Utility of Landsat Thematic Mapper data for mapping site productivity in tropical moist forests

Postprint of: PHOTOGRAMMETRIC ENGINEERING AND REMOTE SENSING, Vol. 56, No. 10, October 1990, pp. 1383-1388

Utility of Landsat Thematic Mapper Data for
Mapping Site Productivity in Tropical Moist
Forests

Jerome K. Vanclay and Robert A. Preston
Queensland Forest Service, G.P.O. Box 944, Brisbane, Queensland 4001, Australia

ABSTRACT: Regression analysis was used to develop a relationship between Landsat 5
Thematic Mapper data and site quality of permanent sample plots. The thermal band (6) produced
misleading results, but the ratio of band 4 on 5 showed promise when combined with geological
information. Best results were obtained when regression analysis was restricted to ensure that at
least one each of the visible (1 to 3), mid-infrared (5 and 7) and near-infrared (4) bands were
included. Results suggest that Landsat TM may be useful for mapping site quality in tropical
moist forest in north Queensland. Interpolation of site productivity within a single pass seems
feasible, but attempts to extrapolate results to other passes or to imagery obtained on different
dates may be unreliable.
INTRODUCTION
One of the great challenges in managing the tropical moist forest (TMF) is efficient multiresource
inventory in these areas of rugged, inaccessible, and often poorly mapped terrain. One of the key
steps in this process is defining homogeneous units for sampling, mapping, and planning.
Although this is conceptually simple and is easily completed for well surveyed and mapped
forests such as industrial plantations, it remains a costly and difficult exercise in the TMF. Where
cloud free satellite imagery is available, remote sensing may provide a solution.
Stratification into homogeneous units for multiresource inventory is often on the basis of stand
density, productivity, and accessibility, although other factors may also be considered. Spectral
analysis of remote sensing imagery provides little help with assessing accessibility; this is the
domain of geographic information systems and digital terrain models where suitable data are
available, and of air photo interpretation and local knowledge elsewhere.
Forest managers are interested in several measures of stand density. Numbers and volumes per
unit area of certain species (usually commercial) are often desired, but cannot readily be
determined by remote sensing. Stand basal area, a forester's measure of biomass of trees, may be
significantly correlated with spectral signature, but the correlation is often so weak that the
relationships so established are of little practical use. Total (above ground) biomass may exhibit a
stronger correlation in temperate forests, but the correlation remains weak in TMF (Sader et al.,
1989). In any case, total biomass is notoriously difficult to measure in TMF and is rarely used in
management and planning.
Site productivity is of economic and ecological importance, indicating the potential for
commercial timber production as well as giving an indication of the species composition of the
forest. Cook et al. (1989) report a significant correlation between Landsat Thematic Mapper
(TM) data and biomass production in three temperature forests. Vanclay (1989) devised an
objective measure of site productivity for TMF in north east Queensland, and found a correlation
between site quality and Landsat TM data. The present study addresses whether those correlations

are specific to Vanclay's (1989) data or are more widely applicable, whether they can be
extrapolated from pass to pass, and how sensitive they are to season.
DATA
The present study concerns the TMF of northeast Queensland, between 16 and 19° S (Figure 1).
The environment and dynamics of these forests have been described by Stocker and Unwin
(1989). These forests have been managed for conservation and timber production for more than
80 years, and the Queensland Department of Forestry's (1983) research program has provided a
database of 250 permanent sample plots with a measurement history of up to 40 years (Vanclay,
1990). The plots range in size from 0.04 to 0.5 hectares and sample virgin, logged, and
silviculturally treated forests on a variety of soil types. Soil type is difficult to map, but geological
data are readily available, are correlated with growth patterns, and may be used as an indicator of
soil type (Queensland Department of Forestry, 1983, 1985). Vanclay's (1989) index of site quality
has been found to be significant in predicting the growth of these forest stands (Vanclay, 1991),
and has been operationally proven in yield forecasting (Vanclay and Preston, 1989).


FIG. 1. Location of study and permanent growth plots.

Site quality was estimated for each plot using Vanclay's (1989) Equation 13: i.e.,

where GI is the growth index of the plot; Dij is the diameter_ (breast high or above buttress, over
bark, in cm) of tree j of species i; DIij is its diameter increment (cm/ann); OBAij is its
"overtopping basal area," the basal area of trees within the plot that are bigger than tree ij
(m2/ha); BA is the plot basal area (m2/ ha); and a, /3 are parameters estimated by linear
regression. This equation estimates growth index, a measure of site productivity based on the
diameter increment adjusted for tree size and competition, of all trees of 18 reference species
(Acronychia acidula, Alphitonia whitei, Argyrodendron trifoliolatum, Cardwellia sublimis,
Castanospora alphandii, Cryptocarya angulata, C. mackinnoniana, Darlingia darlingiana,
Elaeocarpus largiflorens, Endiandra sp. aff. E. hypotephra, Flindersia bourjotiana, F.
brayleyana, F. pimenteliana, Litsea leefeana, Sterculia laurifolia, Syzygium kuranda, Toechima
erythrocarpum, and Xanthophyllum octandrum) using all available remeasures for the plot
(except that, where plots were remeasured more frequently, remeasurements were selected to
achieve approximately 5-year intervals). The (3s were estimated by fitting the equation
Log (DI + a) = Spp + D. Spp + Log(D).Spp + Log(BA) .Spp + OBA. Spp + Log (D) –Plot
(where Spp and Plot are qualitative variables) simultaneously for all these reference species in the
development data set (80 plots; a further 64 plots were used for validation studies). The value
0.08808 was subjectively determined to scale the growth indices into the range 0 to 10.
Fifteen plots were rejected and omitted from this analysis because the standard error associated
with the estimate of growth index exceeded 2 or because the growth index exceeded the
reasonable range (0 to 10). Other plots had insufficient reference species for the growth index and
its error to be computed. Six valid plots were outside the Landsat coverage, leaving 215 plots for
the present analysis.
Clearly, one cannot afford to maintain a permanent plot on each site for 30 years in order to
estimate the growth index, so more immediate methods for estimating and mapping site quality
are required. Vanclay (1989) developed equations to enable the growth index to be predicted
from the presence or absence of indicator species recorded at the site, e.g.,
G1= 5.882 + 0.973 S1 + 2.514 S2 + 1.545 S3 + 1.352 S4 - 0.970 S5 + 1.253 S6 - 0.951 S7-1.523 S8
where Si, i=1,2,...,8 are dummy variables which take the value one if present, zero if absent, and
represent the tree species Beilschmiedia obtusifolia, Helicia australasica, Lethedon setosa, Opis-
thiolepis heterophylla, Podocarpus neriifolius, Scolopia braunii, Syzygium wesa, and Xylopia
maccreai, respectively. Vanclay (1989) also estimated growth index from Landsat TM data,
suggesting that strata for timber inventory and yield forecasting could be prepared from Landsat
images. The present study further examines the utility of Landsat to stratify on the basis of site
quality.
Four Landsat scenes from two passes were available to the present study. These scenes were not
specifically selected for the study, but were two cloud-free scenes freely available to the authors.
The eastern pass included the Innisfail (95-72) and Ingham (95-73) scenes captured on 7 October
1987, and the western pass recorded the Cape Tribulation (96-71) and Atherton (96-72) scenes on
24 August 1986. None of the plots were logged during the period between the two passes. The
region experienced a cyclone during February 1986, in which some of the forest in the eastern
pass was damaged (Preston, 1987). The forests in the western pass suffered no damage, but
received widespread rain. Rainfall during March to August 1986 was erratic and below average,

with July particularly dry. August 1986 experienced average rainfall and temperature (Anon.,
1986). Good rainfall was experienced during the months prior to the October 1987 image, with
the four months June to September all experiencing above average rainfall (Anon., 1987). Natural
color images prepared from the three visible bands (1 to 3) revealed the October 1987 images to
be much greener, particularly for the pastoral country to the west of the TMF.
Ninety-six of the permanent sample plots with valid estimates of site quality occurred in the
overlap region of the two Landsat passes, 20 plots were to the east of the overlap region, and 99
plots were to the west. Thus, the data enabled an analysis of both the spatial and temporal
portability of equations predicting site quality from Landsat TM data. Parameters for equations
were estimated using the 96 plots from the overlap region, and the remaining 119 plots were used
to validate the utility of the equations. Only four of the six geological types recorded on
permanent sample plots occur in the overlap region. Thus, extrapolations to other geological
types inevitably result in a poor correlation. This is a realistic test of efficacy, as few managers
are fortunate enough to have complete information for all geographic data, as well as reliable
sample data in all strata.
Vanclay (1989) used 44 plots develop his equations, and a further 12 to validate them. All these
plots are included in the present database, and all but eight of his development plots lie in the
overlap region.
ANALYSIS
It is common practice to register remotely sensed images onto a map, and to transform the image
to the appropriate projection. However, to avoid the possible loss of spectral integrity," the
Landsat scenes were not transformed. Rather, the back-transformations were applied to the map
projection (UTM) to determine the locations of the permanent plots on the untransformed images.
Most remote sensing analyses employ classification to interpret and extrapolate images, but
regression analysis and other approaches can also be used (e.g., Cook et al., 1989; Sader et a1.,
1989; Vanclay, 1989). For data such as site quality, regression analyses provide the advantage of
predicting a continuum of values rather than a number of discrete classes. Cook et al. (1989) used
the average reflectance values from nine pixels comprising a 3-by 3-pixel window surrounding
and including the ground truth plot. Vanclay (1989) examined the use of the individual pixel
reflectance values, the average and median for a cluster of pixels about the plots, and the use of
some textural measures, but obtained the best results from the median of a 3 by 3 window
centered on the study plot. Thus, for each of the seven bands, the median of the nine pixels
centered on each plot was computed and used in these analyses. Basic manipulation of the
Landsat data was conducted using the DISIMP image processing system (CSIRO, 1988). Data
were then transferred to a general computer system, and were analyzed using two general purpose
statistical packages, GLIM (Payne, 1986) and Genstat (Payne et al., 1987).
Table 1 indicates the correlation matrices for the data derived from overlapping region of the
Landsat scenes. All the correIations between raw bands are significant (p<0.05), and this may
cause problems with stepwise regression analysis. Such correlation among the "independent" or
regressor variables results in a non-orthogonal design matrix. This means that parameter
estimates may change significantly when an additional variable is included in the model, and
detracts from the interpretation of parameter estimates (Neter and Wasserman, 1974). Using the
ratios of the bands reduces these correlations, reducing the problems with multicollinearity. It
also reduces the variation in illumination caused by topography and sun angle and, thus, in
reflectance recorded by Landsat TM, enhancing the portability of results.

TABLE 1. CORRELATION MATRIX FOR THE OVERLAP DATA SET (n=96)
Correlations for the October 1987 Data Set
TM1 TM2 TM3 TM4 TM5 TM6 TM7 Site Quality
1.000 0.550*** 0.697*** 0.236* 0.363*** 0.385*** 0.388*** -0.285** TM1
1.000 0.818*** 0.679*** 0.713*** 0.429*** 0.609*** -0.033 TM2
1.000 0.534*** 0.681*** 0.424*** 0.656*** -0.132 TM3
1.000 0.841*** 0.212* 0.595*** 0.137 TM4
1.000 0.487*** 0.888*** -0.078 TM5
1.000 0.550*** -0.233* TM6
1.000 -0.178 TM7
1.000 Site Quality
Correlations for the August 1986 Data Set
1.000 0.632*** 0.619*** 0.508*** 0.388*** 0.216* 0.326*** -0.160 TM1
1.000 0.734*** 0.646*** 0.652*** 0.360*** 0.495*** -0.212* TM2
1.000 0.416*** 0.471*** 0.220* 0.348*** -0.009 TM3
1.000 0.896*** 0.269** 0.813*** -0.130 TM4
1.000 0.412*** 0.883*** -0.133 TM5
1.000 0.351*** -0.294** TM6
1.000 -0.166 TM7
1.000 Site Quality
Correlations of individual bands between Data Sets
0.293** 0.474*** 0.529*** 0.584*** 0.719*** 0.396*** 0.640*** 1.000 TMi
TM1 TM2 TM3 TM4 TM5 TM6 TM7 Site Quality
***significant at P<0.0001, **P<0.01, *P<0.0
The rightmost column of Table 1 shows the correlations between the raw reflectance values and
the estimates of site quality. Two of the bands in each pass show a significant correlation. The
thermal band (6) is significant in both, band 1 is significant in the October 1987 pass, and band 2
is significant in the August 1986 pass. It is not clear whether the correlation with band 6 reflects a
true relationship with site quality. The TMF in north Queensland occur in a region of strong
relief, and plots used in the compilation of Table 1 range in altitude from sea level to 1200
metres. In the lowlands, forests are confined to the poorer soils, the better soils having been
converted to agriculture, whereas in the higher more rugged areas, better soils are still dominated
by forests. The correlation between altitude and site quality for the 96 plots in the overlap region
was + 0.14, and + 0.39 for all 215 plots. As band 6 may reflect altitude (r = + 0.23 in the October
1987 pass, r=- 0.07 in the August 1986 pass for the overlap region; r=- 0.29 in the October 1987
pass, r=- 0.30 in the August 1986 pass for the whole study region), it may be correlated with site
quality because of historical land-patterns rather than any causal relationship.
The bottom line of Table 1 shows how the reflectance spectra changed over time. All bands were
significantly correlated with the previous/subsequent image. The most similar reflectance was in
band 5 (r = 0.719), and the greatest difference in band 1 (r = 0.293).
Vanclay (1989) reported two equations for estimating site quality from Landsat TM. One was
based on a stepwise regression analysis of all raw bands, several transformations of these bands,
and all possible ratios of bands: i.e..

where Gl is the growth index, TM6 is the thermal Landsat TM band, and TM3 and TM, are the red
and near-infrared bands, respectively. Estimates derived from this equation were found to reflect
altitude and this was thought to be an artifact of the distribution of the permanent plots. This
equation performed poorly in the validation using independent data, and Vanclay (1989) does not
recommend its use.
Vanclay's (1989) preferred equation incorporated geology and used only ratios of Landsat bands:
i.e.,

where AL, AV, BV, CG, SM, TG are dummy (0,1) variables representing Alluvial, Acid
Volcanic, Basic Volcanic, Coarse grained Granite, Sedimentary-Metamorphic, and Tully (fine
grained) Granite geology, respectively, and where TM,, TMs, and TM, are the near-infrared, and
two mid-infrared Landsat TM bands, respectively. This equation provided a better fit to his data
(44 plots) than did Equation 1, and performed well in the independent validation. Vanclay (1989)
suggested that this equation may be useful for mapping site quality in the Landsat scene used in
his study.
The ability of these equations to predict site quality from other Landsat scenes is indicated in
Table 2. The leftmost column of correlations in Table 2 (overlapping region, some pass) indicates
the goodness-of-fit of the model to the calibration data. The four rightmost columns test the
validity of the calibrated equation under different scenarios: a different image over the same plots
(overlapping region, other pass), a different part of the same image (non-overlapping, same pass),
a different location and different image (non-overlapping, other pass) (i.e., extrapolating results to
the west or south-east of the study region, as well as using image recorded in a different season),
and an overall correlation using all data (including calibration data) (overall, both passes). It is
obvious that Equation 1 cannot be relied upon to provide accurate estimates of site quality, except
for the particular part of the image for which it was developed. Even within an image, Equation 1
may not give reliable results when extrapolated beyond the immediate locality of the ground truth
data. Table 2 supports Vanclay's (1989) contention that Equation 1 is unreliable.
TABLE 2. CORRELATIONS BETWEEN PREDICTED AND ACTUAL SITE QUALITY
Calibration Overlapping Region Non-overlapping Region Overall
Equation Data Same Pass Other Pass Same Pass Other Pass Both Passes
Using Landsat Data Only
Equation 1 44 plots (r=.51***) .02 -.28** -.31† -.23* -.10
Equation 1 October 1987 .32*** .19* .14† .08 .10
Equation 1 August 1986 .31** .14 .16 .28† .10
New Best Fit Equation
T4/T5 + T5/T4

October 1987

.45***

.002

-.li†

-.02

.10
T2/T3 + T3/T1 August 1986 .26** -.10 -.23* .002† -.02
1/T6 + T6 August 1986 .41*** -.21* .05 -.25† -.06
Mid-infrared, Near-infrared, Visible
T4/T5 + Tl (Eqn 3) October 1987

.45***

.14

.13†

.25**

.16
1/T2 + T5/T7 § August 1986 .26** .11 .41*** .12† .18
T4/T5 + T4/T2 § October 1987 .39*** .04 .03† .19* .01
T7/T3 + T2/T4 § August 1986 .21* .20* -.13 .31† .03
Using Landsat and Geographical (Geology) Data
Equation 2 44 plots (r=.75***) .35*** ** .46*† -.16 .12
Vanclay's (1989) Eqn 7 44 plots (r=.47**) .35*** .35*** .44*† -.21* .11
Equation 2 October 1986 .45*** .27** .50*† -.22*† .09
Equation 2 August 1986 .38*** .41*** -.19† .46*† .15
New Best Fit Equation
Geol + T4/T5 + T5/T4 October 1987

.53***

.14

-.42†

-.20

.15
Geol + T7/T3 + T2/T3 August 1986 .48*** .30*** - .25**‡ .40† -.08
Geol + 1/T6 + T2/T3 + T6 August 1986 .56*** -.16 -.0002† -.08† -.04
Mid-infrared, Near-infrared, Visible
Geol + T4/T5 + T1 (Eqn 4) October 1987

.50***

.33***

.47*†

.22*†

.19*
Geol + T7/T3 + T4/T2 § August 1986 .47"** .39*** -.20*‡ .54*† .11
Geol + T4/T5 + T4/T2 § October 1987 .46*** .30** .52*† -.21*† .10
Geol + T7/T3 + T2/T4 § August 1986 .46*** .39*** -.17‡ .52† .12
§ These variables not significant (p>0.05)
† These values based on 20 plots only; all others >95 plots.
‡ These values include estimates for "unknown" geological types. *** Significant at P<0.0001, ** P<0.01, * P<0.05

New parameters were estimated for Equation 1 by fitting it to the database for the overlap region.
Parameters were estimated separately for both passes, to enable validation of the relationship
using data from the other pass. Parameters were also estimated using the pooled data set for the
overlap region, but the results were disappointing, and equations based on a single pass generally
performed better. The correlation coefficients obtained for these fits were not as good as those
previously obtained. This is probably due to the larger and more diverse data set used in the
present study (96 plots versus 44 used by Vanclay 1989). The resulting equations, when extrap-
olated to other times and locations, provided rather low correlations (median 0.14), but they were
all positive, an improvement on Equation 1.
Equation 2, which includes geology, provided better results than Equation 1, although the
correlations are still not high. Equation 2 provides a reasonable estimate of site quality for areas
within the same pass, and within the same region in other passes, but cannot be reliably
determined in the west of the study region. Equation 2 performs better than estimates based solely
on geology (Vanclay, 1989, Equation 7), which also cannot be extrapolated to the west.
There were some difficulties in estimating new parameters for Equation 2 from the current
database, as two of the six geological types known to occur in the study area were not represented
in the data from the overlap region used to calibrate equations. The missing types (Alluvial and
Tully fine-grained Granite) were assumed to have the same characteristics as the Sedimentary-
Metamorphics, the default provided by the statistical package used and a reasonable assumption
(see Equation 2). Estimates prepared from the two re-calibrated versions of Equation 2 provided
rather good estimates on four of the geological types, but negative correlations where the two
missing types were included. This suggests that the two ratios, band 4 on band 5 and band 4 on
band 7, may provide useful results for investigating site quality, but only where reliable ground
truth and accurate geological maps are available. Figure 2 illustrates the correlation between site
quality and the ratio of band 4 on band 5.

FIG. 2. Scatterplot of site quality and ratio of bands 4/5 (October 1987). (× acid volcanic, ○ basic
volcanic, ∆ coarse granite, ● sedimentarymetamorphic)

New stepwise regression analyses were performed to investigate different formulations which
may provide better predictions. These regressions were performed on the data from the overlap
region only, for data from both passes separately, and the results validated with data from the
other pass, and from the non-overlapping region. Initially, regression analyses were performed
with only the ratios of bands. Additional analyses examined the use of raw reflectances, and of
several transformations of these. For the October 1987 data set, only the ratios were significant,
but the thermal band (6) featured prominently in the August 1986 data set.
The new "best fit" equations (Table 2) provide rather poor estimates. The equations containing
the thermal band (6) are unreliable, and it is suggested that this band should not be used when
using Landsat data to map site quality in the TMF of Queensland. The equations employing
geology give better results than those based solely on Landsat data, and allow more robust
estimates when extrapolated to other images. If prediction equations employ geology, it is
essential that training sites encompass the full range of geology in the region, as reliable
predictions cannot be made for the missing geological types.
Previous researchers (Badhwar et al., 1986; Benson and DeGloria, 1985; Sheffield, 1985) have
recommended one each of a mid-infrared (5 and 7), near-infrared (4), and visible (1 to 3) band be
employed to maximize information on an ecosystem. When regression analyses were constrained
to include at least one and not more than two terms containing these bands, the results were
markedly better. The best two equations were derived from the October 1987 scene: i.e.,

where AV, BV, CG, and SM are dummy (0,1) variables representing Acid Volcanic, Basic
Volcanic, Coarse grained Granite, and Sedimentary-Metamorphic geology, respectively, and
where TMl, TM4, and TM5 are the blue, near-infrared, and midinfrared Landsat TM bands,
respectively. The goodness-of-fit to the calibration data was only slightly worse than the best un-
restricted stepwise regression result, but these equations performed much better in validation with
data from different passes and different locations. This procedure also minimized bias due to
incorrectly classified geology. Although there are some advantages with using ratios of bands, no
advantage accrued from further restricting the regression analysis to include only ratios of bands.
The last two equations reported in each part of Table 2 were restricted to ratios, excluding ratios
of visible bands and of the mid-infrared bands, but provided no advantage or improvement.
Cook et al. (1989) found the ratio of like bands from different passes helpful. They used the ratio
of the June band 3 to the September band 3 reflectance which has a correlation of + 0.379 with
biomass production. In the present study, band 4 provided the only such ratio from the pooled
overlap data set. It exhibited a significant correlation (+0.319, p<0.01) but did not enter the
regression analysis as a significant variable.
Table 2 shows some differences in predictions based on the two passes. Generally, the October
1987 pass seems to provide better results. It consistently provides a better correlation coefficient

for the initial fit to the data. These equations also perform well when validated with independent
data, especially the equations including geology. Equations based on data from the August 1986
pass provided good results when validated against data from the October 1987 pass. This could
imply that October may be a more favorable time for assessing site quality, it may be an artifact
of the February 1986 cyclone, or it could imply that better results can be obtained when the
forests are not suffering moisture stress.
CONCLUSION
This study demonstrates that regression analysis approaches may be a useful way to analyze,
interpret, and extrapolate Landsat TM data when ground truth is available. It suggests that
Landsat TM may offer potential for mapping site productivity in TMF, but that the results may
depend upon the season. Best results may be obtained when other data sources (e.g., geology) are
also employed. The Landsat TM thermal band 6 produced misleading results in Queensland
TMF, and should not be used to extrapolate site quality data without independent validation
studies. The ratio of band 4 on 5 seemed to show most promise for site quality mapping,
especially when used in conjunction with geology. Best results were obtained when simple
stepwise regression was not used, but when regression analysis was restricted to ensure that at
least one each of the visible (1 to 3), mid-infrared (5 and 7), and near-infrared (4) bands were in-
cluded. The resulting equations could be expected to give reasonable site quality estimates within
the pass for which they were calibrated, but should not be extrapolated to other passes without
further validation.
ACKNOWLEDGMENTS
Many officers of the Queensland Department of Forestry have been involved in the
establishment, measurement, and maintenance of the permanent plots and the associated database.
Tim Danaher of the Queensland Sunmap Centre assisted in processing the Landsat data.
REFERENCES
Anon., 1986. Monthly Weather Review, Queensland, August 1986. Bureau of Meteorology,
Department of Science.
Anon., 1987. Monthly Weather Review, Queensland, September 1987. Bureau of Meteorology,
Department of Science.
Badhwar, G. D., R. B. MacDonald, and N. C. Mehta, 1986. Satellite-derived leaf-area-index and
vegetation maps as input -to global carbon cycle models-a hierarchical approach. Int. J. Remote
Sens. 7:265281.
Benson, A. S., and S. A. DeGloria, 1985. Interpretation of Landsat-4 thematic mapper and
multispectral scanner data for forest survey. Photogrammetric Engineering & Remote Sensing
51:1281-1290.
CSIRO, 1988. DISIMP Device Independent Software for Image Processing, Version 4.1, System
Reference Manual for Unix DISIMP. CSIRO Division of Information Technology, Center for
Spatial Information Systems, Canberra, Australia.
Cook, E. A., L. R. Iverson, and R. L. Graham, 1989. Estimating forest productivity with thematic
mapper and biogeographical data. Remote Sens. Environ. 28:131-141.

Neter, J., and W. Wasserman, 1974. Applied Linear Statistical Models: Regression, Analysis of
Variance, and Experimental Designs, Irwin, Illinois, 842 p. ,
Payne, C. D. (ed.), 1986. The GLIM System. Release 3.77 Manual. Numerical Algorithms Group,
Oxford, U.K. 3 vols.
Payne, R. W., P. W. Lane, A. E. Ainsley, K. E. Bicknell, P. G. N. Digby, S. A. Harding, P. K.
Leech, H. R. Simpson, A. D. Todd, P. J. Verrier, R. P. White, J. C. Gower, G. Tunnicliffe Wilson,
and L. J. Paterson, 1987. Genstat 5 Reference Manual. Oxford Science, Clarendon, Oxford. 749
p.
Preston, R. A., 1987. Use of Landsat-5 MSS imagery to map cyclone damage to rainforests in
north Queensland Fourth Australasian Conference on Remote Sensing (D. Bruce, ed.) Adelaide,
1987, pp. 250260.
Queensland Department of Forestry, 1983. Rainforest Research in North Queensland. A position
paper issued by the Queensland Department of Forestry. 52 p.
Queensland Department of Forestry, 1985. Research Report No. 5. Queensland Department of
Forestry, 100 p.
Sader, S. A., R. B. Waide, W. T. Lawrence, and A. T. Joyce, 1989. Tropical forest biomass and
successional age class relationships to a vegetation index derived from Landsat TM data. Remote
Sens. Environ. 28:143-156.
Sheffield, C., 1985. Selecting band combinations from multispectral data. Photogrammetric
Engineering & Remote Sensing 51:681-687.
Stocker, G. C., and G. L. Unwin, 1989. The rainforests of northeastern Australia-Their
environment, evolutionary history and dynamics. Tropical Rainforest Ecosystems:
Biogeographical and Ecological Studies. (H. Leith and M.J.A. Wenger, eds.) Elsevier,
Amsterdam.
Vanclay, J. K., 1989. Site productivity assessment in rainforests: an objective approach using
indicator species. Proceedings of the Seminar on Growth and Yield in Tropical Mixed/Moist
Forests (Wan Razali Mohd., H.T. Chan, and S. Appanah, eds.) 20-24 June 1988, Kuala Lumpur.
Forest Research Institute Malaysia, pp. 225-241.
Vanclay, J. K., 1991. Aggregating tree species to develop diameter increment equations for
tropical rainforests. For. Ecol. Manag. 42:143-168.
Vanclay, J. K., 1990. Effects of selection logging on rainforest productivity. Australian Forestry
53:200-214.
Vanclay, J. K., and R. A. Preston, 1989. Sustainable timber harvesting in the rainforests of
northern Queensland. Forest Planning for People, Proceedings of 13th biennial conference of the
Institute of Foresters of Australia, Leura, NSW, 18-22 September 1989, pp. 181-191.
(Revised and accepted 12 June 1990)