Forests provide a wide range of valuable ecosystem goods and services [1]. Understanding forest ecosystems and their underlying processes help with forecasts using scenarios under global change conditions and supports the establishment of sustainable management strategies [2,3,4,5]. The three-dimensional distribution of geometric objects and their topology within a forest canopy (termed “canopy structure” [6,7] is considered a particularly crucial constituent of forest ecosystems’ functioning and processes. Canopy structure can, among other things, serve as an indicator of forest stand resistance to disturbances [8,9], enable the estimation of the conservation potential for biodiversity [10,11], or allow the identification of recruitment limitations [12,13]. Canopy structure itself is not a measurable quantity, but the properties of canopy structure can be described by means of a wide variety of structural variables, such as tree height, tree diameter distribution, or canopy layering [14]. The assessment of canopy layering, in particular, is of great importance. Canopy layers provide essential structural components of wildlife habitat [15], affect stand productivity [16], and are used to manage forest stands and to identify forest structural stages [17]. From the various existing definitions of canopy layer (cf. [18]), we follow Franklin and Spies [19] and understand canopy layer as continuous vertical distribution of foliage within the canopy.

Traditionally, canopy layering is assessed by conventional fieldwork in relatively small areas, which is time-consuming and occasionally subjective [20,21]. Advances in Earth observation systems and analysis techniques have greatly improved the ability to characterize canopy structure over large areas in not only the horizontal but also the vertical dimension [22,23,24]. In particular, light detection and ranging (LiDAR) systems, especially airborne laser scanning (ALS), are suitable to provide very detailed vertical and horizontal information on canopy structure based on the physical measurement principle of active sensing and full-waveform digitization [25,26,27,28]. This also means that definitions of canopy layering which, for example, focus more on composition of life forms or age classes cannot be taken into consideration, as the types of physical measurement possible with ALS do not meet these definitions.

ALS-based assessment of canopy layering can be either derived from LiDAR waveform data (mostly the case for large or medium-footprint LiDAR systems) [29,30,31], or based on ALS point clouds directly [32,33] or indirectly, e.g., by generating a kind of synthetic waveform using classified echo height distributions or height distribution probability functions [34,35,36]. The assessment of canopy layering with ALS is mainly performed using area-based approaches (ABA). The spatial scale of ABA is currently often driven by user requirements and usually limited by the features present in the available data [37]. However, the choice of a specific spatial scale affects the canopy layer analysis, as a decrease in horizontal resolution will result in more mixing of horizontal and vertical structure elements within each spatial unit [34,35,38,39]. Most canopy-structure components, however, have inherent spatial scales and the choice of the spatial scale should be made considering the investigated structural component [40,41]. Nevertheless, an acceptable compromise often needs to be found between the technical capabilities of ALS and user needs, such as the application domain or cost-benefit considerations [42,43]. In this study, ABAs represented in regularly spaced grids were used, allowing flexible analysis in terms of spatial scales and cross-scale comparisons (e.g., testing reproducible up- and down-scaling methods).

ABA’s were investigated by Wulder et al. [25,44], White et al. [45], and Næsset [46] who found the concept to be sufficiently robust for operational use. The assessment of canopy layering, however, is rarely considered. Existing approaches to assess canopy layers often require a substantial set of prior information about stand characteristics or rely on pre-defined height thresholds for an existing stratification of the canopy. Whitehurst et al. [30], for example, on one hand used the cumulative canopy cover and pre-defined height thresholds to distinguish between understory (0 to 5 m), midstory (5 to 15 m), and overstory (>15 m). On the other hand, the peaks of the LiDAR waveform were used to determine canopy layers, assuming that each peak represents the mean layer height. The peak definition parameters, however, were determined by visual analysis [30]. Ferraz et al. [32] applied a mean shift segmentation algorithm to determine the modes of the distribution functions of echo heights and afterwards defined thresholds to divide the canopy into overstory, understory and ground vegetation. In this context, the definition of these thresholds remains critical as one component of the algorithm is linked to a function of canopy layer depth and thus depends on the tree species and/or forest biome [32]. Wang et al. [35] also used the modes of echo height distribution functions to divide the canopy into single-layered and double-layered canopies. Furthermore, the length of the determined layers was estimated, whereby only changes in absolute point density (i.e., inflexion points in the height distribution function) were used as the criterion to define the borders of a layer, which are not necessarily coincident. Maltamo et al. [36] also distribution functions of echo heights, in a manner comparable to Whitehurst et al. [30], Ferraz et al. [32], and Wang et al. [35], to relate the modes of the function (i.e., unimodal or multimodal) to the amount of canopy layers. In contrast, however, a fundamental part of their algorithm was the prior definition of a threshold to separate the dominant tree layer and understory trees, which limits the determination of canopy layers to a maximum of two layers. Morsdorf et al. [33] applied a cluster analysis to echo heights and echo intensities to stratify the canopy into three layers, whereby each layer consists almost entirely of oak, pine or shrubs, respectively. Although the proposed approach results in high accuracy values, a successful transfer of the method to dense, mixed temperate forests is unlikely. Zimble et al. [43] used a LiDAR-based canopy height model for the identification of individual trees and to determine their respective heights. The variances in tree heights were subsequently analysed on a 30 × 30 m cell size and used to distinguish between single-layered and multi-layered canopies. Accordingly, only trees forming the top canopy layer were considered and the resulting canopy stratification is more related to the composition of age classes and less to the actual vertical foliage distribution. Possible canopy layers below the top canopy layer, however, could not be assessed.

Therefore, most of the existing approaches are limited in their transferability to other sites and they tend not to be directly comparable. The inherent spatial scales of canopy-structure components are barely considered and, to our knowledge, a multi-scale approach regarding canopy layering has so far only been presented by Wang et al. [35]. To overcome these limitations, we propose using an automated and transferable ABA to provide quantitative assessments of canopy layering at different scales.

For the Laegern site, the validation shows promising results for all canopy-structure descriptors. The validation results for the canopy-structure descriptors are listed in Table 3, including the differentiation between the ALS and reference data used. The different ALS and reference data used influence the accuracy values of the derived canopy-structure descriptors only to a small degree. Nevertheless, all canopy-structure descriptors derived from the ALS data of 2010 show higher accuracies except for canopy_type. On average, the validation with the forest inventory data of the Laegern site results in slightly higher accuracies in comparison to the validation results based on the stand polygons of the Canton of Aargau. With 2.4% (cf. Table 3), however, the mean difference in overall accuracies caused by the used reference can be more or less neglected. The same applies for differences in the overall accuracies caused by the different ALS data used with a mean difference of 3.1%. Particularly in terms of the determination of canopy_type and canopy_{heterogeneity}, the developed algorithm shows a good performance. The classification of canopy_type leads to high overall accuracies up to 89.5%. The highest overall accuracy for canopy_{heterogeneity} is reached with 81.5%. Overall accuracies up to 66.0% are achieved for canopy_length, whereby misclassifications mainly occur in 2- or multi-layered forest stands. Canopy layering (canopy_layer) is derived with an overall accuracy of 67%, ranging from 80% (1-layered canopies) to 53.3% (multi-layered canopies). The highest overlap occurs between the 2- and multi-layered canopy classes. In general, most classification uncertainties are caused by continuous transition zones between forest stands differing in age and structure. For these regions an accurate separation (e.g., in terms of a definition of a borderline between a 1-layered and a 2-layered canopy) is complex, even more so in the field.

Figure 4 shows resulting maps of canopy-structure descriptors for a subset of the Laegern site.

The validation for the entire Canton of Aargau shows comparable results to the one of the Laegern site, with slightly higher overall accuracies. For canopy_type we achieve an overall accuracy of 90.7%. Canopy layering (canopy_layer) is derived with an overall accuracy of 69.2%. An overall accuracy of 70.3% is obtained for canopy_length. The detailed validation results are given in Table 4.

Figure 5 shows a larger subset of canopy_layer as an example of the derived canopy-structure descriptors for Canton of Aargau. In total for Canton of Aargau, 61% of the forested area is characterized by a 1-layered canopy, 38% by a 2-layered canopy, and only 1% by a multi-layered canopy. 81% of the forested area is covered with “deciduous canopies” and thus 19% with “evergreen canopies”. Long canopies occur in 88% of the forested area.

The analysis of leaf-on/leaf-off conditions on the RFDs showed that the added value of leaf-off data is relatively low if leaf-on data are available, in particular in terms of characterizing vertical foliage distribution which does not include the wooden elements in the forest (e.g., stems and branches). This finding, however, largely depends on the combination of the underlying grid-cell size and the respective point density and only refers to the impact on the RFDs. It is important to stress this because in contrast to our findings, other studies reported an added value of leaf-off acquisitions over leaf-on acquisitions for area-based canopy-structure estimates (cf. [64,65,66]). Moreover, Kim et al. [67] pointed out that leaf-off data only is also suitable to discriminate between evergreen and deciduous canopies, which can be an important issue in terms of decreasing costs for ALS-based forest monitoring. Nevertheless, the ALS data used in these studies, had much lower point densities, and the direct comparability with their results is therefore limited.

In terms of the point density, we also need to point out that the structural information expressed in the RFDs depends on the point density in connection with underlying grid-cell sizes [68]. Lim et al. [69] raised the issue of whether or not different point densities influence canopy-structure characterization, which is a contentious one. In general, RFDs based on lower point densities and calculated on smaller grid-cell sizes are difficult to interpret as a possible undersampling of the actual canopy structure increases. With increasing grid-cell sizes, however, the relevance of the point density on the RFDs decreases (cf. [34]). We showed in [34] that, for example, for a grid-cell size of 10 m × 10 m and a point density of 10 pts/m², the derived RFDs match to ~95% with RFDs based on much higher point densities (i.e., >60 pts/m²). A decrease in the point density to 5 pts/m² or 1 pts/m² decreases the match to 90% and 74%, respectively. For smaller grid-cell sizes, for example 5 m × 5 m, the match of the RFDs based on point densities of 5 pts/m² or 1 pts/m² becomes increasingly poor, at 85% and 59%, respectively. Whether this loss of information is significant or important still depends on the research question and/or the planned application. For example, Wilkes et al. [70], Hawbaker et al. [71], and Treitz et al. [42] pointed out that lower point densities (<5 pts/m²) are sufficient for canopy-structure analysis regarding tactical forest management, but these findings are only valid for canopy-structure descriptors derived on larger scales than the one we used in this study. Hayashi et al. [72], in contrast, concluded that in complex forests, low-density ALS data was ineffective in estimating the defined canopy-structure descriptors with sufficient accuracy. This is in accordance with our results and highlights the need and advantage of higher point densities. We thus recommend point densities >10 pts/m², particularly when the small-scale assessment of the vertical canopy structure is also of interest. Although ALS data with such high point densities are increasingly available for many Cantons in Switzerland, we are aware that this is not the standard. We are nevertheless convinced that in future nationwide ALS data with high point densities >5 pts/m² will be more and more available. For the second national height map of the Netherlands (AHN2), for example, LiDAR data with an approximate mean point density of 9 pts/m² were acquired already in 2011/2012 [73].

What should be also taken into account is a possible overestimation of upper canopy parts or an undersampling of lower canopy parts when using RFDs. This could be addressed by, for example, a weighting of echoes according to return number or intensity. However, as we only used the frequencies to filter the RFDs (i.e., determining filled and empty bins), the influence on the resulting canopy-structure descriptors is judged to be very small. Nevertheless, for canopy-structure descriptors which are more related to bio-physical variables, such as leaf area density, a proper weighting of the echoes should be taken into account [74].

The validation of the canopy descriptors proved good accuracies, particularly in terms of the seasonal and horizontal canopy variety (canopy_type and canopy_{heterogeneity)}. The seasonal canopy variety was very well covered with canopy_type and the reached overall accuracy is comparable to the highest obtained accuracies in existing studies (cf. [67,75,76]). Nevertheless, the validation of canopy_type for Canton of Aargau could be influenced by misclassification of deciduous conifers, such as European Larch, as the reference data contains the distinction between coniferous and deciduous trees only. As European Larch covers less than 2.5% of the total area, the overall accuracies were likely only marginally influenced. The validation of the vertical canopy stratification (canopy_layer and canopy_length) showed that misclassifications mainly occurred for forest stands which are very heterogeneous, such as multi-layered canopies with a long topmost canopy layer. For these forest stands, occlusion by the dense vegetation layer hinders a full sampling of the lower canopy parts even for areas with point densities >40 pts/m² available. Moreover, for most applications in forest inventory/management or forest ecology, there is an important difference if a second layer below the topmost canopy layer belongs to the forest floor/ forest understory or still belongs to the tree canopy. Since, based on the proposed method, a distinction between these types of layers would be possible, it should be included in following investigations of canopy layering.

The plot-wise accuracy assessment of the canopy-structure descriptors, however, turned out to be very difficult. The subjective visual evaluation of canopy_layer, canopy_{heterogeneity}, and canopy_length by the forestry experts includes a source of uncertainty, in particular for canopies with more than 2 layers (without a possible forest floor layer). Thus, this information cannot be regarded as an error-free reference: rather, it should be regarded as a data source for cross-comparison and feasibility checking. A possible alternative to create a more objective reference could be the modelling of canopy structure as presented by Scanlan et al. [77], using field measurements and allometric relationships. Furthermore, the canopy stratification approaches used in the forest inventory are more related to the composition of different tree development stages and less focused on the actual continuous vertical foliage distribution. For example, echoes from the vegetation in the lower canopy parts can be either from the forest floor or young growth or caused by low branches of larger trees. For the ALS-based canopy stratification we applied, it is difficult to distinguish between these different sources, even if a combination with methods at the individual tree level could help to address this issue, as, for example, understory and understory trees can be partly detected [32,78,79].

The assessment of the small-scale heterogeneity of the canopy (canopy_{heterogeneity}) showed that specific forest plots are characterized by a higher horizontal variability of vertical canopy structure than others and thus can be used to further distinguish between forest types. The spatial unit defined in most forest inventories to assess the heterogeneity of canopies, however, is with respect to a larger spatial unit (1 ha) than the one we used in this study (10 m × 10 m). It turned out that this change in the spatial unit made it difficult for the forestry experts to apply the usual approach to assess heterogeneity. Although the consideration of the small-scale heterogeneity thus may be not appropriate for forest inventory, these small patches within the forest serve as important indicators for biodiversity estimation or habitat assessment [80]. Therefore, if the available ALS data source facilitates a small-scale analysis it is useful to include LiDAR metrics derived on the smallest grid-cell size possible. Otherwise, a loss of spatial information regarding vertical canopy structure will occur.

The transferability of the method was proved by reaching comparably good results, at least for a similar type of temperate forest. The slightly better results for the canopy-structure descriptors derived for the whole of Canton of Aargau could be explained by the, on average, lower complexity of forest stands in comparison to the forest stands on the Laegern site. What still needs to be explored is the transferability of the method to forests with a highly complex canopy structure such as in the tropics or to forests with a very open/sparse canopy. In complex tropical forests it is likely that the density of the canopy hinders a full sampling of the vertical structure and the RFDs therefore cannot represent the actual canopy structure completely. In very open forests, in contrast, the grid-cell sizes need to be large enough to cover and describe the openness of the forest. In this context, more advanced self-calibration methods to derive grid-cell size in automated manner from the forest type and ALS data available need to be developed (cf. [34]).

For further applications, additional canopy-structure descriptors can be defined and/or adapted according to user’s requirements or, to comply with existing definitions and spatial units, can be aggregated to spatial scales used in the specific forest inventory system, such as the stand level.

In this study we present a method to derive grid-based characterization of canopy layering using ALS data. The resulting canopy-structure descriptors meet stakeholders’ requirements in terms of spatial resolution and information content. We demonstrated the transferability of the method from a small, well-characterized forest patch (800 ha) to a larger patch (180,000 ha), covering all forested area of the Canton of Aargau. Furthermore, we evaluated the advantages of using different grid-cell sizes and temporal information contained in multi-seasonal ALS data for further distinction and characterization of canopy layers. Our method to determine canopy-structure descriptors supports and improves operational, ALS-based canopy-structure monitoring in terms of inter-annual comparability.

Nevertheless, some limitations still need to be considered: (i) objective methods for field-based, non-destructive assessments of vertical foliage distribution are necessary in order to perform a more meaningful accuracy assessment; (ii) the transferability of the proposed method to very complex or open/sparse canopies remains unknown to date; and (iii) for the assessment of the small-scale structure, only ALS data with point densities >10 pts/m² are appropriate to obtain credible information. Additionally, proper interpretation of the canopy-structure descriptors still requires substantial forestry expert knowledge. Moreover, there is a need to link ground measurements and ALS-based canopy-structure descriptors considering differences in the used definitions and spatial scales.

The definitions of “significant/non-significant” in terms of temporal variations (canopy_type) and small-scale heterogeneity (canopy_{heterogeneity}) of the canopy still need to evolve, but can be used to represent stakeholder requirements, such as their use in forest inventory or forest ecology applications. We see the potential of our approach, however, in the field of forest ecology, for example wildlife habitat assessment (cf. [15]), rather than in the implementation in traditional forest inventories. The next steps will be to transfer the method to a broader variety of forests (including boreal and/or tropical forests) to ensure consistent application of the method. Once implemented, it can serve as a proxy for ecosystem-service-based assessments in forests, which includes the long term goal of monitoring forest change with high fidelity and accuracy.