Amaro A., Reed D., Soares P. (editors) Modelling Forest Systems

Подождите немного. Документ загружается.

34Amaro Forests - Chap 30 1/8/03 11:54 am Page 356

356 F.-J. Knauft

dynamic models in science resembles competition more than mutual support. It is

clear that interfaces that remain useful between two (or even among all three) of

these groups are highly non-trivial.

While growth is simulated in individual-based growth models at the level of

individual plants, its evaluation is commonly done at the level of stands or even

landscapes. Ther

efore, upscaling becomes mere aggregation. The individual-tree-

based forest growth simulators

TRAGIC++, SILVA and BWIN realize a solution for the

step from the individual plant to the stand. The next level of aggregation requires

different tools, such as forest information systems for management support, or the

visualization system

VIWA for communication support.

The general decision for a modular system and wide integration of available

softwar

e components eased the transfer of the methodology from the ﬁrst

VIWA ver-

sion up to the

TRAGIC environment. While preserving the general concept, an

exchange of a case-adapted simulation or visualization tool with another user

required only small changes within the interface tools. Further, it enabled us to

enhance the possibilities of the systems from case to case.

The major step from the ﬁrst

VIWA version to the second was the integration of

the full system on a single computer under a common GUI, therefore greatly

enhancing its efﬁciency. Simultaneously, it reduced the necessary technical knowl-

edge for the user. This goes even further in the

TRAGIC environment. Here, the inter-

active graphical interface providing the necessary technical knowledge for its

application is similar to other desktop applications. This increases the group of

potential users on a large scale. The server–client design opens the way for an inter-

active communication between the users over the graphical interface.

In science in general, and in ecosystem research in particular, a dynamic sys-

tems appr

oach is the typical solution to this task, i.e. the behavioural (temporal

changes) and structural complexities are derived from a realistic representation of

stand dynamics. This model type brings the perceptional ‘bias’ of scientists close to

the aesthetics of the public: more structural resolution (mapping of perceived com-

plexity) is regarded as a path to more realistic dynamics, but the addition of com-

plexity for its own sake, without an accompanying increase in predictive capacity or

utility, almost never results in an improved model.

In forest management, another (speculative) type of grounding has been

described (Hauhs and Lange, 2001), in which gr

owth behaviour and response is the

ultimate foundation of modelling (here empirical). The reduced complexity of forest

growth as seen from the perspective of forestry occurs at a relatively coarse resolu-

tion: the rotation period in the temporal and the stand/group development in the

spatial perspective. The standards of silvicultural management describe an interac-

tive situation in which the current structure of a stand is used to assess and decide

about the appropriate interference. It is typical that the goal of any given interfer-

ence is a regulation of competition among trees that leaves the maximum number of

choices for the next round of interference (rather than relying on a prediction far

into the future of this particular stand). In contrast to many other modelling

schemes, the visualization interface becomes the centrepiece through which other

components are linked.

While the proposed three-component concept is not intended to solve all the

oblems of data transfer between practical forestry and modelling (it will be a

future challenge to foresters and scientists as well), nor does it try to substitute

other concepts of forestry information systems or silvicultural decision support

systems, it has served us as a learning environment to study the various methods

of interaction and information exchange around the topic of forest growth and

management.

34Amaro Forests - Chap 30 1/8/03 11:54 am Page 357

357 Landscape Visualization with Forest Growth Simulators

Acknowledgements

This work was supported by the German Ministry of Research (BMBF) (contract no. PT BEO 51-

0339476 C), the Lower Saxony Forestry Research Station and the Faculty of Forestry and Forest Ecology

at the University of Göttingen.

References

BITOEK (2002a) Documentation for the Forest Growth Simulator TRAGIC++. Bayreuth Institute

for Terrestrial Ecosystem Research, Ecological Modelling. Available at: www.bitoek.

uni-bayreuth.de/mod/html/tragic/documentation

BITOEK (2002b) Homepage of the Interactive Forest Growth Simulator Interface JTRAGIC. Bayreuth

Institute for Terrestrial Ecosystem Research, Ecological Modelling. Available at:

www.bitoek.uni-bayreuth.de/mod/html/webapps/jtragic

Hauhs, M. and Lange, H. (2001) Lessons from sustainable development. In: von Gadow, K.,

Pukkala, T. and Tomé, M. (eds) Sustainable Forest Management. Kluwer, Dordrecht, pp.

69–98.

Hauhs, M., Rost-Siebert, K., Kastner-Maresch, A.E. and Lange, H. (1993) A New Model Relating

Forest Growth to Input and Output Fluxes of Energy and Matter of the Corresponding Ecosystem:

TRAGIC (Tree Response to Acidiﬁcation of Groundwater in Catchments). Final Report of EC

Project EV4V-0032-D(B), Institut für Bodenkunde und Waldernährung der Universität

Göttingen.

Hauhs, M., Kastner-Maresch, A. and Rost-Siebert, K. (1995) A model relating forest growth to

ecosystem-scale budgets of energy and nutrients. Ecological Modelling 83, 229–243.

Hauhs, M., Dörwald, W., Kastner-Maresch, A. and Lange, H. (1999) The role of visualization in

forest growth modelling. In: Amaro, A. and Tomé, M. (eds) Empirical and Process-based

Models for Forest Tree and Stand Growth Simulation. Edições Salamandra, Lisbon, pp.

403–418.

Kahn, M. and Pretzsch, H. (1997) Das Wuchsmodell SILVA: Parametrisierung der Version 2.1

für Rein- und Mischbestände aus Fichte und Buche. Allgemeine Forst- und Jagdzeitung 168

(6–7), 115–123.

Knauft, F.-J. (2000) Entwicklung von Methoden zur GIS-gestützten Visualisierung von

Waldentwicklungsszenarien. Dissertation thesis, Fakultät für Forstwissenschaften

und Waldökologie der Georg-August-Universität Göttingen. Available at:

webdoc.sub.gwdg.de/diss/2000/knauft

Knauft, F.-J. and Sloboda, B. (2000) Visualisierung virtueller Waldlandschaften durch

Integration individuenbasierter Modelle. In: Tagungsband Herbstkolloquium 1999.

Deutscher Verband Forstlicher Forschungsanstalten und Internationale Biometrische

Gesellschaft Deutsche Region, Ljubljana, pp. 293–308.

Kurth, W. (1994) Growth Grammar Interpreter GROGRA 2.4: a Software-Tool for the 3-Dimensional

Interpretation of Stochastic, Sensitive Growth Grammars in the Context of Plant Modelling:

Introduction and Reference Manual. Berichte des Forschungszentrum Waldökosysteme 38.

Göttingen.

Kurth, W. (1999) Die Simulation der Baumarchitektur mit Wachstumsgrammatiken.

Habilitationsschrift an der Georg-August-Universität Göttingen. Wissenschaftlicher

Verlag, Berlin, 327 pp.

Lindenmayer, A. (1968) Mathematical models for cellular interaction in development. Journal of

Theoretical Biology 54, 3–22.

McGaughey, R.J. (2000) EnVision – Environmental Visualization System. Homepage of the USDA

Forest Service Paciﬁc Coast Research Station. Available at: forsys.cfr.washington.edu/

envision.html

Nagel, J. (1997) BWIN: Programm zur Bestandesanalyse und Prognose. Handbuch

Niedersächsische Forstliche Versuchsanstalt Abteilung A, 41 pp.

de Reffye, P. and Blaise, F. (1993) Modélisation de l’architecture des Arbres. Applications

Forestières et Paysagères. Revue Forestière Française 45, 128–136.

34Amaro Forests - Chap 30 1/8/03 11:54 am Page 358

358 F.-J. Knauft

Schanz, H. (1994) Forstliche Nachhaltigkeit aus der Sicht von Forstleuten in der Bundesrepublik

Deutschland. Working paper 1994, Institut für Forsteinrichtung und Forstliche

Betriebslehre, Universität Freiburg.

Scheerer, B. (2000) Protokollierung von Durchforstungseingriffen in einem

Waldwachstumssimulator. Diploma thesis, University of Bayreuth, 113 pp.

Seifert, S. (1998) Dreidimensionale Visualisierung des Waldwachstums. Diploma thesis,

Fachbereich Informatik der Fachhochschule München, Munich.

Windhager, M. (1999) Evaluierung von vier verschiedenen Waldwachstumssimulatoren.

Dissertation thesis, Universität für Bodenkultur Wien, Vienna, 217 pp.

35Amaro Forests - Chap 31 25/7/03 11:09 am Page 359

31 How Good is Good Enough?

Information Quality Needs for

Management Decision Making

David D. Reed

and Elizabeth A. Jones

Abstract

There are many aspects to information quality, and everything from measurement method

documentation to data entry errors to metadata documentation falls under this topic. Even

with good data quality assurance procedures in place, uncertainty exists because of random

variation in the population and the sampling design utilized to collect the data. If data are

used as the basis for model projections, further uncertainty is introduced due to model predic-

tion errors. Adequacy of information quality can only be evaluated in the context for which it

is used. The focus of this chapter deals with the uncertainty in information and its impact on

decision making. The consequences of management decisions can only be assessed through

consideration of both the cost of obtaining more precise information as well as the cost of mak-

ing a poor decision.

Introduction

All resource management decisions are made in the face of uncertain and incom-

plete knowledge. Even if information regarding the current state of the system is

known with high precision, future conditions are uncertain, to at least some degree,

and this uncertainty typically increases with time. Managers must deal with these

numerous sources of uncertainty when making decisions. Furthermore, many good

managers have an intuitive understanding of their system of interest and the associ-

ated uncertainties, but they often cannot explicitly identify all of the sources of

uncertainty that affect their information base. It is, therefore, impossible to develop

perfect knowledge of current, much less future, conditions.

How, then, can managers utilize all of this imprecise information to improve

their decision-making pr

ocesses? How can managers act to minimize their total risk

exposure? In many ways, resource management comes down to risk management.

In trying to avoid an obvious source of risk due to uncertain information, a manager

may inadvertently be exposed to a greater level of risk from an unidentiﬁed source.

School of Forest Resources and Environmental Science, Michigan Technological University, USA

Correspondence to:ddreed@mtu.edu

Department of Mathematical Sciences, Michigan Technological University, USA

35Amaro Forests - Chap 31 25/7/03 11:09 am Page 360

360 D.D. Reed and E.A. Jones

The key to decision making is a full understanding of the risks encountered as a

result of imprecise information. Managers must balance the cost of obtaining more

precise information, where this is possible, with the cost of making a poor decision.

Decision Complexity

The simplest type of decision is a binary choice. Statisticians are most familiar with

this thr

ough hypothesis testing. There is a factual statement (e.g. µ

= 0) with the

necessary decision being whether or not to reject the veracity of this statement. The

usual types of uncertainty are known as Type I (α) and Type II (β) errors (Fig. 31.1).

Type I error (the probability of rejecting the statement when it is true) is controlled

by the analyst through statistical methodology. Type II error (the probability of not

rejecting the statement when it is false) is understood to be inversely related to Type

I error, but it is rarely explicitly considered in the decision-making process. The

complexities of this simple decision-making scenario are immense and are speciﬁc

to the context at hand – what is the appropriate statistical procedure? what are the

implicit and explicit assumptions required for the procedure to be appropriate?

Unfortunately, there are virtually no resource decisions that truly ﬁt this relatively

simple decision-making scenario.

A more complex decision-making scenario occurs when the choices are no

longer binary

, but there is a ﬁnite, identiﬁable set of choices. In this case, the deci-

sion does not consist of determining the veracity of a single statement, but in choos-

ing among a ﬁnite group of identiﬁed alternatives. There is risk associated with each

alternative, and the challenge is to identify the alternative that carries the minimum

total risk to the manager. Again, the complexities are enormous and there are few

resource decisions that ﬁt this decision-making scenario.

The most complex scenario involves a choice among a very large number of

alternatives, wher

e most or all of the alternatives do not have explicitly deﬁned

measures of risk. This is, unfortunately, the most common situation in resource man-

agement. Managers often simplify the decision process by eliminating a large num-

ber of possible alternatives, resulting in a situation resembling scenario 2 (e.g. a

decision between planting species A, B or C on a particular habitat type). While

ignoring many possible alternatives, the process is then simpliﬁed. There is an

implicit assumption that the alternatives considered in the reduced set carry less

Decision

Hypothesis

Statement Statement

True False

Statement Correct Type II Error

True

Statement Type I Error Correct

False

Fig. 31.1. Binary c

hoice decision scenario, similar to simple statistical hypothesis testing.

35Amaro Forests - Chap 31 25/7/03 11:09 am Page 361

361 Information Quality and Management Decision Making

total risk than do any of the eliminated alternatives. Decisions are often simpliﬁed

even further into a situation resembling the binary scenario (e.g. should this stand

be harvested during this management period or should it be left alone and evalu-

ated again at some time in the future?).

As illustrated in the following examples, the available information may be ade-

quate at a particular scale, such as a harvest scheduling application involving many

stands within an ownership, but inadequate at a dif

ferent scale, such as the choice of

a management regime for a single stand. Since the cost of obtaining information is

often substantial, it is imperative that the intended uses of the information and the

management decisions that the information will support be identiﬁed prior to

expending the resources to collect the information.

Inventory Example

As noted previously, it is not only important to consider the cost of improving the qual-

ity of information available to support a decision, but it is also important to consider

the cost of making a poor decision. Cochran (1977) presents the use of the cost-plus-

loss method in a sampling context, while Burkhart et al. (1978) present the use of this

method in a multi-resource sampling setting. In this context, the total expected loss

(T(X)) is a function of the cost of sampling to obtain a given precision (C(X)) plus the

expected loss resulting from imprecise knowledge of the quantity of interest (L(X)):

T(X) = C(X) + L

(X) (1)

In a simple application, C(X) is a linear function of the number of samples taken,

(X) is a squared error loss function that accounts for both the sampling error and

the value of the resource in question, and X is the decision variable, which, in this

case, is the number of samples to install in the population of interest.

The question of sample size determination involves a balancing of the cost of

installing n samples with the expected loss due to the uncertainty associated with a

sample of size

n. Burkhart et al. (1978) extend this to the situation of multiple inven-

tories with multiple subsequent management decisions. The method can be

extended to even more complex situations; Reed et al. (1984), for example, use the

cost-plus-loss method to examine the question of whether or not to sample in order

to adjust individual tree volume equations for application to a speciﬁc stand. Ståhl et

al. (1994) used the cost-plus-loss framework to examine the timing and precision of

inventory activities, explicitly recognizing that the loss due to imprecision increases

with the time since the inventory, and that the cost of the inventory is strongly

related to the resulting precision of the estimates.

The important point here is that decision makers must function as risk man-

agers, not just risk avoiders. The explicit consideration of the loss function for

ces

managers to consider not just the costs of implementing a particular option, but also

both the risks of poor performance associated with an option and the opportunity

costs associated with foregoing the same option. Good managers implicitly incorpo-

rate these considerations into their decision-making process, but forcing their explicit

consideration can lead to signiﬁcant insights even for experienced managers.

Harvest Scheduling Example

The simplest cases of the prior example involve a decision regarding the inventory

method or timing for a single stand. A second example is a harvest scheduling

35Amaro Forests - Chap 31 25/7/03 11:09 am Page 362

362 D.D. Reed and E.A. Jones

application that combines inventory results from many stands. The goal is to

Proportion of maximum

develop a schedule of management activities, such as harvests, for a number of dif-

ferent stands over several management periods. In this particular example, an

inventory of 200 quaking aspen (Populus tremuloides) stands forms the resource base.

Inventory information for these stands was generated from a sample of the US

Forest Service Forest Inventory and Analysis data from the state of Minnesota as

retrieved from the Eastwide database (Hanson et al., 1992). A stand size is assigned

to each stand from an independent inventory of stands. In other words, the data

used in this example are a composite from several sources in the north-central

region of the USA and are not derived from any particular area or ownership. From

the inventory data, annual volumes were estimated using the

FVS,LS (Bush and

Brand, 1995) stand simulator. In this example, each management period is 1 year in

length, only clearcut harvesting is considered, and the goal is to obtain a schedule of

stands to harvest each year over a 30-year planning horizon.

The

HABPLAN (Van Deusen, 1999) harvest scheduling program was used. In con-

trast to harvest scheduling optimization programs that use linear programming

techniques to obtain a single, optimum schedule that maximizes a given objective

function (total volume production in this example),

HABPLAN uses the Metropolis

algorithm to iteratively evaluate schedules and identify those that produce the

largest harvested volumes.

HABPLAN was used to iteratively evaluate harvest timing

for each stand in an attempt to identify the combination of stand activities that

resulted in the largest harvested volume over the 30-year planning horizon.

Schedules were evaluated over 10,000 iterations.

Figure 31.2 shows the 200 (of 10,000) schedules produced by

HABPLAN for the

example data that have the largest total volumes over the planning horizon. The

total volume harvested over the 30-year period is expressed as a proportion of the

maximum produced over the 10,000 iterations. Over the 200 iterations shown here,

there was a difference of 0.02% between the volume production of the maximum

and minimum schedules. There are, in essence, a very large number of schedules

that produce essentially equivalent volume production over the planning horizon.

While there are certainly schedules that produce far less volume than those identi-

ﬁed here, there are many alternative schedules that produce essentially equivalent

results.

1.0000

0.9995

0.9990

0.9985

0.9980

21 41 61 81 101 121 141 161 181

Iteration

Fig. 31.2. Total volume production over the 30-year management horizon for 200 HABPLAN iterative

solutions, expressed as a proportion of the largest value.

35Amaro Forests - Chap 31 25/7/03 11:09 am Page 363

363 Information Quality and Management Decision Making

What happens if the original inventory data contain sampling errors? To inves-

tigate this, three scenarios were examined. Using Monte Carlo techniques, the initial

inventory for each stand was randomly varied by ±5%, ±10% and ±20% (assuming a

uniform error distribution). Each stand was assigned a random error factor and this

was applied to all projected volumes for the stand as well as to the initial starting

volume. The resulting harvest schedules did not differ (P < 0.05) in total volume

production from those obtained from the original data. In other words, random

symmetrical errors applied to individual stands had virtually no impact on the esti-

mated total volume harvested from the harvest scheduling exercise.

Discussion

Resource managers are concerned with very complex systems, where they are pre-

sented with an essentially inﬁnite variety of choices. It is important that managers

do not just consider the cost associated with improving the quality of information

available to support decision making, but that they also understand and consider

the risk associated with making a poor decision. Avoiding the relatively easy-to-

quantify costs of obtaining higher quality information may result in greater total

risk due to the loss of possible gains from management options which are either not

explicitly considered or are rejected on the basis of the poorer quality information

which is available.

At the stand level, this is manifest in the cost of obtaining more precise volume

estimates as opposed to the potential gains that might be r

eceived if better quality

information led to the consideration or implementation of different management

regimes. In a ﬁnancial analogy, there is relatively little risk of loss to investments in

ﬁxed-interest-bearing instruments, but there is considerable, and often poorly quan-

tiﬁed, risk that such ‘safe’ investments may really lead to an erosion of value in an

inﬂationary environment. The additional information necessary to evaluate this

requires some cost to obtain, but it may lead to reduced total risk by allowing more

adequate consideration of both the cost and the potential loss associated with a par-

ticular investment decision.

In the harvest scheduling example, the examination of stand-level errors in vol-

ume estimates pr

ovides some important insights. If the individual stand estimates

are unbiased (i.e. symmetrical about the estimate with an expected value of the error

equal to zero), then over the planning horizon, there is little impact of precision in

the individual stand estimates on the ﬁnal estimated total volume production. The

presence of bias in the estimates (reﬂected through non-symmetrical errors about

the estimate with an expected value of the error not equal to zero) could impact the

ﬁnal estimated total volume production over the planning horizon. While error in

the stand-level volume estimates is important at the scale of selecting a management

regime for the individual stand and can result in suboptimal harvest schedules for

individual stands, it has little impact on the larger scale question of selecting a har-

vest schedule for an area consisting of many stands. Hierarchical considerations

may be important in understanding the impact of imprecise information in decision

making, analogous to their impact on physiological processes, where variables and

timesteps affecting individual cell or organ processes are quite different from those

affecting populations or ecosystems.

The distribution of the uncertainties can also affect their impact on decision

making. It is logical to consider sampling err

ors as being symmetrical about the

mean. Model projection errors, however, are often not symmetrical since many

resource models are logarithmic. If the usual regression error assumption is met

35Amaro Forests - Chap 31 25/7/03 11:09 am Page 364

364 D.D. Reed and E.A. Jones

using the logarithmically transformed developmental data (ε~N(0, σ

)), then con-

version of the predicted values back to arithmetic units results in a skewed, non-

symmetrical distribution of errors in the ﬁnal predicted values, with the modal

arithmetic predicted value being less than the mean value (Finney, 1941; Baskerville,

1972). In this case, it is logical to expect a greater impact on the total harvested vol-

ume estimated from the harvest scheduling exercise. A harvest scheduling algo-

rithm may tend to preferentially select stands and treatment schedules for which the

predicted volumes are overestimates. While symmetrical individual stand errors did

not impact the ﬁnal total volume production over the management period, asym-

metrical errors could have an impact on the estimated total volume production over

the 30-year period.

In most forest resource situations, inventory precision goals are usually

expr

essed either as percentages or as units per unit area. In a total volume or value

context, an error of ±5% in a large-area stand has a much greater impact than an error

of ±5% in a smaller stand. When individual stands are combined, as they are in the

harvest scheduling example discussed above, estimates for large-area stands may

have a proportionally greater impact on a quantity such as total volume production

over a 30-year period than estimates of similar precision for a small-area stand, even

though both are unbiased. In the harvest scheduling example, it might make sense to

have greater precision requirements for large-area stands than small-area stands. This

would obviously need to be balanced with cost, but many simulation studies have

assumed equal stand sizes, and many managers may not fully appreciate the advan-

tages that might be received in a harvest scheduling situation by relaxing precision

requirements for small stands while maintaining or increasing precision goals in

larger area stands. In other words, the appropriate quantity to use to estimate the

needed sample size may not be the precision expressed as a percentage of the

mean/ha, but in many situations might be the precision of total volume or value

expressed in absolute numerical units of total volume or value at the stand level.

Summary

Natural resource managers are concerned with very complex decision-making sce-

narios; there are often an essentially inﬁnite variety of options from which to select,

and precision of available information concerning the risk associated with each

option is difﬁcult to assess. It is important for managers to consider and balance

both the cost of obtaining more precise information and the cost of making a poor

decision. These costs vary by situation, and often precision needs differ for the same

piece of information when it is used to support different decisions. In a harvest

scheduling context, the sampling errors associated with estimates of stand volumes

have little impact on the ﬁnal projected harvest volume provided that the estimates

are unbiased, while error associated with stand-level inventories can have large

impacts on the consequences of individual stand management decisions. Managers,

therefore, must also consider the context in which information will be used to sup-

port decision making in order to judge the adequacy of the available information for

supporting the decision at hand.

References

Baskerville, G.L. (1972) Use of logarithmic regression in the estimation of plant biomass.

Canadian Journal of Forest Research 2, 49–53.

35Amaro Forests - Chap 31 25/7/03 11:09 am Page 365

365 Information Quality and Management Decision Making

Burkhart, H.E., Stuck, R.D., Leuschner, W.A. and Reynolds, M.R. (1978) Allocating inventory

esources for multiple-use planning. Canadian Journal of Forest Research 8, 100–110.

Bush, R.R. and Brand, G.J. (1995) The Lake States TWIGS variant of the Forest Vegetation

Simulator. US Department of Agriculture, Forest Service, North Central Forest

Experiment Station, St Paul, Minnesota.

Cochran, W.G. (1977) Sampling Techniques, 3rd edn. John Wiley & Sons, New York, 330 pp.

Finney, D.I. (1941) On the distribution of a variate whose logarithm is normally distributed.

Journal of the Royal Statistical Society, Series B 7, 155–161.

Hansen, M.H., Frieswyk, T., Glover, J.F. and Kelly, J.F. (1992) The Eastwide Forest Inventory Data

Base: Users Manual. General Technical Report NC-151, US Department of Agriculture,

Forest Service, North Central Forest Experiment Station, St Paul, Minnesota, 48 pp.

Reed, D.D., Jones, E.A. and Green, E.J. (1984) Sampling to adjust individual tree volume equa-

tions for application to a speciﬁc stand. In: La Bau, V.J. (ed.) Proceedings of the International

Conference on Inventorying Forest and Other Vegetation of the High Latitude and High Altitude

Regions, pp. 289–292.

Ståhl, G., Carlsson, D. and Bordensson, L. (1994) A method to determine optimal stand data

acquisition policies. Forest Science 40, 630–649.

Van Deusen, P.C. (1999) Multiple solution harvest scheduling. Silva Fennica 33, 207–216.