1. Introduction
Remote sensing (RS) methods, such as optical and microwave satellite imaging, digital aerial photography and light detection and ranging (LiDAR), are highly useful in various forest-monitoring tasks. Until recently, knowledge of forest biomass and its changes has been based on ground measurements and coarse or medium resolution satellite images. Therefore, the accuracy of biomass estimations, especially at the local level (e.g., forest stands or sample plots), is poor. Stand biomass is highly correlated with tree heights that can be determined accurately by means of LiDAR, e.g., [
2]. It is expected that LiDAR applications will enhance the accuracy of forest biomass estimates at all levels from single-tree to nationwide inventory applications.
LiDAR measurements can be divided into profiling and scanning approaches. Airborne scanning LiDAR (small-footprint) has two main methods for deriving forest information: an area-based approach (ABA, [
3]) and individual tree detection (ITD, [
4]). LiDAR is a promising technique for efficient and accurate biomass detection, due to its capacity for direct measurement of vegetation structure and forest characteristics at different levels (from tree to stand level), e.g., [
5,
6]. With LiDAR’s ability to directly measure forest structure, including canopy height (CH) and crown dimensions, it is increasingly being used for forest inventories at different levels. Previous studies have shown that LiDAR data can be used to estimate a variety of forest inventory attributes, including tree, plot, and stand level estimates for tree height [
4,
7–
9], stem volume (VOL) [
3,
10,
11], basal area (BA) [
12–
14] and tree species [
15–
19]. In relation to biomass changes, LiDAR is also a promising method for monitoring forest hazards and defoliation, due to its ability to derive vegetation structure properties [
20–
24].
Inventory of stand aboveground biomass (AGB) can be based on LiDAR data at single time points and multitemporal LiDAR can be used in monitoring biomass changes. Lefsky
et al. [
12] showed that LiDAR-derived features such as quadratic mean CH could explain 80% of the variance in AGB. The structure of the forest canopy and leaf area index (LAI) affect the penetration of laser pulses in tree crowns [
25]. The LAI value represents the area of leaf surface per unit area of ground surface. Changes in AGB have also been estimated, using changes in LAI. The ground truth for LAI can be determined, using special measuring devices, or by estimation from the LiDAR data [
21,
25–
27]. Popescu
et al. [
28] studied the accuracy of individual tree crown diameter at the plot-level and its effect on AGB estimation from LiDAR data. The study area was covered by coniferous, deciduous and mixed stands of varying age classes. The effect on biomass estimation was notable; the estimate of crown diameter alone explained 78% of the variance in biomass and the R
2 values and root-mean-squared errors (RMSEs) improved by up to 0.24 and 7 t/ha, respectively.
Popescu
et al. [
29] combined small-footprint LiDAR and multispectral airborne data to estimate the plot-level VOL and AGB in deciduous and pine (
Pinus L.) forests, using ITD in which the average VOL ranged between 123 and 163 m
3/ha. The maximum R
2 values for AGB were 0.32 for deciduous trees and 0.82 for pines. The respective RMSEs were 44 t/ha and 29 t/ha. Bortolot and Wynne [
30] also used ITD for AGB estimation in young forests (ages between 11 and 16 yr) in which the correlation (r) varied from 0.59 to 0.82 and RMSEs from 13.6 to 140.4 t/ha. Van Aardt
et al. [
31] estimated the VOL and AGB with LiDAR point height metrics as predictors in a per-segment estimation in deciduous, coniferous and mixed forests. The adjusted R
2 and RMSE values for deciduous AGBs were 0.58 and 37.41 t/ha. Næsset [
32] used regression methods to estimate AGB for 143 sample plots in young and mature coniferous forests. The sample plot data were divided into three strata (I: young forest, II: mature forest with poor site quality and III: mature forest with favourable site quality). The regression models explained 92% of the variability in the AGB for all the forest types. Jochem
et al. [
33] used a semiempirical model that was originally developed for VOL estimation to estimate the AGB in Norway spruce (
Picea abies L.) dominated alpine forests. The model was extended with tree canopy transparency parameters (CTPs) extracted from LiDAR. The model was calibrated, using 196 selected sample plots. The R
2 values for the fitted AGB models were 0.70 with no CTP and varied from 0.64 to 0.71 with different CTPs. The standard deviations (stds) varied from 87.4 t/ha (35.8%) to 101.9 t/ha (41.7%). Latifi
et al. [
34] tested the ABA in southwestern Germany in VOL and AGB mapping. They found that the random forest method was superior to other nearest neighbour (NN) methods and achieved relative errors of 23.3%–31.4% in plot-level VOL and 22.4%–33.2% in AGB prediction, depending on the feature sets and feature selection used.
Breidenbach
et al. [
35] introduced a method, semi-ITC (Individual Tree Crown), that combines a modified ITD approach with ABA. The study introduced a practical solution for solving the problems with tree detection and the estimation bias of ITD. The semi-ITC method predicts volume for the entire segment, which can include none, one or more trees. The use of ABA in forest parameter estimation requires large amounts of training data, which are currently collected mainly with field measurements. Since field measurements are expensive and slow to make, new applications for acquiring training data are needed. Vastaranta
et al. [
1] introduced a method that combined ITD and ABA for forest parameter retrieval. They also used visual interpretation to diminish tree detection problems in ITD. Breidenbach
et al. [
36] used a new data collection strategy employing high-density scanning LiDAR from overlapping flight strips to improve the accuracy of
k nearest neighbour (kNN) estimates in the areas with low-density scanning LiDAR. The semi-ITC method was used with high-density LiDAR to make forest parameter predictions and the accuracy was tested, using external data with AGB as a response variable. The relative RMSEs varied from 26.7% to 41.56%, depending on the selection of reference plots used.
The objective of this study was to use a combination of two airborne scanning LiDAR-based applications, ITD [
4] and the ABA [
3,
14], in a manner similar to that used in [
1], to map AGB and VOL estimates in the same study area [
1]. The ABA was trained separately with field and ITD measurements. The main goal of the study was to find the usability and accuracy of the previously introduced method in AGB and VOL mapping. The accuracy of the biomass estimation was evaluated in Finland, where single-tree-level biomass models are available. The models used were evaluated with values of tree biomass measured in the laboratory. The study focuses on developing a new cost-efficient biomass mapping that would provide spatially explicit estimates.
4. Discussion
The estimation of forest AGB and VOL with airborne scanning LiDAR was tested in the present study. The results showed that LiDAR is capable of retrieving AGB accurately at the plot level. Validated current tree-level biomass models proved to be useful in AGB calculations, providing accuracies similar to those of general tree VOL models. Systematic shifts were found in validation of the current tree-level biomass models for biomasses larger than 250 kg, which meant that the large biomasses were underestimated. The bias in larger trees can be the reason for the increasing deviance that is seen in
Figure 6. Therefore the usability of these models should be studied in forests with varying development levels.
The AGB imputation accuracies (RMSEs) varied in this study from 24.9% (23 t/ha) to 34.9% (32.3 t/ha) at the plot level. The results were in line with previous studies, e.g., [
29,
30,
55,
56], in which the RMSEs varied from 14% to 42%. Latifi
et al. [
34] achieved plot-level accuracies of 22.2%–45.5%, using LiDAR data, and the best results were achieved with a random forest approach in comparison to other NN methodologies. The AGB imputation accuracy [
34] for the k-MSN method varied from 43.3% to 45.8%, depending on the variable set used. The results for the AGB imputation with the ABA
ITD-methods were, in general, similar to those in Latifi
et al. [
34], but to make adequate comparison of the NN methods used, all external factors (e.g., site-specific or the selection of model features) should be removed.
The best features for each ABA method were selected, using LASSO regression. The number of features selected was dependent on the penalty value (lambda), which was searched using cross-validation. The values selected were quite small and therefore the number of features in the models was large. This could cause overfitting, especially if highly correlated feature groups are present. With higher lambda value, fewer features were selected for the model, but this resulted in rapid increase in mean squared error (MSE) in cross-validation. Detailed comparison of the various feature selection methods when LiDAR features are used would be an interesting research topic since the correlated features here were selected for the optimal feature set. The impact of variable selection was also investigated by calculating the estimates for ABAITDauto and ABAITDvisual using selected features from ABAfield. Only small increase in bias and RMSE was found, e.g., the relative RMSE increased 1.6% and 0.5% for ABAITDauto and ABAITDvisual, respectively.
The AGBs imputed with the various ABA methods explained 71%–72% of the variation at the plot level. The R
2 values for the Scots pine, Norway spruce and deciduous plots varied between 0.73 and 0.76, 0.68 and 0.69 and 0.71 and 0.72. In comparison to previous studies, the total R
2 values were similar or somewhat better, e.g., [
28,
30].
The accuracy of tree VOL imputation at different levels (tree, plot or stand) is one of the most studied of forestry subjects, e.g., [
4,
13,
14,
34]. The accuracy of plot-level VOL imputation in the present study varied from 26.4% (48.4 m
3/ha) to 34.0% (62.4 m
3/ha), depending on the ABA training data used. Latifi
et al.’s [
34] imputation accuracies for VOL varied between 23.3% and 31.4%, depending on the feature set and selection method used. Popescu
et al. [
29] achieved VOL RMSEs of 52.84 m
3/ha and 47.9 m
3/ha for deciduous and pine plots, respectively. The VOL accuracy obtained in the present study with the ABA
ITD methods was similar to that in the above-mentioned studies. The best result was from ABA
field, but the results from the ABA
ITD methods were not far behind.
In large-scale forest biomass inventories, acquisition of extensive field references may be a major source of expense. In the present study, the training datasets for ABA were measured from LiDAR data and the imputation accuracies were compared with the accuracies imputed, using traditional field data. This type of approach could provide savings, especially in areas with difficult terrain or sparse road networks, and still maintain the spatial accuracy of the estimates. However, it is crucial to obtain unbiased estimates, which often presents a bottleneck if ITD is used in acquiring training data for the ABA. In the present study, ABA
ITDvisual resulted in even smaller bias for AGB and VOL than ABA
field. The relative bias was 0.8%–1.2% smaller for ABA
ITDvisual. This was most likely caused by the growth simulation used with the sample plots measured in 2007. It should be noted that the knowledge of tree species has a large impact on ABA estimation results and it was given as an auxiliary input data for the ITD training datasets in this study. The bias estimates might therefore be too low since tree species were not estimated using the LiDAR data. However, multiple promising studies have been published on this subject [
16–
19] and tree species classification should be integrated to the method in operational forestry.
Tree detection has been a major error source in ITD. Under Nordic forest conditions, tree detection accuracies with ITD
auto methods have varied from 20% to 90% [
57]. Other sources of bias in ITD methods include prediction of DBH and tree species classification [
58]. In this study, the composition of tree species was relatively simple. In more diverse forests, the biomass estimation and especially biomass model development would be more challenging. We used visual interpretation to reduce bias caused by commission and omission errors of tree detection. The bias for AGB was −13.9% with ABA
ITDauto. Visual interpretation reduced the bias to −3.1%. The bias in ITD could also be reduced by using plot-level biomass calibration data measured in the field or more sophisticated DBH prediction models [
24,
58,
59].