Performance of Percentile Based Diameter Distribution Prediction and Weibull Method in Independent Data Sets

Diameter distribution is used in most forest management planning packages for predicting stand volume, timber volume and stand growth. The prediction of diameter distribution can be based on parametric distribution functions, distribution-free parametric prediction methods or purely non-parametric methods. In the fi rst case, the distribution is obtained by predicting the parameters of some probability density function. In a distribution-free percentile method, the diameters at certain percentiles of the distribution are predicted with models. In non-parametric methods, the predicted distribution is a linear combination of similar measured stands. In this study, the percentile based diameter distribution is compared to the results obtained with the Weibull method in four independent data sets. In the case of Scots pine, the other methods are also compared to k-nearest neighbour method. The comparison was made with respect to the accuracy of predicted stand volume, saw timber volume and number of stems. The predicted percentile and Weibull distributions were calibrated using number of stems measured from the stand. The information of minimum and maximum diameters were also used, for re-scaling the percentile based distribution or for parameter recovery of Weibull parameters. The accuracy of the predicted stand characteristics were also compared for calibrated distributions. The most reliable results were obtained using the percentile method with the model set including number of stems as a predictor. Calibration improved the results in most cases. However, using the minimum and maximum diameters for parameter recovery proved to be ineffi cient.


Introduction
Different theoretical distributions, for example beta, Weibull and Johnson's SB functions, have been widely used to describe tree stock mathematically (e.g.Loetsch et al. 1973, Bailey and Dell 1973, Hafl ey and Schreuder 1977).In Finland, researchers have mainly applied the Weibull distribution (e.g.Kilkki and Päivinen 1986, Kilkki et al. 1989, Maltamo 1997).Two main methods have been used to predict a parametric distribution function, namely the parameter prediction method and the parameter recovery method (Hyink 1980, Hyink andMoser 1983).The characteristics considered in parameter recovery can be either moments or percentiles of diameter distribution (see Knoebel and Burkhart 1991).It is also possible that some of the parameters are predicted and others are solved using a parameter recovery approach (Burkhart et al. 1982, Maltamo 1998).
Distribution-free methods do not rely on any predefi ned functional form.For example, Borders et al. (1987) developed a percentile based diameter distribution prediction method.In this method, the diameter distribution is obtained by fi rst predicting diameters at certain percentiles and then interpolating the distribution function between these percentiles.Maltamo et al. (2000) used the percentile based approach to predict irregular stem frequency diameter distributions of stands in a natural state.Kangas and Maltamo (2000b) presented two sets of models for each species in Finland, to predict 12 percentiles of the basal area diameter distribution.
Another possibility to describe diameter distribution is to rely purely on non-parametric methods.For example, Haara et al. (1997) and Maltamo and Kangas (1998) predicted the basal area diameter distribution using nearest neighbour based approaches.The predicted diameter distribution was a weighted average of the distributions of the sample plots that are most similar to the stand of interest.These plots were selected from a database of previously measured stands.
In most applications, the predicted basal area diameter distribution is scaled to the measured basal area.However, there is no guarantee that the other stand characteristics obtained from the predicted diameter distribution correspond to the measured stand characteristics (except for characteristics that are used for parameter recovery of theoretical distributions or measured percentiles used in the percentile method).Kangas and Maltamo (2000a) calibrated stem frequency diameter distribution with basal area using an approach adopted from sampling theory, calibration estimation, originally presented by Deville and Särndal (1992).In their study, both number of stems and basal area were known but the predicted distribution did not give correct estimates for both of these stand characteristics.With calibration, correct estimates for all known stand characteristics were obtained.
In this study, the performance of the percentile based method is compared to the parameter prediction of the Weibull distribution, and, in the case of Scots pine, to nearest neighbour based method.The models are thoroughly tested in varying conditions, with available independent data sets.The calibration estimation method is applied to basal area diameter distribution when additional information is assumed to be available.

Test Materials
The models were tested with several independently measured data sets.These include anglecount sample plot data from southern Finland, fi xed area sample plot data from mineral soils across Finland, peatland stands and different kinds of mixed stands.The purpose was to examine the accuracy of the models in different situations.These data sets also include stands (mixed stands and stands in northern Finland) which have been found to be especially problematic for current diameter distribution models (e.g.Siipileh to 1999).Independent data sets were restricted standwise according to the following rules: the basal area of the considered tree species had to be over 1 m 2 /ha and the number of stems over 50 per hectare; the number of measured trees in the sample plots had to be at least 10; the basal area median diameter had to be over 5 cm; and the range between both minimum and median and maximum and median diameter had to be over 2 cm.
A private forest enterprise collected the fi rst test data (see e.g.Suutarla 1985).Six to twelve relascope (angle-count) sample plots were systematically located in each stand.Diameter at breast height (dbh) was recorded in 1 cm classes from each tree included in the sample plots using the basal area factor two (m 2 /ha).The basal area diameter distributions were formed by summing the trees in the relascope sample plots in the stand.This data set is later referred to as ANGLE.
The other test data sets include the permanent sample plots (INKA and SINKA) measured by the Finnish Forest Research Institute (FFRI), originally for growth modelling purposes (Gustavsen et al. 1988, Penttilä andHonkanen 1986).The INKA sample plots were established on mineral soils across Finland.The SINKA sample plots are mainly located on ditched peatland (later called DITCHED) in northern Finland.A small data set including unmanaged peatland stands was also considered (later called UNMANAGED).The data sets include sample plots which consist of a cluster of three circular plots within a stand.When testing the diameter distribution prediction methods, these circular plots were combined.The diameter of all trees within a plot were measured to the nearest 0.1 cm.In the case of the INKA material, the test data was also divided by geographical area.Finally, the models were tested in mixed stands.These data sets include pine-spruce sample plots (MIXED 1), pine-birch sample plots (MIXED 2) and spruce-birch sample plots (MIXED 3).The mixed pine-spruce stand data was originally collected for studying the productivity of mixed forests stands (Pukkala et al. 1994).The sample plots from mixed pine-birch and spruce-birch stands were originally from the growth and yield studies of Mielikäinen (1980Mielikäinen ( , 1985)).These stands are located in southern and eastern Finland.In the case of pine-birch stands three circular plots were located within a stand.In the current study these plots were combined.In the case of sprucebirch stands one circular plot, and in the case of pine-spruce stands one rectangular sample plot was measured from each stand.In this data diameter at breast height was measured in 1 mm classes from all trees.
A summary of the sample plot type of different data sets, number of observations and the mean values of the most important stand characteristics are presented in Table 1.

The Compared Methods
In the percentile method, the diameters at certain percentiles of the distribution function are predicted with models.These diameters characterise an empirical distribution function.In this study, the percentiles are defi ned with respect to basal area in a stand.By interpolating between the predicted diameters, a basal area diameter distribution function is obtained.From this distribution, the basal area and number of stems in desired diameter classes can be calculated.Interpolation is done using Späth's rational spline, in order to obtain a monotone distribution (see Kangas and Maltamo 2000b for details).The models used to predict the diameters at different percentiles are those estimated by Kangas and Maltamo (2000b).
They estimated two model sets: in the fi rst set the number of stems was not included as a predictor (Percentiles 1) and in the second set it was (Percentiles 2).
The percentile based basal area diameter distribution models were compared to the Weibull distribution parameter prediction models presented by Maltamo (1997).These models have been estimated separately for Scots pine and Norway spruce from the same empirical data which was also used for estimating the models for diameters at different percentiles.The parameter models of Scots pine were used for birch species.Stand characteristics measured by tree species (e.g.basal area and basal area median diameter) were used as predictors in these parameter models.In the case of ditched peatland also the Weibullmodels presented by Hökkä et al. (1991) were tested for Scots pine and birch distributions.These models have been estimated for Scots pine and birch dominated stands earlier using the same peatland data (SINKA).The nearest neighbour based approach for predicting diameter distributions presented by Maltamo and Kangas (1998) was applied to the Scots pine stands of the INKA material, since there were no distance functions available for other tree species.The K-nearest neighbour method is based on distance-weighted nearest neighbour estimation, where k most similar stands are used for predicting the diameter distribution of the target stand.Before the nearest neighbour method can be applied, 1) the form of distance function to be used to fi nd the most similar reference stands; 2) the number of nearest neighbours to be used; and 3) the form of the weight function for weighting the reference stands must be established.In this study the results of the study by Maltamo and Kangas (1998) were applied.The used form of the distance function was based on absolute differences between stand characteristics.Basal area median diameter and median height of the stand were the stand characteristics used when searching for nearest neighbours.The coeffi cients of these variables were 5 for basal area median diameter and 1 for median height.The number of nearest neighbours used was 14.The weighting parameter of the chosen nearest neighbours was 4.8.
In nearest neighbour approach, INKA-stands were used as target stands, for which diameter distributions are calculated.The INKA stands were chosen as target stands in order to get a large scale alternative, which could then be compared with other diameter distribution prediction methods.Both INKA and ANGLE data were used as databases from which the reference stands were chosen.When the stands of the INKA data were used both as target and as reference stands, the target stand was excluded from the reference stands.Using the ANGLE data as a reference database can be justifi ed by the idea that there is common inventory data which can be used in many applications.Using the INKA material as a reference database can be justifi ed by the idea that one should search for the most similar neighbours from stands which correspond to the target stands as closely as possible.
In the application stage, the estimate of the relative basal area in each diameter class [d 1 ,d 2 ] was calculated from the cumulative distribution of diameters F as F(d 2 )-F(d 1 ).For the Weibull method, the cumulative distribution is obtained analytically.For the percentile method, the value of the empirical distribution F was obtained by interpolating the percentiles as a function of the predicted diameters with Späth's rational spline interpolation (Späth 1974, Lether 1984, see Maltamo et al. 2000).In the case of the k-nearest neighbour method, the original diameter classes of the chosen reference stands were used.In each case, the relative basal areas were scaled to the measured basal area in the stand to obtain an absolute value of basal area b k in each diameter class k.Finally, the frequency f k in each diameter class k was calculated from the class basal area by dividing it with the basal area of the mean tree in this class: Volumes for each diameter class were calculated with Laasasenaho's models (1982), using diameter at breast height as a predictor.Saw timber volume was defi ned as the volume of trees larger than or equal to 16.5 cm.

Calibrating the Predicted Basal Area Diameter Distribution
The predicted basal area diameter distribution was calibrated with an approach presented by Deville and Särndal (1992).Kangas and Maltamo (2000a) used this approach to calibrate the predicted frequencies of a diameter distribution.In the present study, the calibration estimator was used to modify the predicted basal area b k of each diameter class k.The modifi cation was carried out so that the modifi ed class basal areas w k are as close as possible to the predicted basal areas b k , while respecting the calibration equation(s).The calibration equation for number of stems is where N is the number of stems per hectare, g k is the basal area of the mean tree in diameter class k (m 2 /ha) and K is the number of 1-cm diameter classes.
Further, it was required that the basal area, which was used in scaling the relative basal area, remained correct also after calibration, by using a constraint where G is the stand basal area (m 2 /ha).In addition, the basal area below the midpoint of d gM diameter class was set to half of the total basal area as where The distance measure used was the logarithmic distance (Deville and Särndal 1992) Minimising this distance measure (7) while respecting the calibration equation(s) (3, 4, and 5) is a constrained non-linear optimisation problem.
The problem was solved by reformulating it using Lagrange multipliers (see Deville and Särndal 1992 for details).The resulting group of nonlinear equations was then solved using IMSL subroutines.
If the minimum and maximum diameters were assumed to be known, the distribution obtained with percentile method was re-scaled to the correct interval.The minimum (maximum) was set to the observed value, and the other diameters between minimum and mean diameter (mean and maximum diameter) were scaled according to where d min is the observed minimum diameter, d0 is the predicted minimum diameter, di is the predicted diameter at ith percentile and ˆ* d i is the re-scaled diameter.In the case of Weibull distribution, the minimum and maximum diameters were utilised using parameter recovery approach.The calibration was performed for ANGLE, INKA and DITCHED datasets.

Performance of the Methods
The performance of the different approaches (percentile, Weibull and k-nearest neighbour; both calibrated and uncalibrated), was examined by calculating the root mean square errors and biases of stand volume estimates (m 3 /ha) obtained with these methods.The absolute root mean square error (RMSE) was calculated as where n is the number of sample stands, V i is the true volume of stand i and Vi is the volume of stand i estimated from the predicted distribution.The relative RMSE of the volume estimate was calculated by dividing the absolute RMSE by the true mean volume V of the stands.The bias of the predictions was calculated as bias In addition to stand volume, the saw timber volume and number of stems were considered.The saw timber volume was not calculated for data sets including mostly pulpwood trees (DITCHED, UNMANAGED and Oulu region of INKA birches).Finally, it was calculated what proportion of the 25th and 75th percentiles had been correctly predicted into their diameter class in INKA dataset.This was done to examine the ability of different methods to predict stand structure characteristics.The 25th and 75th percentiles were chosen because they are not directly predicted when using the percentile based method.

Uncalibrated Test Results
Some examples of predicted distributions of different tree species and models are presented in Figs.1-3.All examples are from the INKA data set.In the case of Scots pine (Figs.1a-f) the fi t of the percentile based method which uses number of stems as a predictor (Percentiles 2) is good and it can also reproduce stand structure quite precisely.Both the Weibull method and especially the percentile based method (Percentiles 1), neither of which uses number of stems as a predictor, produce overestimates of small trees.The situation is quite similar in the case of Norway spruce   (Figs.2a-f).The fi t of the Percentiles 2 is good.
The Percentiles 1 method overestimates small trees and the Weibull method cannot describe the shape of the distribution properly.Finally, in the case of birch species (Figs.3a-f) the fi t of the Percentiles 2 is not as good as for other tree species.The situation is same for the Percentiles 1 method but the fi t of the Weibull function is good.
The relative root mean square errors (RMSE) and the absolute biases of stand volume (m 3 /ha), stand timber volume (m 3 /ha) and number of stems (per hectare) in four independent test data sets are presented in Tables 2, 3 and 4. On average, the results for Scots pine were the most reliable, and the results for birch were the least reliable.With all species, the most reliable results were obtained with the Percentiles 2 method.When compared to the Percentiles 1 method, the Percentiles 2 method produced on average 23% smaller RMSE The percentile based method not using number of stems as a predictor and the Weibull method produced nearly equally reliable results for pine and spruce.For birch, however, the results obtained with Weibull method were clearly more accurate than those obtained with Percentile 1 method.It is worth noting that the Weibull results for birch were calculated using models estimated for pine.The number of stems estimates were in many cases badly biased if number of stems was not used as a predictor.
The performance of k-nearest neighbour estimates with the Scots pines in the INKA material was close to that of the other methods not using number of stems as a predictor (Table 2).When the INKA stands were used as reference stands (K-nn 1), the accuracy of the predicted stand characteristics was slightly better than with the Percentiles 1 method and the Weibull method.When ANGLE stands were used as reference stands (K-nn 2), the accuracy of volume estimates was worse than with other methods.For saw timber volume and number of stems the results were quite similar to the Percentiles 1 and Weibull methods.
The results obtained using peatland data, SINKA, were generally poorer than the results obtained in mineral soils.This is especially true for the unmanaged peatland and for birch.However, the poorer results can partly be explained by the fact that the SINKA sample plots were measured from more northern areas than the modelling data.Weibull models based on the DITCHED material (Weibull2, Hökkä et al. 1991) produced only slightly more accurate results than general Weibull models.It is also notable that with respect to peatland the percentile based method which used number of stems as a predictor produced in some cases worse results than the one not using number of stems.This was especially true for birch species.
The effect of the geographical areas on the RMSE of the different forest variables is presented in Table 5 using the INKA data: the results were clearly poorer in the northern areas.Although the Percentiles 1 method and the Weibull method produced generally nearly equally reliable results, it seems that in extreme conditions, such as on peatland which has not been ditched (UNMANAGED) and most of northern Finland (INKA in Lapland), the results are usually more accurate with the percentile based method.
For the INKA data set, the proportion of 25th and 75th percentiles that were predicted to be in the correct diameter class are presented in Table 6.The results are presented only for INKA as the differences between the data sets were not great.However, there were notable differences between the model sets and different percentiles.The Percentiles 2 method produced the percentiles considered most accurately.The other methods produced better results only for the 75th percentile of basal area in the case of spruce and for both basal area percentiles in the case of birch.The 75th percentile with respect to number of stems and the 25th percentile with respect to basal area could be predicted correctly more often than the other percentiles.This can be explained by the fact that a small deviation from the true distribution in the smallest diameter classes can change the number of stems considerably, but not the basal area.On the other hand, a deviation from the true distribution in the largest diameter classes may affect the basal area considerably, but may have a negligible effect on the number of stems.One reason for the good results in predicting the 75th percentile of number of stems is also that this diameter is usually very close to the basal area median diameter, which was assumed to have been assessed and thus known.
In the mixed data sets (Table 7) the results for Scots pine were very good with all methods.However, the results for Norway spruce and birch were generally much poorer.This may be partly due to the nature of the species: Scots pine does not form understorey in the same way as Norway spruce does (e.g.Siren 1955).It has also been found that the natural forest structure is still retained in many managed Norway spruce stands (Esseen et al. 1997).The distribution of nearly even-aged dominant trees is easier to predict than that of the suppressed trees of uneven-aged stands.The data of Norway spruce and especially pubescent birch may contain more dominated or suppressed trees than that of Scots pine.The different tree storeys were not separated in this study.
Excluding the understorey trees would probably have improved the results for Norway spruce.

Calibration Results
Calibration was performed for the percentile based methods and for the Weibull method.K-nearest neighbour estimates or alternative Weibull models on peatland were not calibrated, since these methods were only used in one case.Calibration was carried out in all different data sets except UNMANAGED and different mixed stands.In each data set with each method, the distribution was calibrated using number of stems (Table 8).Also basal area and d gM were used in calibration, to ensure correct values for these variables also after calibration.Observed minimum and maximum diameters were used to re-scale the predicted distribution to the correct interval in the case of percentile based method (Table 9).If minimum and maximum diameters were assumed to be known, these were used to recover the Weibull parameters instead of using prediction models (Table 9).Calibrating the basal area diameter distribution with number of stems proved to be problematic (Tables 8 and 9).Feasible solutions to the calibration equations were not found in all cases using logarithmic distance function.With the quadratic distance function (see Kangas and Maltamo 2000a), a feasible solution was found in all cases, but the obtained basal area diameter distribution included negative frequencies, which are not acceptable.Thus, if feasible solution was not found in a stand using the logarithmic distance function, the original predicted distribution was used for this stand.
In addition, the calibrated results were in some cases worse than without calibration, especially in the case of birch.Usually, calibrating the distribution obtained without using number of stems as a predictor (Percentiles 1) produced poorer results than could have been obtained by using the number of stems as a predictor (Percentiles 2).Calibrating a distribution obtained with Percentiles 2 method with respect to number of stems seems to be a more reasonable alternative.Number of stems as such is not a good predictor for stand volume, it needs to be related to some information concerning the size of the trees; in the Percentiles 2 models it was related to the basal area of the stands.
In average, calibrating with number of stems reduced the RMSE of stand volume by 19% in  -f), for Weibull method, the effect of calibration is minor in this stand.In the case of birch stand (Figs.3a-f) the Weibull prediction is modifi ed to slightly bimodal distribution The calibration also improved the proportion of correctly predicted percentiles, especially for the Percentiles 1 method (Table 6).In most cases, the calibrated Percentiles 1 method provided as accurate or even more accurate results than the Percentiles 2 method, which uses number of stems as a predictor.In any case, measuring the number of stems is worthwhile if an accurate estimate of the stand structure is of interest.

Discussion
The prediction of diameter distribution worked fairly well in most cases, the largest RMSEs of stand characteristics occurring in the northernmost Finland and in the peatland.However, the results were quite accurate also in this data; the biggest relative RMSEs were still less than 6% for stand volume.From the stand characteristics considered, stand volume could be predicted most accurately and number of stems least accurately.There were 5%-7% RMSE in the estimate of number of stems even if it was assumed to be known.With regard to the tree species, the Scots pine results were the most accurate, and the birch results the least accurate.The results obtained were quite similar to those obtained in the modelling data set (Kangas and Maltamo 2000b), which indicates that the models are reliable in varying conditions.
The accuracy of the percentile based method using number of stems as a predictor (Percentiles 2) was superior compared to other existing models.The accuracy of the percentile based method not using number of stems as a predictor (Percentiles 1) and the accuracy of the Weibull method were close to each other.This is natural because these models utilise the same amount of information.However, their superiority differed in different data sets.This same phenomenon has been established also in earlier studies (e.g.Maltamo 1997, Siipilehto 1999).However, it seems that percentile based methods are more fl exible with respect to varying stand structure and therefore produce more accurate results in extreme conditions.
In this study number of stems per hectare was obtained either by counting all the trees in a plot (circular and rectangular sample plots) or by computing from treewise basal area (angle-count sample plots).A number of stems estimate, especially one obtained from an angle-count sampling plot, may not be precise.This may also explain the great biases and errors which were obtained in the predictions of number of stems with data based on angle-counts.In addition, if number of stems is measured in forest inventories all trees may not be considered.For example in advanced stands, the smallest trees may not be measured.Then the number of stems is smaller than that which was used in this study.All in all, it is obvious that the measurement and use of number of stems require further investigation.Especially, the effects of the measurement errors of number of stems on the prediction of the distribution and on its calibration were not considered in this study and remain to be studied.
K-nearest neighbour estimates were calculated for Scots pine stands of the INKA material.These estimates were based on distance functions, number of nearest neighbours and value of weighting parameter from the results of the study by Maltamo and Kangas (1998).If these characteristics had been optimised from material used in the current study the results would have been even more accurate.The use of number of stems in distance functions would have further improved the results.Presumably, the accuracy of non-parametric estimates would then have been close to the accuracy of the percentile based models which include number of stems as a predictor, since the amount of information seems to be more important than the prediction method.Also other non-parametric methods, such as most similar neighbour (MSN) approach (Moeur and Stage 1995), can be used in description of diameter distribution.
With respect to number of stems, the results can be compared to those of Siipilehto (1999).When number of stems was not used as a predic-tor the accuracy of percentile based method was worse than that of the corresponding Johnson's SB models of Siipilehto (1999).If number of stems was used as a predictor, the percentile based method produced quite similar estimates except for the birch.For stand volume and timber volume, the results cannot be directly compared, because of different usage of height models: the height estimate can have a profound effect on the accuracy of volume estimates.Siipilehto (1999) used a height model, whereas in this study height model was not used.
Calibration seems to be worthwhile especially with respect to stand structure.However, calibration with number of stems did not always improve the accuracy of the results.The diameter distribution in some stands may be so heterogeneous that more information, for example about median diameters, would be required to make the calibration worthwhile.Re-scaling the predicted distribution with extreme diameters seemed to be worthwhile, especially with respect to saw timber volume.However, using these extreme diameters for parameter recovery in the case of Weibull method did not work so well.
The calibration of predicted distributions using additional information measured from the corresponding stand seems to be a very promising method.The results of this study and those of Kangas and Maltamo (2000a) indicate that good estimates are obtained in most cases.The calibration can be carried out for any diameter distribution prediction method, including non-parametric estimates.The calibration possibilities are not restricted purely to number of stems.If stem frequency diameter distributions are used, the calibration using basal area is very effective (Kangas and Maltamo 2000a).In irregular stands, assessing other mean diameter characteristics than basal area median diameter, e.g. the arithmetic mean diameter, may help.In tropical countries, other stand mean characteristics, such as dominant diameter, may be used.Also the basal area or number of stems in certain diameter classes, e.g. for trees bigger than any predefi ned value, could be useful in calibration.
In the case studied, feasible solutions for calibration with number of stems were not found if the number of stems estimate obtained from the original distribution was very poor.In those cases, the original predicted distribution was used.However, if the minimum and maximum diameters were assumed known, the feasible solution was found more often.Feasible solutions were always found with a quadratic distance function, but they included negative frequencies.Consequently, different forms of distance function may be useful in future studies (see e.g.Théberge 1999).
More attention should also be paid to prediction of tree heights.In this study, the height information was not utilised, since the tree volumes were predicted using a volume model with diameter as the only predictor.However, the effect of errors in tree heights can be important, especially when the volumes for the different timber sortiments are needed.In the future, the whole procedure from data acquisition to the calculation of the stand results needs to be considered.

Fig. 1 .
Fig. 1.An example of predicted and calibrated Scots pine diameter distributions.Subfi gures a, c and e show basal area diameter distributions and subfi gures b, d and f corresponding stem frequency diameter distributions.Subfi gures a and b show uncalibrated ( ) and calibrated ( ) Percentiles 2 method, subfi gures c and d show uncalibrated ( ) and calibrated ( ) Percentiles 1 method, and subfi gures e and f uncalibrated ( ) and calibrated ( ) Weibull method.Subfi gures a, c and e show the histogram of empirical basal area diameter distributions and subfi gures b, d and f the histogram of empirical stem frequency diameter distributions.

Fig. 2 .
Fig. 2.An example of predicted and calibrated Norway spruce diameter distributions.Subfi gures a, c and e show basal area diameter distributions and subfi gures b, d and f corresponding stem frequency diameter distributions.Subfi gures a and b show uncalibrated ( ) and calibrated ( ) Percentiles 2 method, subfi gures c and d show uncalibrated ( ) and calibrated ( ) Percentiles 1 method, and subfi gures e and f uncalibrated ( ) and calibrated ( ) Weibull method.Subfi gures a, c and e show the histogram of empirical basal area diameter distributions and subfi gures b, d and f the histogram of empirical stem frequency diameter distributions.

Fig. 3 .
Fig. 3.An example of predicted and calibrated birch species diameter distributions.Subfi gures a, c and e show basal area diameter distributions and subfi gures b, d and f corresponding stem frequency diameter distributions.Subfi gures a and b show uncalibrated ( ) and calibrated ( ) Percentiles 2 method, subfi gures c and d show uncalibrated ( ) and calibrated ( ) Percentiles 1 method, and subfi gures e and f uncalibrated ( ) and calibrated ( ) Weibull method.Subfi gures a, c and e show the histogram of empirical basal area diameter distributions and subfi gures b, d and f the histogram of empirical stem frequency diameter distributions.
large, which is to be expected when the age on peatlands is considered.However, as in forestry practise the stands with missing values cannot be dismissed, they were not dismissed in this test.In such a case, the use of imputed values is reasonable.Consequently, the estimated model was used to predict the stand age on such peatland stands, where it was not measured.
where d gM is basal area median diameter and TS temperature sum, d.d.(+5 °C threshold).The degree of determination of the model is 0.30 and standard error 24.87.The standard error of the model is quite

Table 1 .
The method of data collection, number of observations (n), and the mean value of basal area (G), number of stems (N), basal area median diameter (d gM ) and age in each data set by species.

Table 2 .
The results of the prediction of basal area diameter distribution of Scots pine in four independent test data sets (ANGLE, INKA, DITCHED and UNMANAGED).

Table 3 .
The results of the prediction of basal area diameter distribution of Norway spruce four independent test data sets (ANGLE, INKA, DITCHED and UNMANAGED).

Table 4 .
The results of the prediction of basal area diameter distribution of birch species in four independent test data sets (ANGLE, INKA, DITCHED and UNMANAGED).

Table 5 .
The results of the prediction of basal area diameter distribution in three different geographical areas of the INKA data.

Table 6 .
Relative proportion of diameter quartiles of number of stems and basal area which are predicted to correct diameter class.Results are presented in INKA test data set with (C) and without (U) calibration.

Table 7 .
The results of the prediction of basal area diameter distribution in data of mixed stands of Scots pine, Norway spruce and birch species (MIXED1, MIXED2 and MIXED3).