircraft positions, were corrected as much as possible by the system software ofIters Research Ltd. Although CASI data were converted to radiance values, theterm DN is still used in this paper due to uncertainty about the sensor calibration.DN showed a linear relationship with radiance values, and hence showed a linearrelationship with reflectance factors on flat terrain for all channels.

2.4 Reference data

Black and white aerial photos at scales 1:20,000 and 1:24,000, which weretaken in 1990 and 1991 by the Forestry Agency and the Geographical SurveyInstitute of Japan, respectively, were used to check stand conditions.

Color aerial photos of scale 1:20,000, which were taken on August 2, 1993by NASDA using a Hasselblad camera, were also used. Plot locations weremarked on the color aerial photos according to the forest planning maps of scale1: 20,000 of Tomakomai and Eniwa forestry offices and the location documents.

Fig. 1. Location map of the study site.

Potential of Woody Carbon Stock Estimation 429

3. METHOD

3.1 Plot data compilation

Two tree height-diameter curves were utilized to estimate tree height, onefor conifer trees (Eq. (1), Awaya et al., 2000) and the other for broad-leaved trees(Eq. (2), Nakashima, 1946).

H = DBH2/(2.709 + 0.1621·DBH)2 + 1.3 (1)

H 0.67474 DBH 0.74271 +1.3 21.44326 DBH= × × ( )

where DBH was the diameter at breast height and H was the tree height.The tree height was calculated using the curves for each tree in all plots. Stem

volume was estimated for each tree using the measured DBH, the calculatedheight and stem volume equations for spruce, fir and broad-leaves (HokkaidoForestry Bureau, 1970). Four different groups of trees were considered in eachplot. The average height of the highest 10% of the trees in each plot (Htop) wascalculated. Groups T1, T2 and T3 consisted of trees whose heights were at leastthree fourths, one-half and one-fourth of Htop, respectively, and group Tallconsisted of all trees.

Average tree height and stem volume per hectare were computed for the fourgroups in each plot. They were assigned symbols H1 (top layer height), H2, H3and Havr (average tree height) for tree height, and V1, V2, V3 and Vha (total stemvolume per hectare) for stem volume in T1, T2, T3 and Tall, respectively.

CASI9309 CASI9410 CASI9506

Observation date Sep. 2, 1993 Oct. 3, 1994 Jun. 12, 1995

Observation time 10:35 10:00 10:08

Nominal flight height 3000m

Flight direction 325� 325� 330�

Solar azimuthal angle 153.7� 150.6� 131.3�

Solar zenith angle 37.1� 50.5� 26.5�

Field of view 35� 35� 42�

Nominal spatial resolution about 3.7 m x 3.7 m

Pixel spacing cross truck 3.6 m 3.6 m 4.4 m

along truck 4.3 m 3.4 m 3.9 m

Quantization levels 12bit (4096 levels)

Spectral resolution about 28nm

Observation channels*

(nm)

446, 476, 507, 537, 567, 596, 626, 656,

687, 716, 746, 776, 806, 837, 867, 902**

Table 1. CASI data specification.

*The center wavelength in 1995. The center positions were different by about a few nanometersin 1993 and 1994.

**Not observed in 1993.

430 Y. AWAYA et al.

Relatively homogeneous stands, in which more than 70% of the trees were ingroup T2, were selected for the analysis.

3.2 Selection of plot areas in the images

Plot locations were identified on all images by comparison with each otherand by reference to plot locations on the color aerial photos. Fifty one trainingareas of the plots were carefully selected to choose the same areas in all images.The size of these training areas was large enough to cover the plots. The average(DNavr), minimum (DNmin) and maximum DN (DNmax) of each channel werecomputed for each plot. Because of cloud cover and because the images covereddifferent areas, the three images contained the same 17 plots (common plots),which were measured in 1993, 1994 and 1995. CASI9309, CASI9410 andCASI9506 contained other 16, 16 and 20 plots of the 51 plots, respectively. Thecommon 17 plots were used for the correlation and regression analyses, and theother plots were used for accuracy validation (validation plots). Annual changesin forests, such as growth and thinning, were neglected in the analysis.

3.3 Analysis

Relationships between DN parameters (DNavr, DNmin, DNmax) and thestand parameters of the 17 plots were analyzed as follows.

The correlation coefficients between the DN and the stand parameters werecalculated for a visible channel at 567 nm and an NIR channel at 806 nm. Basedon the results, the most influential stand parameter on DN was analyzed. The mosteffective DN parameter showing stand parameters or stand structures was alsoevaluated. The best channel (wavelength) and image (season) were evaluated bycorrelograms, which showed correlation coefficients between H1 and DNmin orVha and DNavr.

Linear regression models were developed for H1 and DNmin, and for Vhaand DNavr for a visible and a NIR channels according to the results of correlationcoefficient evaluations. Accuracy of the estimates obtained from the regressionmodels was judged using the validation plots. The spatial distributions of theparameters were estimated using the derived models and the CASI images. Theimages were geometrically rectified with 3.5 meter resolution using groundcontrol points and second-order polynomials. A minimum and a low pass filterwith circular moving windows with a diameter of 7 pixels were applied to therectified images. The minimum filter gives local DNmin and the low pass filtergives local DNavr. The minimum filtered image and low pass filtered image wereused for H1 and Vha estimation, respectively.

The spatial patterns of H1 and Vha were evaluated as follows. The evergreenconifer areas were extracted visually for prints by comparing the images with theforest planning maps. The estimation results were compared with each othervisually. Sixteen areas, which were located near the nadir in all images, wereselected with tree size differences being considered (comparison areas). Theareas each consisted of approximately 150 pixels. The average of H1 and Vha

Potential of Woody Carbon Stock Estimation 431

were calculated for all areas. Relationships among the three seasons wereanalyzed by scattergrams and linear regression analysis.

4. RESULTS AND DISCUSSION

4.1 Correlation coefficients

A negative correlation was observed in all combinations of DN and the standparameters (Table 2). Correlation coefficients were significant (P > 0.01) exceptfor the 806 nm channel in 1995. Negative relationships are common between thestand parameters (LAI, DBH, tree height, woody volume and so on) and DN (orradiance, or reflectance factor) in visible and NIR channels (Franklin, 1986;Peterson et al., 1987; Spanner et al., 1990; Franklin and McDermid, 1993;Baulies and Pons, 1995). We discuss the results, except for the 806 nm in 1995case, in this sub-section.

4.1.1 Effects of layersCorrelation coefficients of four layers for height or volume were not

significantly different each other (P > 0.05) within the same statistical items(average, maximum or minimum) of the same channel in the same image. EitherH1 or H2 mostly showed the greatest correlation among the four layers. On theother hand, unlike tree height, Vha showed the greatest correlation coefficient inmost cases with stand volume. Although there is no clear reason for the difference,the tree height cases probably suggest that the upper layer canopy affected mainlyDN. Since older or higher plantations had more under-story trees, total volume(Vha) may show a difference of canopy structure more than the case of only top

Table 2. Correlation coefficients between height, volume and DN of CASI.

The correlation coefficients are significant at the 0.05 level (*) and at the 0.01 level (**).

Vari-

able

Wave-

lengthAverage

Maxi-

mum

Mini-

mumAverage

Maxi-

mum

Mini-

mumAverage

Maxi-

mum

Mini-

mum

H1 567nm -0.892**

-0.769**

-0.917**

-0.883**

-0.880**

-0.915**

-0.761**

-0.634**

-0.789**

H2 567nm -0.900**

-0.770**

-0.914**

-0.881**

-0.886**

-0.891**

-0.767**

-0.655**

-0.763**

H3 567nm -0.884**

-0.755**

-0.889**

-0.866**

-0.869**

-0.880**

-0.734**

-0.656**

-0.759**

Havr 567nm -0.881**

-0.751**

-0.886**

-0.865**

-0.864**

-0.885**

-0.746**

-0.658**

-0.757**

H1 806nm -0.705**

-0.711**

-0.810**

-0.815**

-0.835**

-0.788**

-0.169 -0.121 -0.424

H2 806nm -0.750**

-0.752**

-0.823**

-0.823**

-0.864**

-0.747**

-0.205 -0.158 -0.424

H3 806nm -0.741**

-0.750**

-0.799**

-0.815**

-0.861**

-0.732**

-0.205 -0.165 -0.434

Havr 806nm -0.739**

-0.748**

-0.799**

-0.812**

-0.848**

-0.746**

-0.201 -0.169 -0.439

V1 567nm -0.801**

-0.576*

-0.780**

-0.801**

-0.848**

-0.787**

-0.695**

-0.715**

-0.670**

V2 567nm -0.870**

-0.686**

-0.853**

-0.860**

-0.892**

-0.804**

-0.801**

-0.749**

-0.721**

V3 567nm -0.886**

-0.737**

-0.847**

-0.888**

-0.920**

-0.819**

-0.823**

-0.788**

-0.755**

Vha 567nm -0.894**

-0.756**

-0.860**

-0.903**

-0.923**

-0.841**

-0.836**

-0.708**

-0.793**

V1 806nm -0.689**

-0.707**

-0.724**

-0.752**

-0.820**

-0.689**

-0.215 -0.271 -0.383

V2 806nm -0.789**

-0.773**

-0.824**

-0.832**

-0.874**

-0.701**

-0.326 -0.311 -0.442

V3 806nm -0.798**

-0.802**

-0.834**

-0.858**

-0.885**

-0.728**

-0.347 -0.348 -0.492*

Vha 806nm -0.780**

-0.781**

-0.844**

-0.864**

-0.864**

-0.768**

-0.336 -0.325 -0.517*

CASI DN in 1995CASI DN in 1993 CASI DN in 1994

432 Y. AWAYA et al.

layer trees (V1). H1 was related to the canopy size, and hence to the surfaceroughness of the stands, while Vha was the total woody biomass, or the sum ofthe three-dimensional biomass concentration in the stands.

4.1.2 Difference in minimum, average and maximum DNDNmin showed the greatest correlation coefficients for tree height, except

in the case of the 806 nm channel in 1994. This probably showed the influence ofcanopy shadows on DN. However, there was not any clear trend for the cases ofvolume. DNavr showed either greatest or second greatest correlations, except inthe case of 806 nm channel in 1993. It may be better using the average DN toestimate stem volume with stable accuracy than using the DNmax or DNmin.

DNmin contained larger parts of canopy shadows than DNmax and DNavr.Pixels of DNmin must consist of conifer canopies and shadows. Therefore, theshadow area within the pixel was changed according to the upper tree size, andcaused the linear relationship in 567 nm for H1. If the ground resolution of CASIdata was finer, there would be an asymptotic value, which would be DN of pixelscontaining only shadows of canopies. This suggests that there are differentappropriate ground resolutions for different forest types for tree height estimation.

The higher reflectance factors of broadleaved trees than those of evergreenconifers caused errors in the stand parameter estimation in mixed stand condition.However, the minimum filtering would select only dark pixels of either conifertrees or shadows. Thus errors due to deciduous trees can be reduced by a filtering.

4.1.3 Difference in seasons and channelsThe visible channels gave higher correlation coefficients than the NIR

channels (Table 2, Fig. 2). The highest correlation was observed in the greenchannels at 567 and 596 nm in CASI9309 and CASI9410, although in the red

Fig. 2. These correlograms show correlation coefficients between stand and DN parameters indifferent wavelength. (a) shows the correlation between HI and DNmin, and (b) shows thecorrelation between Vha and DNavr. Correlation coefficients were higher in visible channelsbetween 537 and 687 run than the those in NIR channels between 776 and 902 rim channels.Correlation coefficients between –0.6 and –1.0 are significant at 0.01 level.

Potential of Woody Carbon Stock Estimation 433

channels, it was observed at 656 and 687 nm in CASI9506. Similar to Fig. 2, thered and green channels give high correlations with the stand parameters (Petersonet al., 1987; Spanner et al., 1990; Franklin and McDermid, 1993, Franklin, 1986).On the other hand, NIR channels give high correlations in some cases (Bauliesand Pons, 1995; Trotter et al., 1997) and poor ones in others (Franklin, 1986;Peterson et al., 1987; Spanner et al., 1990). Figure 2 showed similar results tocases in the literature, namely, high correlation coefficients were observed in thered and green channels steadily, but poor or high correlation coefficients appearedin the NIR channels. The difference in the correlation coefficients in these imagesprobably shows effects of seasonal spectral changes of tree species.

Although there were some differences in correlation coefficients betweenchannels, the spectral regions between 600 and 690 nm showed similar correlations.Therefore, a channel with wider spectral rage than CASI channels may givealmost the same accuracy in stand parameter estimations with single CASIchannel.

Seasonal spectral changes of tree species in the study area were examinedusing Landsat TM images after adjustment for intensity of DN and correction ofatmospheric effects (Fig. 3, Awaya and Tanaka, 1999). Reflectance factors in theNIR (TM4) increase greatly between late April and late May due to leafdevelopment in spring, then reached a peak in early summer for both spruce andbroadleaved forests. On the other hand, reflectance in the green channel (TM2)of deciduous forest increased in mid-April and steadily decreased until earlysummer due to leaf development and maturation. Due to lack of TM images inSeptember, spectral changes in this period are not clear, but reflectance in thesechannels is probably stable before autumnal coloring in October, according to a

Fig. 3. Spruce and broad-leaved forests showed significant seasonal changes in their spectralreflectance. The reflectance factors were estimated using TM images between 1985 and 1993. TM2and TM4 are the green and the NIR channel, respectively.

434 Y. AWAYA et al.

ground measurement (Kodani et al., 2002).The seasonal changes of reflectance factors suggest that reflectance factors

of spruce trees were probably not homogeneous in places in the study site inCASI9506. Because the period between May and early June was the leafdevelopment season, trees had new leaves at the time of image acquisition. Suchdiversified stand condition was probably the cause of difference in correlationcoefficients of CASI9506 from those of other images. Therefore, the leafdevelopment period is probably not suitable for estimating stand parameters. Onthe other hand, reflectance considerably changes season-by-season in the NIR,but hardly in the blue and red regions (Howard, 1991), which indicates that treeshad more similar reflectance to each other in the blue and red channels than in theother channels in CASI9506. This may be a reason for higher correlation betweenDN and the stand parameters in the green than that in the red channel in the caseof CASI9506.

Since absorption of tree leaves is strong in the visible region, light scarcelypenetrates the leaf layers (Awaya et al., 1999). Therefore, DN mainly providedthe information of the canopy surfaces in the visible region, and the surfaceinformation appeared in H1. Meanwhile, high transmittance allowed light topenetrate the leaf layers in the NIR region (Awaya et al., 1999), and DN wasinfluenced by trees not only in the upper layers but also in the understory.Variations of reflectance factors among trees and among species (Awaya andTanaka, 1999) appeared to be greater in the NIR than in the visible regions, thusaffecting the relationship between the stand parameters and DN in the NIRregion. This can be explained as the cause of the differences in H1 and Vha in therelationship at 567 nm and 806 nm in Table 2.

4.2 Estimation of stand parameters

According to the results in the previous sections, stand parameters and DNvariables were determined to estimate spatial distributions. Linear models weredeveloped for the green channel (567 nm) in CASI9309 and CASI9410 and forthe red channel (656 nm) in CASI9506 for H1 and Vha with the 17-plot data(Table 3). H1 and Vha were estimated using those models and CASI images.

Parameter Image Wavelength Slope Constant R2

H1 CASI9309 567nm -0.137 34.2 0.841**

H1 CASI9410 567nm -0.0287 38.6 0.837**

H1 CASI9506 656nm -0.0205 32.7 0.802**

Vha CASI9309 567nm -0.271 674 0.800**

Vha CASI9410 567nm -0.545 761 0.816**

Vha CASI9506 656nm -0.421 675 0.771**

Table 3. Simple regression equations for H1 and Vha estimation.

H1 (m) or Vha (m3/ha) = Slope × DN + Constant.The coefficients of determination are significant at 0.01 level (**).

Potential of Woody Carbon Stock Estimation 435

Since the study site was of an ideal condition with a flat terrain and quitehomogeneous evergreen conifer stands, the coefficients of determination ofequations, which were use for H1 and Vha estimation, were one of the highestamong some previous studies of this field. The most accurate estimation ispossible for homogenous evergreen conifers with the coefficients of determinationof 0.89 by a log-linear model (Peterson et al., 1987), 0.51 by a log-linear model(Spanner et al., 1990) for LAI, and 0.56 by a linear model for DBH (Flanklin andMcDermid, 1993). The coefficient of determination is lower in complex speciescomposition, 0.29 by a log-linear model for the basal area (Franklin, 1986). Ourresults probably show the potential of high spatial resolution imagery for standparameter estimation.

4.2.1 ValidationVha was underestimated in higher volumes, but H1 corresponded quite well

with the measured value in the range of validation plots (Fig. 4). The differencemay be related to the characteristics of H1 and Vha, the surface roughness and thethree-dimensional biomass concentration, respectively. H1 was a better parameterthan Vha for achieving finer accuracy in our results.

A numerical simulation showed a curvilinear relation between timber volumeand DN in the red and NIR channels, when volume exceeded 300 m3 per ha(Gemmell, 1995). Vha was saturated at about 230 m3 per hectare for the case ofCASI9410 (Fig. 4) and the saturation was lower than the model prediction. Stemvolume about 250 m3 seemed be a limitation to estimate volume accurately usingvisible near infrared radiometer data alone for the case of evergreen conifers.

4.2.2 Spatial patternAs regards the spatial distribution, both H1 and Vha were obviously

Fig. 4. Scattergrams between the measured and estimated stand parameters using CAS19410. a) Therelationship for the case of top layer height (H1), and b) the relationship for the case of stem volumeper ha (Vha). H1 has a near one-to-one relationship between the measured and estimated H1,however, Vha is probably saturated over 280 m3. The vertical lines show the confidence intervals atthe 0.95 level. The slope coefficients of the two equations are significant at 0.01 level.

436 Y. AWAYA et al.

overestimated in the north-eastern edge of CASI9506 (Fig. 5). The sun’s rayswere almost parallel to the flight directions of CASI9309 and CASI9410, and theillumination condition was almost same within each image. On the other hand, thesun shone about 20° from the east and behind the airplane in the case ofCASI9506. Being the sunny side (backscattering side) of the canopy, DN in theright half (Fig. 5) was greater than on the shade side (forward-scattering side) in

Fig. 5. Spatial distribution of the estimated parameters. left: CASI images (Blue: ch7, Green: ch15,Red: ch3), middle: H1, right: Vha, upper: September 1993, middle: October 1994, and bottom: June1995. H1 and Vha were overestimated at the right half of the June image. Spatial patterns are similaramong three results in the left half of the June image and others. The white lines are the flight centers.Evergreen stands were delineated visually. Numbers in the upper middle are stand age in 1993.

Potential of Woody Carbon Stock Estimation 437

Fig. 6. Relationship between the estimated stand parameters. a) The case of top layer height (H1)and b) the case of stem volume per hectare (Vha) between CASI9309, CASI9410 and CASI9506. Theslope coefficients of the four equations were significant (P > 0.01).

the left half. Due to such directional reflectance properties (Kleman, 1987;Ranson et al., 1994), H1 and Vha were estimated higher in the right half ofCASI9506.

For this reason, narrower viewing sensors are desirable for stand parameterestimation. Sensors with an oblique viewing capability, such as IKONOS andQuickBird, will cause a serious problem in stand parameter estimation. It will bedifficult to obtain stable, accurate estimates using these satellite images.

Spatial patterns except those on the north-eastern side (right side in Fig. 5)appeared similar in each estimation. Comparison areas were selected near thenadir of images to avoid the effects of directional reflectance properties. Therewas a linear relationship and the correlation coefficients were high between theresults in the comparison areas, but CASI9506 showed smaller correlationcoefficients with CASI9410 than CASI9309 (Fig. 6). As shown in Fig. 6, theresults of CASI9506 are over estimations to the results of CASI9410. HoweverFig. 6 suggests that if season and viewing geometry of images are chosenproperly, tree height and volume estimations will become more stable as seen inthe cases between CASI9410 and CASI9309.

4.3 Best season for stand parameter estimation

As discussed above, leaf maturation affected the spectra in the early growingseason. The reflectance factors exceeded the maximum in early July and droppedremarkably in early August in NIR in Tomakomai (Fig. 3, Awaya and Tanaka,1999), and then the effects of leaf maturation decreased after July. The correlogramsshowed higher correlation coefficients in all channels in the late growing seasonthan early growing season except for the two red channels (Fig. 2). Therefore, the

438 Y. AWAYA et al.

late growing season before autumnal coloring will be the best time for obtainingimages.

5. CONCLUSIONS

The following points became clear from the results of this study, whichevaluated stand parameters of spruce using three CASI images:

(1) According to the results, local DNmin showed the strongest correlationwith H1. Projected shadow with larger trees was probably darker than smallertrees. Therefore, DNmin was useful for H1 estimation. Finer spatial resolutionthan traditional satellite sensors such as TM is probably effective in tree heightestimation. There must be a suitable spatial resolution for different canopy sizes.

(2) A visible spectral region between 550 and 690 nm showed the largestcorrelation coefficients between DNmin and H1 or between DNavr and Vha.Narrower spectral resolution is probably favorable. However, a wider spectralchannel between 600 and 690 nm probably gives similar results as narrowerchannels in this spectral region.

(3) Utilization of NIR channels will be affected very much by phenologicalspectral changes of trees. A later growing season must be selected for standparameter estimation, if NIR channels are used. Even regarding the green and redchannels, the late growing season seemed to be the best season.

(4) The bidirectional reflectance characteristics of conifer canopies resultedin significant errors in stand parameter estimation at off-nadir angles.

These results suggest that the specifications of current high spatial resolutionsensors are appropriate for stand parameter estimation in terms of the spatialresolution and channel setting. However, off-nadir observations will causeserious problems. Future studies will be necessary on the effects of topography,the usefulness of multi-off-nadir observations and so on. Stem volume can beestimated based on canopy diameter size, which is observed using very highresolution images. A comparative study of radiance-based and diameter-basedmethods would be necessary to understand effects of canopy size on spectra andto validate the usefulness of radiance-based analysis.

Acknowledgements—The authors would like to express their heartfelt gratitude to JapanAerospace Exploration Agency former NASDA for its support of the airborne and fieldsurveys. We are also thankful to Dr. Mitsuo Inose (Forestry and Forest Products ResearchInstitute, FFPRI), Dr. Norihiko Shiraishi (The Univ. of Tokyo), Dr. Satoshi Ishibashi(FFPRI), Dr. Makoto Sano (FFPRI), Mr. Nobuhiko Tanaka (FFPRI), Mr. KunihiroTanaka (FFPRI), Dr. Nobuyuki Abe (Niigata Univ.), Dr. Masato Katoh (Shinshu Univ.),Mr. Anssi Lohi (previously Technical Research Center of Finland) and many other peoplewho have helped in the forest surveys.

REFERENCES

Awaya, Y. and Tanaka, N. (1999): Seasonal spectral changes in cool temperate forests: An Analysisusing Landsat TM data. Japan Journal of Photogrammetry and Remote Sensing, 38, 35–46 (inJapanese with English summary).

Awaya, Y., Miller, J. R. and Freematle, J. R. (1992): Background effects on reflectance and

Potential of Woody Carbon Stock Estimation 439

derivatives in an open-canopy forest using airborne imaging spectrometer data. InternationalArchives of Photogrammetry and Remote Sensing, ISPRS, Washington, D.C., XXIX B7, 836–843.

Awaya, Y., Tanaka, K., Sawada, H. and Takao, G. (1999): Seasonal changes of transmittance spectrain a beech stand—Measurement method and outline of seasonal changes—. Tohoku Journal ofForest Science, 4, 1–7 (in Japanese with English summary).

Awaya, Y., Tanaka, N., Tanaka, K., Takao, G., Kodani, E. and Tsuyuki, S. (2000): Stand parameterestimation of artificial evergreen conifere forests using airborne images: An evaluation ofseasonal difference on accuracy and best wavelength. Journal of Forest Research, 5, 247–258.

Babey, S. K. and Anger, C. D. (1989): A Compact Airborne Spectrographic Imager (CASI).Proceedings of International Geoscience and Remote Sensing Symposium, IGARSS ’89,Vancouver, Canada, 12, 1028–1031.

Baulies, X. and Pons, X. (1995): Approach to forestry inventory and mapping by means of multi-spectral airborne data. Int. J. Remote Sensing, 16, 61–80.

Blackburn, G. A. and Milton, E. J. (1995): Seasonal variations in the spectral reflectance ofdeciduous tree canopies. Int. J. Remote Sensing, 16, 709–720.

Chen, J. M. and Cihlar, J. (1996): Retrieving leaf area index of boreal conifer forests using LandsatTM images. Remote Sens. Environ., 55, 153–162.

Cohen, W. B. and Spies, T. A. (1992): Estimating structural attributes of Douglas-fir/westernhemlock forest stands from Landsat and SPOT imagery. Remote Sens. Environ., 41, 1–17.

Cohen, W. B., Spies, T. A. and Fiorella, M. (1995): Estimating the age and structure of forests in amulti-ownership landscape of western Oregon. Int. J. Remote Sensing, 16, 721–746.

Curran, P. J. (1983): Multispectral remote sensing for the estimation of green leaf are index. Phil.Trans. R. Soc. Lond. A, 309, 257–270.

Curran, P. J., Dungan, J. L. and Gholz, H. L. (1990): Seasonal LAI in Slash pine estimated withLandsat TM. NASA TM-102278, 19 pp., National Aeronautics and Space Administration, U.S.A.

Franklin, J. (1986): Thematic Mapper analysis of coniferous forest structure and composition. Int.J. Remote Sensing, 7, 1287–1301.

Franklin, S. E. and McDermid, G. J. (1993): Empirical relations between digital SPOT HRV andCASI spectral response and lodgepole pine (Pinus contorta) forest stand parameters. Int. J.Remote Sensing, 14, 2331–2384.

Fritz, L. W. (1996): The era of commercial earth observation satellites. Photogrammetric Engineeringand Remote Sensing, 62, 39–45.

Gemmell, F. M. (1995): Effects of forest cover, terrain, and scale on timber volume estimation withThematic Mapper data in a Rocky Mountain site. Remote Sens. Environ., 51, 291–305.

Hokkaido Forestry Bureau (1970): Standing Tree Volume Table. 50 pp., Rinya Kohsaikai, Sapporo(in Japanese) (the English titles are tentative translations from the original Japanese titles by theauthors of this paper).

Holben, B. N., Tucker, C. J. and Fan, C. (1980): Spectral assessment of soybean leaf area and leafbiomass. Photogramm. Eng. Remote Sens., 46, 651–656.

Howard, J. A. (1991): Remote Sensing of Forest Resources. 420 pp., Chapman and Hall, London.Huete, A. R. (1987): Soil influences in remotely sensed vegetation-canopy spectra. In Theory and

Applications of Optical Remote Sensing. Asrar, G. (ed.), 734 pp., John Willey & Sons, NewYork, 107–141.

Kleman, J. (1987): Directional reflectance factor distributions for two forest canopies. Remote Sens.Environ., 23, 83–96.

Kodani, E., Awaya, Y., Tanaka, K. and Matsumura, N. (2002): Seasonal patterns of canopy structure,biochemistry and spectral reflectance in a broad-leaved deciduous Fagus crenata canopy.Forest Ecology and Management, 167, 233–249.

Nakashima, H. (1946): Standing Volume Table for Some Forest Types and Areas of Hokkaido.Hokkaido Branch of Kohrinkai, 4 (in Japanese) (the English titles are tentative translations fromthe original Japanese titles by the authors of this paper).

Nilson, T. and Peterson, U. (1994): Age dependence of forest reflectance: analysis of main drivingfactors. Remote Sens. Environ., 48, 319–331.

440 Y. AWAYA et al.

Peterson, D. L., Spanner, M. A., Running, S. W. and Teuber, K. B. (1987): Relationship of ThematicMapper simulator data to leaf area index of temperate coniferous forests. Remote Sens. Environ.,22, 323–341.

Ranson, K. J., Irons, J. R. and Williams, D. L. (1994): Multispectral bidirectional reflectance ofnorthern forest canopies with the Advanced Solid-State Array Spectroradiometer (ASAS).Remote Sens. Environ., 47, 276–289.

Spanner, M. A., Pierce, L. L., Peterson, D. L. and Running, S. W. (1990): Remote sensing oftemperate coniferous forest leaf area index The influence of canopy closure, understoryvegetation and background reflectance. Int. J. Remote Sensing, 11, 95–111.

Strahler, A. H., Wu, Y. and Franklin, J. (1988): Remote estimation of tree size and density fromsatellite imagery by inversion of a geometric-optical canopy model. The 22nd InternationalSymposium on Remote Sensing of Environment, 227–384.

Trotter, C. M., Dymond, J. R. and Goulding, C. J. (1997): Estimation of timber volume in aconiferous plantation forest using Landsat TM. Int. J. Remote Sensing, 18, 2209–2223.

Y. Awaya (e-mail: awaya@ffpri.affrc.go.jp), S. Tsuyuki, E. Kodani and G. Takao