This study evaluated whether the MS-NFI pixel-level data can be used in generating the tree characteristics of individual stands and compared alternative prediction methods for diameter distribution. In general, the MS-NFI based methods (k-NN_trees, k-NN_stand) performed well for the total stand volume, and the average difference between the predicted and validation volumes ranged from 8–14%. The RMSE% using the satellite-based MS-NFI data varied between 25% – 30%, which was much smaller than in previous studies using Landsat imagery. Indeed, e.g., Mäkelä and Pekkarinen (2004) reported considerably larger 48% RMSE in total volume at stand-level using Landsat imagery. However, in Mäkelä and Pekkarinen (2004) the standwise results were predicted directly whereas in our study the standwise results were computed from predictions for pixels in the stand. Other previous studies have also reported RMSE% from 42–50% for total volume using satellite-based data for stand-level volume (Hyyppä et al. 2000; Hyvönen 2002; Muukkonen and Heiskanen 2005).
Satellite-based predictions have usually resulted in larger RMSE% than those based on ALS. In the preliminary study, the plot-level (32m × 32m) accuracy in stand characteristics were validated for DG, G and N with varying k (1 to 5) in k-NN method. At their best RMSEs were 5.1 cm (26%), 7.4 m2ha− 1 (32%) and 1304 ha− 1 (100%), respectively. In the previous study using the same field data, Tomppo et al. (2017) reported the respective best RMSEs of 2.0 cm (13%), 4.2 m2ha− 1 (18%), and 802 ha− 1 (36%) using fixed k = 5 and ALS instead of satellite imagery. Thus, the reliability in stand characteristics was considerably better, especially for N, when more accurate ALS was used instead of satellite-based data.
The Kolmogorov-Smirnov goodness-of-fit test at a 10% risk level rejected as much as 24% – 48% of the predicted dbh distributions for the 32 m × 32 m ground truth plots. The main reason for this rejection was the overestimated proportion of small trees (see Fig. 2), which, on the other hand, had a minor effect on the stand volume. Indeed, the 3-NN_trees was rejected most often, but simultaneously it was one of the best methods when validating the volume characteristics (lowest RMSE for pine and broadleaves). The reason for the difficulty in predicting the smaller trees may be due to the fact that this feature could not be detected, e.g., in spectral channel radiances. On the other hand, no published results exist on the accuracy of the Trestima approach to identify the smallest trees of a stand. Vastaranta et al. (2015) studied the Trestima approach for a sample plot measurement (basal area, mean diameter, mean height), but unfortunately not the accuracy in stem number. According to Siipilehto et al. (2016) and Ruusunen (2020), the bias in the stem number was between 2% and 5%, while the RMSE% was between 32% and 34% using the Trestima approach. According to Dunaeva (2017), stem number estimate by the Trestima approach was much more accurate than the estimates by forest experts in a preharvest field inventory. Also, the species composition was accurately estimated by Trestima compared to the harvester-based validation data (Dunaeva 2017).
In addition to the number and size of trees, information on tree species composition is a key parameter describing the stand structure for a wide variety of applications in forest management and conservation. The classification of stands by tree species composition is a complex task using satellite- or ALS-based data (Holopainen et al. 2008; Hovi et al. 2017), and the results of the present study were considered satisfactory. In this study, spruce was the most abundant tree species, pine was slightly less frequent, and broadleaves clearly a minority species. The main tree species was satisfactorily predicted, and the species proportions were generally quite well estimated, especially for broadleaves. Relatively good results for tree species dominance using Landsat TM and Sentinel-2 with NFI plots have been reported by Tomppo et al. (2009) and Breidenbach et al. (2020), especially for conifer species.
The validation data in the present study was small (27 stands) and the differences in the total and species-specific volume estimates were larger than in previous studies based on ALS (e.g., Maltamo et al. 2009, Tuominen et al. 2014). Indeed, the average differences in total volume were 13% for Scots pine and 35% for Norway spruce. Nevertheless, the stem volume for broadleaves in this study was at its best almost unbiased (0.1% – 1%) and otherwise at the same level (7% – 16%) to that by Packalén and Maltamo (2007), namely a bias of 11%. It is presumable that the smaller number of stands and different methods (satellite imagery vs. ALS) are the main reasons for the relatively high bias in this study. In addition, the stands of the present study were selected from stands with an irregular stand structure, in which an estimation with remote sensing methods usually leads to large RMSEs.
Maltamo et al. (2009) and Tomppo et al. (2017) reported increased accuracy in stand characteristics using fixed area instead of relascope plots as training data. Note that only the latest fixed area NFI plots were applied in this study. Also, one reason for the relatively similar results with the prediction methods was the update of the data set from 2015 to inventory year 2020. Even if the update period was considerably short, updating has an averaging feature. For example, Fig. 6 shows how much the first 10-year period decreased the initial differences.
When Holopainen et al. (2010) studied uncertainty in timber assortment estimates predicted from forest inventory data, the main source of error was forest inventory, either stand-wise field inventory or ALS based inventory, while the effects of generated stem distribution errors were minor. In Finland, when the stand-wise management-oriented field inventory was changed to the ALS based inventory, the mean stand characteristics were also changed from basal area-median tree dimensions (DGM, HGM) to weighted means (DG, HG). Recent results showed that DG is more stable in parameter recovery than DGM (Lee et al. 2021). Thus, the parameter recovery method makes the distribution errors negligible.
In the k-NN methods, the increasing k decreased the level of error in most of the evaluated variables. On the other hand, each value of k (1, 3, 4, 5) was best for at least one validated feature. However, after 30-year simulation k = 1 never provided the best result but instead, provided most frequently the worst volume estimates. In Holopainen et al. (2009), the best results were generally given by k = 4, but also the applied k = 3 or k = 5 provided the best results for some characteristics. According to review by Chirichi et al. (2016), the most frequently applied values of k were 1, 5, and 10. Maltamo et al. (2009) and Tuominen et al. (2014) ended up using k = 6. Nevertheless, the increasing k also considerably increased the required computational capacity and simulation time with the k-NN_tree methods. The present study found it too time consuming to simulate more than three times (k > 3) the number of trees (N) when generating a dbh distribution: when k was 3, the maximum number of 21,360 trees were simulated for a stand of 6.35 ha for the 3-NN_tree method. Because the results with the k-NN_tree and k-NN_stand methods were close to each other, the practical solution was to use only the k-NN_stand method when k was greater than 3 and generate only N/k trees from each k distribution. By doing so, the number of trees in the simulation was restricted to the number of trees per hectare multiplied with the area of stand compartment. In the case of the present study, the maximum number of simulated trees using the k-NN_stand method was 6,780 trees. Finally, each k species specific stand characteristic was used for the parameter recovery and thus this successfully mimicked the realization of measured trees. Previously, Maltamo and Kangas (1998) found clearly better results with the k-NN empirical distribution (i.e., k-NN_trees) compared with the prediction of Weibull function. However, the applied prediction model for the Weibull distribution (Maltamo 1997) was not as flexible as the parameter recovery method used in this study (see Siipilehto 2022).
The stand development was simulated over time and the effects of biases was compared in the initial stand structure by the different prediction methods. The 30-year simulation showed that differences between the prediction methods occur as long as the initial state has some influence on stand development, though selecting the prediction method for the initial stand became less significant over time. Even though the rank between the methods did not change much over time, the differences between the methods decreased. Similar results have been reported by Siipilehto (1999), Kangas and Maltamo (2003) and Mäkinen et al. (2010), with the reason most likely due to the feature that individual tree models predict higher diameter growth and less mortality for stands in which stocking level were initially underestimated and vice versa. If the initial state was accurate, the differences could even increase during simulation. Accordingly, Kangas and Maltamo (2003) noticed that a calibrated, more accurate initial state did not improve the accuracy of the predicted future volumes after simulation. The simulations of this study were carried out without intermediate disturbances, e.g., thinnings or damage, which would further reduce the effect of the predicted initial state on stand development. Thus, the selection of the prediction method can be made according to convenience to handle the data.