Modeling individual tree mortality for Austrian forest species
Introduction
Recent statistics from the Austrian National Forest Inventory indicate that at least 40% of forest stands have a mixed-species composition (Schieler and Schadauer, 1993). Furthermore, only 45% of the inventory plots were sufficiently even-aged to allow usual site index determination (Monserud and Sterba, 1996). This lack of a stand age and site index on half of Austria's forests renders existing yield tables increasingly unreliable, and provided the impetus for the development of the individual-tree stand growth model PROGNAUS (Sterba et al., 1995; Monserud et al., 1997). To date, models have been developed for predicting basal area increment (Monserud and Sterba, 1996), height increment (Hasenauer and Monserud, 1997), crown ratio (Hasenauer and Monserud, 1996), natural regeneration (Schweiger and Sterba, 1997; Sterba et al., 1997), and harvesting (Monserud et al., 1997). In addition, the model has successfully passed a validation test (Sterba and Monserud, 1997), and has been the focus of an investigation on simultaneous equation systems (Hasenauer et al., 1998).
The remaining key component is an accurate mortality model (Hamilton, 1990; Avila and Burkhart, 1992). Accordingly, our objective is to develop individual tree mortality models for the major forest species of Austria. We are looking for mortality models that have generality in two senses. First, we want to develop mortality models that are representative of all forest conditions in Austria, and for all major species. Second, we want to develop models that can be used not only in the PROGNAUS simulator, but in any stand simulator that uses a list of individual trees by species.
Error propagation and budgeting analyses have shown that growth predictions are very sensitive to the underlying mortality model; furthermore, the contribution to total variability due to the mortality component increases as the projection period increases (Gertner, 1989). Guan and Gertner (1991a)considered this situation common in stand simulation modeling.
Mortality remains one of the least understood components of growth and yield estimation (Hamilton, 1986). The key to a tree's survival is its genetic makeup and its environment (Spurr and Barnes, 1980). Growth modelers almost universally ignore a tree's genetic status (Monserud and Rehfeldt, 1990), as well as important environmental factors such as climatic extremes (e.g., wind, drought, killing frosts), insects, and diseases. Great detail is paid to environmental competition arising from neighboring trees, however (Buchman et al., 1983), as well as measurable gross physical features of the tree and site. Perhaps mortality would appear less stochastic if relevant environmental variables were measured on permanent plots, and if the genetic status of the trees could be characterized.
The literature on modeling mortality of forest trees is not small, but successes are rare. Realistically, mortality modelers mostly hope to capture the average rate of mortality, and relate it to a few reliable and measurable size or site characteristics. The key then is a large and representative sample of remeasured trees so that a rare event – mortality – can be observed frequently enough to predict it accurately.
A representative sample must reflect both the full range in site variability as well as the diversity of management treatments in a given population (Hamilton, 1980). Because mortality data are most reliably and efficiently obtained from permanent plots, researchers are often forced to live with the limitations of the underlying permanent plot network. As a result, many studies rely on data from either unthinned plots (e.g., Zhang et al., 1997) or only lightly thinned plots (Dursky, 1997). Even if the data contain various treatments, the permanent plots often are clustered spatially and do not necessarily represent the average and dispersion of stands in the entire region of interest (e.g., Hasenauer, 1994).
The most common methodology for modeling individual tree mortality is statistical. Generally, the parameters of a flexible non-linear function bounded by 0 and 1 are estimated using weighted nonlinear regression or a multivariate maximum likelihood procedure (Neter and Maynes, 1970). Although most cumulative distribution functions will work, the most popular is the logistic or logit (e.g., Monserud, 1976; Hamilton and Edwards, 1976; Buchman, 1979; Hamilton, 1986; Vanclay, 1995). Other applications have used the Weibull (Somers et al., 1980), the gamma (Kobe and Coates, 1997), the Richard's function (Buford and Hafley, 1985), the exponential (Moser, 1972), and the normal or probit (Finney, 1971). Monserud (1976)found that both the probit and the logit produced similar results, even though the underlying functional forms are quite different. Vanclay's 1995 recommendation of the logistic for mortality models in tropical forests is surprisingly quite relevant to our Austrian temperate forest situation, for both cover a spectrum of different species mixtures and age structures, precluding the possibility of using either stand age or site index as predictor variables.
Two additional procedures have been used to model individual tree mortality: recursive partitioning and neural networks. Unfortunately, neither has led to significant improvement in our ability to predict mortality using classical statistical methods. Recursive partitioning is best known by the acronym CART (Classification And Regression Trees; Breiman et al., 1984), although it is a similar classification tool to the SCREEN algorithm used by Hamilton and Wendt (1975)for efficiently identifying potential independent variables and relationships. CART employs heuristic methods with binary classification, with results presented as decision trees (Verbyla, 1987). Dobbertin and Biging (1998)compared CART to logistic regression for two species in northern California, and concluded that CART performed somewhat better, although the percentage of dead trees correctly classified was very low. Working in a branch of artificial intelligence, Guan and Gertner, 1991a, Guan and Gertner, 1991bbuilt a neural network mortality model that was as accurate as the corresponding logistic model with similar variables, and was better behaved because the same model form was not constrained to operate in all regions of the data space. Hasenauer and Merkl (1997)have recently compared a neural network to the logit for Austria, and found that both predict equally well, with a slight advantage to the neural network.
Section snippets
Methods
We rely on the remeasured permanent plots of the Austrian National Forest Inventory for mortality and survival data (Forstliche Bundesversuchsanstalt, 1981, Forstliche Bundesversuchsanstalt, 1986). A systematic 3.89 km grid of permanent plots covering Austria was established in 1981–1985. Each year 20% of the grid locations are sampled (1100 clusters) in such a way that all of Austria is covered by the inventory each year. Each location was then remeasured from 1986–1990, exactly 5 years after
Analysis
Individual tree mortality is a discrete event. A datum can have only the value 0 (live) or 1 (dead). A dichotomous dependent variable calls for special consideration not only in parameter estimation (Hamilton, 1974), but in the interpretation of goodness of fit as well (Neter and Maynes, 1970). We take the classical approach to model the probability of mortality, the logistic equation:where b′X is a linear combination of parameters b and independent variables X, and e is the base of
Results
For all species, we found the hyperbolic D−1 to be highly significant in predicting mortality rate (Table 2). Furthermore, it behaved properly in matching the rapid decline from high mortality for the smallest diameters to a more gradual decline in mortality rate for the larger diameters (Fig. 1). It was clearly superior to a simple D term.
Crown ratio CR was also consistently significant, except for oak (Table 2). We likewise found BAL to be a highly significant predictor of mortality rate for
Application
Although our mortality model was developed as a submodel in the distance-independent stand simulator PROGNAUS (Sterba et al., 1995; Monserud et al., 1997), it should work equally well in a simulator with a different architecture (e.g., Sterba, 1983; Eckmüllner and Fleck, 1989; Pretzsch, 1992). The key ingredient for obtaining a good mortality model is a large data set that is representative of the population under study. The Austrian National Forest Inventory data (Forstliche
Acknowledgements
This research was conducted when Monserud was Visiting Scientist and Universitätslektor at the Institut für Waldwachstumsforschung in Vienna, on a grant from the Austrian Ministry of Agriculture and Forestry. Karl Schieler and Klemens Schadauer of the Federal Forest Research Center (Forstliche Bundesversuchsanstalt, Wien) generously made the Austrian Forest Inventory data available. We thank Dave Hamilton, Ralph Amateis, Karl Schieler, and the anonymous reviewers for helpful review comments.
References (66)
- et al.
A crown ratio model for Austrian forests
For. Ecol. Manage.
(1996) - et al.
Biased statistics and predictions for tree height increment models developed from smoothed `data'
Ecol. Modelling
(1997) - et al.
Evidence from nitrogen fertilization in forests of Germany
Environ. Pollut.
(1988) - et al.
Needle, crown, stem, and root phytomass of Pinus sylvestris stands in Russia
For. Ecol. Manage.
(1996) - et al.
A basal area increment model for even- and uneven-aged forest stands in Austria
For. Ecol. Manage.
(1996) - et al.
Natural regeneration establishment of Norway spruce (Picea abies L.) in Austria
For. Ecol. Manage.
(1997) - et al.
Excess nitrogen deposition: issues for consideration
Environ. Poll.
(1988) - et al.
Applicability of the forest stand growth simulator PROGNAUS for the Austrian part of the Bohemian Massif
Ecol. Modelling
(1997) Beiträge zum Baumwachstum – der h/d Wert
Forst und Holzwirt.
(1976)- et al.
Modeling survival of loblolly pine trees in thinned and unthinned plantations
Can. J. For. Res.
(1992)
Die Winkelzählprobe. Allg
Forst-und Holzwirt. Zeitung Wien.
Yield differences between research plots and managed forests
J. For.
A tree survival model with application to species of the Great Lakes region
Can J. For. Res.
Modeling the probability of individual tree mortality
For. Sci.
A logistic model of mortality in thinned and unthinned mixed conifer stands of northern Idaho
For. Sci.
Extending the range of applicability of an individual tree mortality model
Can. J. For. Res.
Tree life tables
Bioscience
Cited by (319)
Recruitment models after reduced impact logging in the Amazon rainforest
2023, Forest Ecology and ManagementAdapting the tree growth model MOSES to manage uneven-aged mixed species forests in Bosnia & Herzegovina
2023, Trees, Forests and PeopleMachine Learning Forest Simulator (MLFS): R package for data-driven assessment of the future state of forests
2023, Ecological InformaticsNorway spruce (Picea abies (L.) H. Karst.) selection forests at Siljansfors in Central Sweden
2023, Trees, Forests and People
- 1
Fax: +43-1-47654-4242; e-mail: [email protected]