Forest growing stock volume (
) is an essential measure used to characterise forest structure, value and is often used to estimate forest above-ground biomass (AGB; t ha−1
). Maps and estimates of forest
and its changes are central to understand the carbon and water cycles, for assessing the climate change mitigation potential of forests and to quantify ecosystem services [1
]. Several earth observation missions, such as Global Ecosystem Dynamics Investigation (GEDI),
NASA-ISRO SAR Mission (NISAR), and BIOMASS, aim specifically at mapping forest AGB and will improve our ability to consistently map and estimate forest resources across the globe [2
]. Still, all operational applications of remote sensing techniques for forest inventory require field observations for calibration of models and their validation [4
]. Due to high costs of field work, the development of methods for forest inventories and mapping that only utilise remotely sensed data without the need for in-situ field data represent a breakthrough for the utility of RS data [5
]. We will use the term ‘direct estimation’ to describe remote-sensing based estimates that are obtained without the use of in-situ field data for fitting or calibrating empirical models.
In the past decade, terrestrial laser scanning (TLS) emerged as a promising data source for directly estimating forest AGB [6
]. Furthermore, airborne laser scanning (ALS) data have also been used to directly estimate AGB [5
]. Recent technical development in miniaturised survey-grade laser scanning sensors suitable for unmanned aerial vehicles (UAV-LS) allows for the collection of detailed three-dimensional (3D) data similar to TLS data with the wall-to-wall characteristics of airborne data. Thanks to the combination of the wall-to-wall capabilities and the finely detailed representation of single trees, UAV-LS data opens new frontiers in direct forest measurement from remote sensing data [9
UAV-LS data are particularly attractive as they allow for dense (1000–20,000 points m−2
) sampling of the 3D structure of the forest canopy, including crowns, stems, branches, and understory vegetation. Thus, UAV-LS data have the potential to be used in a similar manner as TLS to obtain direct estimates of tree-level characteristics such as volume or AGB. While UAV-LS data have more occlusion than TLS data in the lower parts of the canopy, it can be collected on a much larger area and with only a fraction of the time for data acquisition. The possibility to use UAV-LS data for estimating AGB or growing stock is particularly attractive as calibration and validation datasets for space-borne earth observation programs [10
]. In such applications, UAV-LS data may be used to estimate AGB or growing stock for the exact spatial extent and shape of the space-borne sensor footprint.
Early studies by Jaakkola et al. [11
] and Wallace et al. [12
] first introduced the use of UAV-LS for forest inventory. Since then, an increasing number of studies adopted single tree methods to assess biophysical properties such as tree height [13
], tree crown properties [15
], tree density [15
], or diameter at breast height (DBH) measurements [14
] and AGB [20
]. The latter two variables are the most challenging and yet relevant, which potentially can justify the high costs of acquiring UAV-LS data. DBH is a relevant variable as it is widely used in forest inventory for valuation and simultaneously allows to obtain volumetric single tree estimates using existing allometric models. Despite the somewhat encouraging results in terms of DBH measurement accuracy (1–7 cm), the existing studies did not provide methods for how UAV-LS single tree measurements can be used to predict single tree volume. The exception is the recent study of Wang et al. [21
] that adopted a tree-centric method [22
] consisting of using measured height and DBH from UAV-LS data, with species information from a field-plot survey to predict single tree AGB using pre-existing allometric models. They obtained plot-level estimates by excluding the commission errors from the segmentation algorithm, which makes the method reliant on the availability of field data. A general finding of all UAV-LS studies is that DBH measurements are only possible for a sample of the entire population, including mainly dominant or isolated trees.
Based on the reported omission errors from UAV-LS data (15%–80%) and difficulties of measuring DBH for all trees [20
], it is not foreseeable in the near future to obtain UAV-LS based AGB or growing stock predictions for all trees in a population. A key scientific challenge is, therefore, to develop methods for using the available UAV-LS single tree measurements into plot-, stand-, or forest-level estimates while accounting for the non-probability nature of the sampled trees.
Potential ways for direct estimation of AGB without the use of in-situ observations have been proposed [5
] and consist of estimating single tree AGB by applying available allometric models using ALS-derived single tree variables (e.g., top height, crown area, crown base height) as predictors. While the approach developed by Ferraz et al. [5
] is useful as it can be applied over large areas and can account for different vertical layers of the forest canopy, it was designed to segment tree crowns, and its applicability to discriminate between stems and understory vegetation in the lower parts of the canopy is unknown. The substantial increase in point density of UAV-LS data compared to ALS requires methods that, similar to TLS applications, can separate the stems from the other vegetation, especially in the lower parts of the canopy, thus allowing DBH measurements.
The objective of this study is to illustrate a model-based method to use a sample of trees with DBH, tree species and tree height measured from UAV-LS data to predict single tree and to use the predictions to estimate at the plot, stand, and forest levels. The proposed method relies solely on UAV-LS and field data (i.e., 58 plots) were used only for independent validation.
The results of the Random Forests model for the DBH model highlighted the importance of height percentiles, crown shape, and spectral variables to describe the DBH variation (Figure 5
). For the tree species model, the most important variables were the spectral variables, followed by
(i.e., the rate of decrease of point density when going from the ground level to the tree top), and two-point density variables corresponding to vertical slices taken at one-third of the total tree height (d30
The scatterplot of the DBH model’s predictions against the field measurements (Figure 6
) revealed that despite a reasonably good model fit for the observations around the average, there was a tendency of over-predicting small DBH values and under-predicting large DBH values.
The accuracy assessment of the tree species model was performed by comparing out-of-bag predictions with observations (Table 2
). The overall accuracy was 77.1% and the species-specific producer’s accuracy was 79.5%, 86.7%, and 57.8% for spruce, pine, and deciduous species. The poorer performance of the deciduous class was mainly described by the fact that 36% of the deciduous trees in the sample were classified as pine.
The plot-level validation was performed using
= 58 field plots and the scatterplot of the UAV estimated against the field measured volume revealed a strong correlation between the two (Figure 7
was 103.4 m3
or 32% of the mean, and despite the presence of two pine-dominated plots with large residuals, there were no substantial differences among the
valuesby tree species (not shown). However, it was noteworthy that
was underestimated for most of the spruce-dominated plots. The largest residuals were found for plots composed of a dominant layer of large pine trees and a co-dominant layer of spruce and birch. The bootstrap estimates revealed an average
for the UAV-LS plot measurements across all plots of 31.5 m3
, corresponding to 9.7% of the mean value.
The stand-level validation was performed using
= 14 stands. The scatterplot of the UAV stand estimates against the field estimated volume (Figure 7
b) revealed that the averaging of plot or tree level values reduced the volume ranges and thus reduced the
to 93.1 (
= 28.9%). As for the plot-level validation, the
was slightly underestimated for spruce-dominated stands while the same was not evident for the other tree species. It is, however, important to remember that for nearly 50% of the stands the uncertainty of the field estimates was rather large (see Figure 7
) given that it was based on only three field plots for some stands. Furthermore,
was of 26.0 m3
, or 7.6% of the mean on average.
The forest-level validation was performed by comparing the estimated mean based on the volume predictions for all UAV-detected tree crowns (n
= 9111) with the design-based estimate using all field plots (n
= 58). The mean volume from the UAV measurements was 3.5% larger than the field estimated mean volume (Table 3
) but well within the 95% confidence interval of the design-based point estimate. Furthermore, the precision of the UAV estimate was similar (SE = 18.0 m3
) to the design-based estimate (SE = 18.6 m3
) (Figure 7
The analysis of the plot-level absolute differences between the field measured and the UAV-LS estimated
) revealed that the estimates were affected by tree density, dominant species, and species mix. The most accurate results were found for open (0–1000 trees ha−1
) pine-dominated plots. The accuracy decreased with increasing tree density and was found to be smallest for deciduous-dominated plots. Concerning the effect of species mix on the accuracy of the
predictions, we found larger accuracy in pure stands (i.e., stands where the
for one of the tree species > 70% of total plot
) compared to mixed-species stands.
This study shows that, by using UAV survey-grade laser scanning data, it is possible to produce model-based estimates of plot-, stand-, and forest-level growing stocks with a high level of correspondence to traditional design-based estimates with field measurements. Our findings confirmed the results by Ferraz et al. [5
], who first demonstrated the possibility to use single tree measurements from laser scanning point clouds as inputs to existing allometric models to predict AGB at different scales.
The results of this study revealed that the precision of the UAV-LS estimates increased with the spatial scale from plot to stand and forest level and that in none of these cases, were there severe systematic errors. This findings were consistent with previous literature showing the predictive accuracy to increase when increasing the geographical scale due to averaging effects of errors [23
]. At plot and stand level the
was somewhat larger (32.2%) than what can be expected in area-based airborne laser scanning forest inventories in similar forests [49
] and larger than reported by Ferraz et al. [5
] for eucalyptus plantations (
= 17.1%). At forest level, the UAV-based estimates were well within the 95% confidence interval of the estimates based on an intensive field survey and are thus not significantly different from the design-based estimate. Furthermore, the
was on all levels smaller than the
which means that the UAV-based estimates are more precise for estimating growing stock volume on the respective level. Based on the results from the bootstrapping, we found that the precision of the forest-level UAV-LS estimate was of the same magnitude of a design-based estimate using a dense network of field plots. It is essential to remember that our attempt to estimate the precision at forest-level for the UAV-LS estimate was limited by the availability of suitable variance estimators for a case where two Random Forests models are used to predict explanatory variables used then as predictors in allometric models.
The effect of the accuracy and precision of the UAV-LS estimates at different scales has implications on which applications may be most suitable for these data. For forest management, inventories the forest stand (1–10 hectares) represents the smallest management unit. According to the results of this study, the use of UAV-LS data can produce reliable measurements that could reduce the need for fieldwork for stand-level forest management inventories to the acquisition of some data for quality control. Furthermore, our results are encouraging for future use of UAV-LS data in calibration and validation of different types of space-borne remotely sensed data in a similar fashion to TLS data [50
]. It is important to note that tree density, tree species, and tree species mix affected the accuracy of the UAV-LS plot
estimates. The mentioned sources of variation in the accuracy directly affect the accuracy of DBH measurements from UAV-LS and the quality of the crown segmentation. In particular, the accuracy decreased with increasing forest structure complexity (i.e., dense and mixed species plots) and was largest for open pine-dominated plots. While it is rather intuitive that the occlusion rates and segmentation errors increase when increasing the tree density, the effect of different tree species on the ability to directly estimate
using UAV-LS represents a non-trivial issue. While we present some first results under boreal managed forest conditions, it is important in future research to better understand the transferability of the proposed method to more complex forest structures.
This study expands from the studies by Jaakkola et al. [18
], Brede et al. [14
], and Wieser et al. [19
] by including methods to utilise DBH measurements from UAV-LS data together with species and height information to model tree volume. The comparison of our results with previous studies using UAV-LS for direct measurement of tree properties [14
] was not possible since none of these early studies assessed
and because they were mostly conducted in uniform forest areas with relatively low tree density (400–805 trees ha−1
) with limited understory vegetation. In this study, the measured tree density at plot level was in the range 400–4400 trees ha−1
covering a larger variety of forest structures than previous studies, ranging from open pine forest to pure spruce or deciduous plots and including a range of mixes between the three species.
In this study, to predict single tree volume, we used allometric models relying on the input predictions of DBH and tree species. Amongst the explanatory variables describing crown geometry only the ratio between crown area and perimeter (i.e., compactness index) was selected in the final model while the crown area was ranked only 21st
. The latter is often an important explanatory variable to describe the DBH and its lack of importance in the DBH model may be due to a sub-optimal segmentation of the tree crowns or due to the large correlation of the crown area with other variables. Thus, better segmentation methods are likely to increase the importance of variables describing crown geometry and possibly improve the DBH model. Concerning the model’s predictive accuracy, the
was 8 cm, which is slightly larger than what reported by Jaakkola et al. [18
] (2.5–6.8 cm) in a similar study. However, it is important to highlight that the model had a tendency of predicting towards the mean, while under- and over-predicting large and small trees, respectively. Factors such as the crown segmentation quality and the error in DBH measurements from UAV-LS are potential error sources explaining the model performance and further studies should attempt to improve similar DBH models by reducing these sources of error. Nevertheless, the independent validation did not show serious systematic errors caused by DBH model, suggesting that the under-prediction for some of the trees may have been levelled out by the over-prediction of other trees. One important aspect to account for is that different tree species have a different probability of having a reliable DBH measurement from UAV-LS data. The branching and canopy structure characteristic of different tree species affect the amount of occlusion from UAV-LS returns at 1.3 m. This effect may partly explain why we observed a consistent under-estimation in spruce dominated plots and stands (see Figure 7
and Figure 8
). Further research should thus explore ways to include species-specific probabilities of including a DBH measurement in the sample used to fit the DBH model and see how this affects species-specific estimates at different spatial scales.
Concerning the tree species model, seven of the 10 most important variables were calculated from the RGB imagery, highlighting the complementarity of the RGB and LS data. The simultaneous acquisition of RGB images with LS data is a common practice in UAV-LS acquisitions and hardly increases costs. Despite the poorer accuracy found for the deciduous class, the results of the classification were deemed satisfactory and the tree species map representative of the reality.
While the main focus of this study was not to assess the accuracy of the single tree detection and segmentation, it remains important to acknowledge that the results obtained were dependent on the choice of methods for tree crown detection, and segmentation, and for DBH measurement. As a means to ensure a large number of potential candidates of trees with UAV-LS DBH measurements, the local maxima detection was characterised by rather small window sizes. The tree-crown segmentation was done using the CHM, and thus, for the detected suppressed trees, both the crown delineation and height measurement were characterised by errors. The adoption of more advanced segmentation methods allowing users to discriminate dominant and dominated tree crowns could result in better predictor variables for each of the devised models.
An important aspect of this study was that our method relied on multiple parameter values (e.g., minimum distance between trees, minimum DBH) which were determined based either on general knowledge on the forest type (i.e., rare occurrence of DBH > 60 cm in managed boreal forests) or by a trial-and-error approach informed by the UAV-LS data. While such parametrization does not affect the repeatability of the study, which can be replicated using the same data and parameters used in this study, the reduction of parameter should be a focus of future research to enhance the transferability of the method to new data. While some parameters could be determined adaptively from the data itself as a function of, for example, point density, biophysically meaningful parameters are more difficult to estimate automatically. However, methods to predict the latter based on models fitted to a sample of trees have shown to be promising [23
]. Ferraz et al. [23
] proposed a 3D adaptive mean shift (AMS3D) method which allows users to segment single trees crowns and requires minimal parametrization. A key advantage of the method is the possibility to derive tree crown properties such as tree height, crown base height, crown volume, which can be directly plugged into existing allometric models to predict tree DBH and AGB. Because the AMS3D method was developed using sparse ALS data (approximately 10 points m−2
) compared to the UAV-LS data used in this study (1300 points m−2
), it was designed to segment the tree crowns rather than tree stems. The possibility to segment and measure tree stems is a key advantage of UAV-LS data over ALS data as it allows to directly quantify the stem volume, which is the largest portion of the tree biomass. As demonstrated by the complexity of the methods commonly used to segment single trees in TLS data, the retrieval of stem measurement from very dense point clouds may require additional or even different methods to those proposed by Ferraz et al. [23
This study was limited by the lower point density (1130 points m−2
) compared to previous UAV-LS literature, which was at the lower end of the range compared to similar studies using Riegl VUX-1 data (1500–18,000 points m−2
]. In practice, this meant that for many of the detected trees, there were not enough points at 1.3 m to be able to fit a circle and thus to derive DBH estimates. As shown by Schneider et al. [51
], UAV-LS data are characterised by large occlusion rates in proximity to the ground. With an increased point density, the occlusion rates at breast height are reduced thus increasing the probability of reliable DBH measurements from the UAV-LS point cloud. The increase in the number of sample trees to be used to fit DBH models can potentially contribute to improving the model’s predictive accuracy.
A further limitation of this study was the fact that our methods relied on a semi-automated procedure, which included a manual step to visually select trees with reliable UAV-LS DBH measurements and to classify different tree species to use available allometric models. The development of more sophisticated methods to select reference data from the UAV-LS-detected trees could enable the collection of a larger sample of trees, for example, by analysing the whole vertical profile of the data and measure diameters at multiple heights [18
] or fitting of quantitative structure models [52
]. By becoming independent of allometric models, such an approach could be applied to a wide range of forest types. Furthermore, since the parameters of the proposed method were selected based on a trial-and-error approach, the applicability to different forest types than the ones in this study is unknown and could require tuning of the parameters.
Concerning the practical application of the proposed method, our study represents a first stepping-stone toward fully airborne forest inventories. Currently, because of the large costs and limited geographical coverage possible with survey-grade UAV-LS data, i.e., 1–10 km2
], these data should be seen as an alternative to field data (UAV-LS plots) rather than a large-scale mapping tool. In this regard, they could be used to calibrate models using other wall-to-wall remotely sensed data such as ALS or satellite data. An important advantage of UAV-LS over traditional field plots is that they can cover larger areas of any shape and thus can provide better training data for wall-to-wall remotely sensed data. However, UAV-LS is more sensitive to weather (e.g., wind and rain) than traditional field work and administrative regulations may restrict the use of UAVs in general, which needs to be considered under operational applications.
Although we assessed only GSV, UAV-LS data offers unique opportunities to derive a significantly larger pool of measurements compared to traditional field surveys. The possibility to sample the tree crown and upper stem can allow to expand well beyond the variables that are often measured using field data, including information about the shape of tree stems, the assortments obtainable and crown-related variables such as the effective leaf area index. Interestingly we found similar costs for the UAV-LS and field data collections, with the former also providing full-coverage data rather than a 16% sampling fraction by the field data. A fairer comparison of the cost–benefits of two data sources would need to better evaluate the full information potential of UAV-LS data and the development of costs under operational settings. In the future, thanks to technological advances, it is likely that costs for UAV-LS data capture will decrease while increasing their area coverage. As a result, the proposed method could become of interest also for mapping purposes over larger areas.