Incorporating Whole-Stand and Individual-Tree Models in a

Stand-Table Projection System

Quang V. Cao

Abstract: A stand table provides number of trees per unit area for each diameter class. This article presents
three methods to project a current stand table into the future by predicting mortality and diameter growth for each
diameter class by use of an individual-tree model. The stand table was then adjusted to produce the same total
number of trees and basal area per hectare as predicted from the whole-stand model. The three methods evaluated
in this study were stand table adjustment, constrained least squares (LS), and modified constrained LS. Data from
the Southwide Seed Source Study of loblolly pine (Pinus taeda L.) showed that incorporating the individual-tree
model helped improve the projection of stand tables, as compared to a previous approach. The three methods
produced comparable results; error indices from these methods were within 5% of one another. The constrained
LS method consistently provided the best fit (lowest error indices) as compared to the other methods. FOR. SCI.

Downloaded from https://academic.oup.com/forestscience/article/53/1/45/4604362 by guest on 09 March 2024

53(1):45– 49.
Keywords: pinus taeda, loblolly pine, constrained least squares, tree survival, diameter growth

A
STAND TABLE, usually developed from a conven- tion errors among multiple growth periods from the same
tional forest inventory, shows number of trees in plot. These observations constituted the fit data set, to be
each diameter class. By characterizing the distri- used for developing the whole-stand and individual-tree
bution of diameters in a stand, it allows calculations of growth models. Another 100 plots were randomly selected
product volumes and, finally, financial value of that partic- from the rest of the original data to form the evaluation data
ular stand. Growth modeling based on projection of stand set. All possible growth periods from these plots were
tables started with simple assumptions on tree diameter included in the evaluation data set, totaling 300 short-pro-
growth and mortality (Chapman and Meyer 1949, Avery jection (4 –7 yr) periods, 200 medium-projection (9 –12 yr)
and Burkhart 2002). Predictions from regression equations periods, and 100 long-projection (15–17 yr) periods. Sum-
for stand survival and basal area were used to compute mary statistics for the fit and evaluation data sets appear in
future stand tables by Clutter and Jones (1980) and Pienaar Table 1.
and Harrison (1988). These approaches deviated from sim-
ple stand table projection systems in the past, even though Methods
the assumption that trees in each class were uniformly
distributed still remained. Nepal and Somers (1992) devel- The system proposed in this article for stand table pro-
oped a stand table projection algorithm in which trees in jection requires a current stand table, a tree survival and
each class were assumed to follow a truncated Weibull diameter growth model, and a stand-density projection
distribution. The algorithm was later modified by Cao and model. The current stand table is projected into the future by
Baldwin (1999a, b). use of the following steps.
The objective of this study was to incorporate predictions Step 1: Compute Survival.—All mortality is assumed
from whole-stand and individual-tree growth models into a to occur at the beginning of the growing period. The tree
reliable system for projecting a current stand table into the survival equation is used to predict surviving number of
future. trees in each diameter class.
Step 2: Grow Diameters.—All trees in each class in-
crease in diameter during the growing period. This process
Data is simplified by projecting the lower and upper limits of
Data used in this study were from unthinned plantations each diameter class into the future, using the tree diameter
in the Southwide Seed Source Study, which included 15 growth equation.
seed sources of loblolly pine (Pinus taeda L.) planted at 13 Step 3: Reclassify Trees.—After diameter growth in
locations across 10 southern states (Wells and Wakeley step 2, trees in each class either stay in that class or move up
1966). Each plot of size 0.0164 ha consisted of 49 trees to higher classes. Assuming that trees in each diameter class
planted at a 1.8 m ⫻ 1.8 m spacing. Tree diameters were follow a truncated Weibull distribution (Nepal and Somers
measured at ages 10, 15 (or 16), 20 (or 22), and 25 (or 27) 1992), these trees are then reclassified into new diameter
years. A subset of the original data (100 plots) was ran- classes.
domly selected, and one growth period (ranging from 4 to 7 Step 4: Adjust Number of Trees.—The resulting fu-
years) was randomly chosen for each plot to avoid correla- ture stand table does not produce the same stand density in

Quang V. Cao, Professor of Forestry, Louisiana State University, School of Renewable Natural Resources, Baton Rouge, LA 70803—Phone: (225) 578-4218;
Fax: (225) 578-4227; [email protected].

Manuscript received June 16, 2006, accepted September 11, 2006 Copyright © 2007 by the Society of American Foresters

Forest Science 53(1) 2007 45

Table 1. Means (and standard deviations) of stand variables, by data type and age

Number Dominant Number of trees Basal area Minimum Maximum

Age of plots height (m) (no./ha) (m2/ha) diameter (cm) diameter (cm)
Fit data set1
10 25 9.00 (1.44) 2,238 (388) 24.48 (6.34) 5.9 (1.8) 16.4 (2.4)
15 38 12.89 (1.46) 1,773 (556) 31.96 (7.67) 8.1 (2.3) 21.7 (3.1)
16 8 16.38 (1.06) 1,289 (194) 30.92 (3.11) 10.4 (1.7) 23.7 (1.5)
20 45 16.96 (1.96) 1,159 (360) 31.28 (8.50) 11.7 (2.6) 25.4 (3.7)
22 2 15.32 (0.15) 580 (216) 12.96 (6.51) 10.9 (2.2) 20.3 (1.1)
25 17 18.46 (1.64) 1,105 (426) 31.85 (11.06) 12.6 (2.6) 26.1 (3.0)
27 9 20.59 (2.68) 739 (227) 30.64 (10.27) 14.8 (3.4) 30.2 (4.6)
Evaluation data set2
10 100 9.04 (1.61) 2,071 (437) 22.13 (6.12) 5.2 (2.0) 16.2 (2.7)
15 86 13.04 (1.96) 1,735 (463) 30.66 (6.20) 8.2 (2.7) 20.9 (3.2)
16 14 16.16 (1.13) 1,273 (194) 29.89 (3.25) 9.0 (1.8) 23.8 (2.0)
20 95 17.05 (2.21) 1,301 (323) 35.23 (7.70) 11.6 (2.1) 25.5 (3.5)

Downloaded from https://academic.oup.com/forestscience/article/53/1/45/4604362 by guest on 09 March 2024

22 5 14.00 (2.11) 952 (119) 19.11 (8.23) 8.3 (1.1) 22.0 (4.3)
25 72 19.66 (2.46) 1,208 (328) 40.35 (9.31) 13.4 (2.5) 28.5 (4.0)
27 28 20.47 (3.10) 795 (201) 32.35 (6.82) 15.1 (3.3) 30.7 (4.0)
1
The fit data set consisted of 100 plots. Growth periods (one randomly selected per plot) spanned from age 10 to 15 yr (30 plots), 10 to 16 yr (4 plots),
15 to 20 yr (22 plots), 16 to 20 yr (11 plots), 20 to 25 yr (20 plots), 20 to 27 yr (7 plots), and 22 to 27 yr (6 plots).
2
The validation data set consisted of 100 plots. All possible growth periods from each plot resulted in a total of 600 observations: 4 to 7 yr (300), 10 to
12 yr (200), and 15 to 17 yr (100).

terms of number of trees and basal area per hectare as Constrained Least-Squares (LS) Method:
predicted from the whole-stand model. In this step, the
As the name implies, the method involves minimizing
number of trees in each diameter class is adjusted to match
the predicted number of trees and basal area per hectare.
Three adjustment methods were investigated in this study. 冘 共n ⫺ n 兲 ,
*
i i
2
(2)

subject to two constraints that total number of trees and

Stand Table Adjustment Method basal area are the same as predicted from the stand-density
projection model. The constrained LS method was origi-
Number of trees per hectare in the ith diameter class is nally developed by Matney et al. (1990) for allocating
adjusted using Nepal and Somers’ (1992) procedure as mortality and diameter growth to the tree list, and was
follows: modified by Cao and Baldwin (1999a) to adjust the future
stand table. The approach used in this study is slightly
n *i ⫽ n i ␣ exp(␤Di), (1)
different from Cao and Baldwin’s (1999a) method in that a
where n*i and ni are adjusted and unadjusted number of trees third constraint involving average diameter is omitted. The
per hectare in the ith class, respectively, Di is the midpoint adjusted number of trees per hectare in the ith diameter class
of the ith class, and ␣ and ␤ are parameters to be solved is given as
such that total number of trees and basal area match those
predicted from the whole-stand model. n *i ⫽ n i ⫹ ␭ 1 ⫹ ␭ 2 D 2i , (3)

Table 2. Regression equations to predict stand and tree attributes

Attribute Equation1
Dominant height Ht⫹1 ⫽ exp关3.9128 ⫹ (ln(Ht ⫺ 3.9128) (At/At⫹1)0.6609兴 ⫹ ␧ R2 ⫽ 0.92; RMSE ⫽ 0.90 m

Number of trees/ha Nt⫹1 ⫽ exp关(At/At⫹1)ln(Nt)⫹ (1 ⫺ At/At⫹1)(3.8705 ⫹ 16.3614/At ⫹ 0.0008 Nt)兴 ⫹ ␧

R2 ⫽ 0.82; RMSE ⫽ 268 trees/ha

Basal area/ha Bt⫹1 ⫽ exp关(At/At⫹1)ln(Bt) ⫹ (1 ⫺ At/At⫹1){3.1769 ⫹ 11.3393/Ht ⫺ 0.0120 Bt}兴 ⫹ ␧

R2 ⫽ 0.70; RMSE ⫽ 5.34 m2/ha

Tree diameter di,t⫹1 ⫽ di,t ⫹ (5.6043 Dqt⫺0.4350)/关1 ⫹ exp(2.2280 ⫺ 14.6569/At⫹ 0.0760 Ht ⫹ 0.03972 Bt ⫺ 0.1530 di,t)兴 ⫹ ␧
R2 ⫽ 0.94; RMSE ⫽ 1.28 cm

Tree survival pi,t ⫹1 ⫽ pi,t/关1 ⫹ exp(3.7112 ⫺ 0.0254 Ht ⫺ 13.0205 RSt ⫺ 4.2899 di,t/Dqt)兴
probability MD ⫽ 0.0015; MAD ⫽ 0.2862
1
H ⫽ dominant height; A ⫽ stand age in years; N ⫽ number of trees/ha; B ⫽ basal area/ha; Dq ⫽ quadratic mean diameter in cm; RS ⫽ (10,000/N)0.5/H ⫽
relative spacing; subscripts t and (t ⫹ 1) denote measurements at age At and At⫹1, respectively, where At⫹1 ⫽ At ⫹ 1. pi,t and di,t ⫽ survival probability
and diameter of tree i at age At, respectively; ␧ ⫽ error term; RMSE ⫽ root-mean-squared error; MD ⫽ mean difference; and MAD ⫽ mean absolute
difference.

46 Forest Science 53(1) 2007

where ␭j values are Lagrangian multipliers computed from
冘 兩b ⫺ b̂ 兩,
mj

EIBj ⫽ ij ij (7)
共⌺D 2i 兲共⌺n i ⫺ N̂ 2 兲 ⫹ m共B̂ 2 /K ⫺ ⌺n i D 2i 兲 i⫽1
␭2 ⫽ (4)
m⌺D 4i ⫺ 共⌺D 2i 兲 2
where EINj and EIBj are error indices based on number of
and trees/ha and basal area/ha for the jth plot, respectively; nij
and n̂ij are observed and predicted number of trees per
N̂ 2 ⫺ ⌺n i ⫺ ␭ 2 ⌺D 2i hectare of the ith diameter class in the jth plot, respectively;
␭1 ⫽ , (5)
bij and b̂ij are observed and predicted basal area in m2/ha of
m
the ith diameter class in the jth plot, respectively; and mj is
where N̂2 and B̂2 are predicted number of trees and basal the number of diameter classes in the jth plot.
area per hectare, respectively, at the end of the growth
period, K ⫽ ␲/40,000 ⫽ a constant to convert diameter in
cm to basal area in m2, m ⫽ number of diameter classes, and Results and Discussion
the summation sign denotes the sum overall diameter
Equations comprising the whole-stand and individual-
classes.

Downloaded from https://academic.oup.com/forestscience/article/53/1/45/4604362 by guest on 09 March 2024

tree models are listed in Table 2. These equations were
carefully selected after evaluating many alternatives, based
on R2, mean difference, and mean absolute difference. An-
Modified Constrained LS Method nual prediction equations were used in this study because
The surviving number of trees in each diameter class com- they had been demonstrated to work well in projecting both
puted from step 1 was proportionally adjusted to match total stand-level and tree-level attributes (Cao 2000, Ochi and
stand survival predicted from the stand survival model. Cao 2003, Cao 2004). Since cross-equation correlations
Then the diameter growth from step 2 was also proportion- likely existed among error components (Borders 1989),
ally adjusted such that the sum over all diameter classes parameters of the equations to predict number of trees and
produced the same basal area per hectare as projected by the basal area per hectare were simultaneously estimated with
stand basal area prediction model. The next two option SUR (seemingly unrelated regression) of SAS pro-
steps—reclassifying trees and adjusting number of cedure MODEL (SAS Institute Inc. 1993). Individual-tree
trees—were carried out as in the constrained LS method. survival and diameter growth were expressed as functions
of tree diameter and stand-level variables (age, dominant
height, stand density, and quadratic mean diameter).
Evaluation The means and standard deviations of Reynolds et al.’s
(1988) error indices for the three adjusting methods based on
The three methods of incorporating the individual-tree the evaluation data set were presented in Table 3. A low value
model were evaluated on their abilities to project future of error index indicated little difference between the observed
stand tables. Also included in the evaluation was Nepal and and predicted numbers of trees in each diameter class.
Somers’ (1992) method, in which the diameter growth func- Without exception, the constrained LS method consis-
tion was derived from the Weibull distribution. tently produced the lowest mean error index values, based
Two forms of error index proposed by Reynolds et al. either on number of trees or basal area, among the three
(1988) were used to determine how well the three methods methods for all projection lengths. However, the differences
performed at the diameter-class level: among the three methods were small. For EIN, they were

冘 兩n ⫺ n̂ 兩,
mj less than 1% for short and medium projections and 2.5% for
EINj ⫽ ij ij (6) long projections. The differences increased in magnitude for
i⫽1 EIB, from 2.1% for short projections to 4.4% for long

Table 3. Means (and standard deviations) of Reynolds et al.’s (1988) error indices based on number of trees/ha and basal area/ha for four methods

Projection length
1
Method Short (4–7 yr) Medium (9–12 yr) Long (15–17 yr)
2
Error index based on number of trees/ha
Stand table adjustment 627 (230) 692 (209) 729 (207)
Constrained least squares 623 (226) 686 (204) 711 (195)
Modified constrained LS 626 (225) 689 (202) 718 (184)
Nepal and Somers (1992) 688 (225) 778 (186) 821 (171)
Error Index based on basal area/ha2
Stand table adjustment 15.88 (6.38) 20.66 (6.76) 24.67 (7.08)
Constrained least squares 15.72 (6.33) 20.28 (6.83) 23.90 (7.15)
Modified constrained LS 16.05 (6.32) 20.76 (6.64) 24.74 (6.74)
Nepal and Somers (1992) 17.47 (6.99) 23.32 (7.27) 28.43 (7.71)
1
The first three methods are adjusting methods proposed in this study. The last one was developed by Nepal and Somers (1992).
2
For each type of error index and each projection length, numbers in bold denote the smallest mean (best), and underlined numbers denote the largest mean
(worst) among the four methods.

Forest Science 53(1) 2007 47

projections. A summary for error index values was also adjustment method projects diameters from the tree diame-
included in Table 3 for the Nepal and Somers (1992) ter growth function. These changes should result in a more
method, which consistently provided the highest mean error realistic distribution of trees before the adjustment step, and
index values for all growth periods. did provide better predictions for future stand tables. Com-
Table 4 shows mean Reynolds et al.’s (1988) error pared to the stand table adjustment method, mean values of
index based on number of trees per ha (EIN) by diameter error index for Nepal and Somers’ (1992) approach (based
class for each of the four methods. This table confirms on data used in this study) increased between 9 and 11% for
findings from Table 3 that the three new methods pro- EIN and between 8 and 15% for EIB (Table 3). Similar
duced similar results, with the constrained LS slightly findings were also obtained when the implied diameter
better. The Nepal and Somers (1992) method, however, growth function (Bailey 1980) was substituted for the indi-
did not perform well for small-diameter classes (below vidual-tree model in the constrained LS method.
14 cm) and over-extended the future stand table to be- Furthermore, Table 4 shows that the implied diameter
yond 40 cm. Similar results were also obtained when growth function in the Nepal and Somers (1992) method
basal area was substituted for number of trees as basis for over-predicted future diameters for large diameter classes
error index for each diameter class. (over 40 cm). These results confirmed that incorporating the

Downloaded from https://academic.oup.com/forestscience/article/53/1/45/4604362 by guest on 09 March 2024

individual-tree model did improve the projection of stand
tables.
Effects of Incorporating the Individual-Tree
Model
Stand Table Adjustment versus Constrained LS
The benefits of incorporating the individual-tree model
into the stand table projection system were evident when
Method
contrasting the stand table adjustment method to the one Cao and Baldwin (1999a) used observed stand attributes
introduced by Nepal and Somers (1992). These two ap- to evaluate the stand table adjustment approach against the
proaches were similar, except for two major differences. constrained LS method; the latter used number of trees,
First, number of trees in each class was not modified before basal area per hectare, and average diameter as constraints.
the adjustment by Nepal and Somers (1992), but was re- In this study, predicted number of trees and basal area per
duced by use of the tree survival equation in the stand table hectare were used instead. In addition, the average diameter
adjustment method. Second, Nepal and Somers (1992) used constraint was dropped because it seemed to cause difficulty
an implied diameter growth function derived from the in convergence for many plots. Results from Cao and Bald-
Weibull distribution (Bailey 1980), whereas the stand table win (1999a) and this study consistently showed that the

Table 4. Mean Reynolds et al.’s (1988) error index based on number of trees/ha by diameter class for four methods. The four methods are stand
table adjustment (STP), constrained least squares (CLC), modified constrained least squares (MCLS), and Nepal and Somers (NP)

Short (4–7 yr) Medium (9–12 yr) Long (15–17 yr)

DBH
(cm) STP CLS MCLS NP STP CLS MCLS NP STP CLS MCLS NP
2 21 36 19 60 0 24 3 32 0 13 1 21
4 21 30 16 51 2 26 2 29 0 15 0 17
6 24 33 21 50 4 17 5 28 1 10 2 17
8 45 45 45 58 13 17 11 31 3 8 3 20
10 69 67 66 79 34 34 32 46 20 20 15 35
12 73 73 73 74 58 55 56 59 46 44 44 48
14 88 80 87 83 84 78 84 87 84 78 79 80
16 85 82 90 91 94 86 91 89 81 73 81 85
18 81 81 82 86 96 88 99 97 81 74 81 79
20 74 73 74 76 97 93 95 102 108 100 107 110
22 55 55 55 59 68 68 69 71 85 75 81 75
24 54 54 54 52 63 63 64 62 79 74 79 75
26 45 44 44 45 57 56 57 54 66 65 65 67
28 39 39 40 38 53 56 48 50 54 60 58 59
30 41 40 40 34 44 44 45 38 51 55 44 50
32 31 31 32 31 29 29 28 30 31 31 30 32
34 38 38 36 24 26 28 30 22 34 31 34 23
36 25 25 26 21 24 23 14 19 26 25 20 17
38 39 36 61 33 4 1 34 20 30 31 40 17
40 18 16 15 36 15 11 33 15 28 26 30 29
42 11 2 4
44 7 6 2
46 6 4 2
48 2 2
50 1
52 1
For each diameter class and each projection length, numbers in bold denote the smallest mean (best) and underlined numbers denote the largest mean (worst)
among the four methods.

48 Forest Science 53(1) 2007

constrained LS method outperformed the stand table adjust- Literature Cited
ment method in projecting a stand table into the future. AVERY, T.E., AND H.E. BURKHART. 2002. Forest measurements,
5th ed. McGraw-Hill, New York. 456 p.
BAILEY, R.L. 1980. Individual tree growth derived from diameter
Constrained LS versus Modified Constrained distribution models. For. Sci. 26:626 – 632.
LS Method BORDERS, B.E. 1989. Systems of equations in forest stand model-
ing. For. Sci. 35:548 –556.
The “least squares” strategy of the constrained least CAO, Q.V. 2000. Prediction of annual diameter growth and sur-
squares method aimed to make “minimum” changes to the vival for individual trees from periodic measurements. For. Sci.
future stand tables to satisfy the two stand density con- 46:127–131.
straints. One might think intuitively that if the future stand CAO, Q.V. 2004. Annual tree growth predictions from periodic
table was first adjusted to produce totals close to the target measurements. P. 212–215 in Proc. of the South. Silvic. Res.
stand density, then the subsequent least squares adjustment Conf. USDA For. Serv. Gen. Tech. Rep. SRS-71.
would require only minor changes in the stand table and CAO, Q.V., AND V.C. BALDWIN JR. 1999a. A new algorithm for
might lead to better results. Table 3 shows that this extra stand table projection models. For. Sci. 45:506 –511.
CAO, Q.V., AND V.C. BALDWIN JR. 1999b. Projecting a stand table

Downloaded from https://academic.oup.com/forestscience/article/53/1/45/4604362 by guest on 09 March 2024

adjustment included in the modified constrained LS method
through time. P. 515–518 in Proc. of the South. Silvic. Res.
was not necessary, and in this case even undesirable. This
Conf. USDA For. Serv. Gen. Tech. Rep. SRS-30.
method consistently produced higher means of error index
CHAPMAN, H.H., AND W.H. MEYER. 1949. Forest mensuration.
than did the constrained LS, and even ranked last based on McGraw-Hill, New York. 522 p.
EIB. From Table 4, the modified constrained LS method CLUTTER, J.L., AND E.P. JONES JR. 1980. Prediction of growth
performed better for the smaller diameter classes, but lost after thinning of old-field slash pine plantations. USDA For.
ground for the middle and larger diameter classes. Serv. Res. Pap. SE-217. 14 p.
MATNEY, T.G., K.L. BELLI, AND R.M. FARRAR JR. 1990. Param-
eter-free diameter-distribution recovery. P. 95–107 in Division
Summary 4 Proc. International Union of Forestry Research Organisa-
tions (IUFRO) World Congress meet. Pub. FWS-2-90; School
The general stand table projection system presented in of For. & Wildl. Resource., VPI&SU, Blacksburg, VA.
this article involved (1) computing survival of trees in each NEPAL, S.K., AND G.L. SOMERS. 1992. A generalized approach to
diameter class from the tree survival equation, (2) projecting stand table projection. For. Sci. 38:120 –133.
diameters using the tree diameter growth equation, (3) re- OCHI, N., AND Q.V. CAO. 2003. A comparison of compatible and
classifying trees into new diameter classes, and (4) adjusting annual growth models. For. Sci. 49:285–290.
number of trees in each class to match total number of trees PIENAAR, L.V., AND W.M. HARRISON. 1988. A stand table projec-
and basal area per hectare as predicted from the whole-stand tion approach to yield prediction in unthinned, even-aged
model. Results indicated that incorporating the individual- stands. For. Sci. 34:804 – 808.
REYNOLDS, M.R., T.E. BURK, AND W.C. HUANG. 1988. Goodness-
tree model helped improve the projection of stand tables, as
of-fit tests and model selection procedures for diameter distri-
compared to the previous method suggested by Nepal and bution models. For. Sci. 34:373–399.
Somers (1992). The three adjusting methods (stand table SAS INSTITUTE INC. 1993. SAS/ETS user’s guide, version 6, 2nd
adjustment, constrained LS, and modified constrained LS) ed. SAS Institute Inc., Cary, NC. 1022 p.
produced comparable error indices, but the constrained LS WELLS, O.O., AND P.C. WAKELEY. 1966. Geographic variation in
consistently provided the best results as compared to the survival, growth, and fusiform rust infection of planted loblolly
other methods. pine. For. Sci. Monogr. 11. 40 p.

Forest Science 53(1) 2007 49