1. Introduction
Forests are important global resources that affect numerous natural cycles as well as contributing to natural biodiversity, i.e., flora and fauna [
1]. Forested lands also constitute the largest terrestrial carbon sink on the planet, with approximate relative contributions of 80% being made by above-ground biomass and 40% being made by below-ground biomass [
2].
Forest structural information cannot be fully exploited if species information is missing. Indeed, precise species identification is a crucial variable for forest inventories [
3], for the quantification and monitoring of biodiversity [
4], and for the study of forest ecosystems and habitats [
5]. Accurate tree species identification is the information that is most frequently requested by the forestry industry and by government organisations in the elaboration of forest inventories [
6]. However, it is economically unfeasible to sample large numbers of trees in the field. As a consequence, remote sensing is essential not only to supply forest inventories with primary data [
7,
8,
9], but also to address environmental information needs.
High-resolution optical imagery is the most common source of remotely sensed data for species identification. For such images, pixel sizes that are larger than a tree crown may contain foliage from more than one species, leading to ambiguity and frequent identification errors. A small pixel size (e.g., 20–40 cm) implies that a tree crown would necessarily be composed of multiple pixels, leading to a situation where individual pixels will be spectrally representative of neither the tree nor the species. The pixels composing each crown thus include intra-specific spectral variability, which reduces classification accuracy [
10]. For this reason, most studies pertaining to tree species identification use object-based classification, which is frequently denoted Individual Tree Crown (ITC) segmentation or delineation [
11]. Once the tree crown is delineated, the individual pixels are extracted and summary statistics (e.g., mean spectral signature) and a gamut of features (spectral, spatial, and contextual features, among others) are calculated per crown, which reduces intra-specific spectral variation [
12,
13]. Optical imaging sensor methods also suffer from major shortcomings when used for species identification at the individual tree level. First, passive optical methods provide information regarding the top of the canopy, especially in dense broadleaf cover, but yield little to no information regarding vertical canopy structure [
2]. The second shortcoming of optical methods is related to the anisotropy of reflectance (dependent upon sun-sensor viewing geometry relative to the object) causing different spatial radiometric patterns of the spatial objects (e.g., sun-light vs. shaded crowns) [
14,
15]. The fact that the bidirectional reflectance distribution function (BRDF) effect is dependent upon the species further complicates the retrieval of information from optical imagery.
Within the broad range of remote sensing technologies that are available to practitioners, airborne laser scanning (ALS) is particularly well adapted to precision forestry, as it provides detailed structural information (given the laser pulse capacity to penetrate closed canopy) [
16]. Linear ALS systems are composed of a laser emitter (or multiple emitters in the case of multispectral systems) that produces pulses, which are emitted at a repetition rate of hundreds of kHz. The detector requires a flux of at least 500–1000 photons to register the backscattered laser energy from the target [
17]. The detector generates an electronic signal directly and linearly proportional to the backscattered laser energy from the target, hence the name “linear ALS.” The width and amplitude of the returned energy pulse depends upon the target characteristics. Proprietary algorithms transform the multiple peaks of a given waveform into discrete multiple returns. Semi-porous targets such as forest canopies can backscatter multiple peaks corresponding to different components of the canopy (top of crown, leaves, branches, trunk, ground).
The use of ALS data addresses some passive optical sensor limitations that are related to tree species identification. Give that it is an active sensor, lidar signal acquisition is permanently in the hot-spot configuration (the emission angle of the laser pulses is always the same as the viewing angle), which resolves many anisotropy issues [
18]. The ALS returns penetrate the canopy to various depths, sometimes reaching the ground. Therefore, the spatial ALS information (i.e.,
X,
Y and
Z coordinates) provides species-related structural information concerning the crown, branches and leaves [
19]. This canopy penetration and ground resolution capability is the major advantage of linear ALS over other remote sensing methods in the production of enhanced forest inventories.
ALS return intensity values, which are measures of the backscattered laser energy, bring supplemental information about tree species. Intensity values are not only strongly related to the type of foliage [
20] and its spectral signature, but the size, orientation, density, and clumping of leaves as well [
21,
22]. One of the main disadvantages of airborne lidar systems is that there are still many unanswered questions regarding the algorithms that are used to calculate ALS intensity values (they are proprietary to the various instrument manufacturers) and they preclude the comparison of lidar acquisitions that are provided by different sensors and over-flights. Additionally, the linear ALS system used in this study is monospectral, which precludes the use of vegetation indices to improve classification accuracy. In order to address the latter point, multispectral ALS systems is one of the latest major innovations to have developed over the past few years [
23]. The three channels with different wavelengths provide additional intensity features and permit the calculation of ratios analogous to NDVI. The intensity comparison issues between different surveys remain with MSL however, as with all lidar systems.
ALS technology is undergoing rapid evolution. One of the most important variables in ALS acquisition specifications is the point density or the average number of returns per m
2. This density depends upon flight altitude and flight speed for a given pulse repetition frequency. Therefore, there is a direct relationship between the cost (
$/km
−2) of acquisition and the point density. There is also a relationship between ALS point density and classification accuracy for ITC methods. Conversely, methods using the Area-Based Approach (ABA) provide good results at lower point densities and results accuracy that do not necessarily improve proportionally with point density [
24]. Even if the importance of these relationships is well known, it remains unclear what effect ALS point density exerts on ITC identification.
As soon as the first commercial linear ALS systems appeared in the mid-1990s, researchers also started to explore the use of photon-counting instruments, i.e., Single-Photon Lidar (SPL), to address some shortcomings of conventional or linear ALS systems such as the high cost to obtain coverage of an area, even when compared to optical imaging sensors. SPL covers larger areas at comparable densities at much higher flight altitude, potentially reducing costs [
25]. SPL also provides opportunities for more frequent over-flights. SPL instruments utilise beam energy in a more efficient manner than linear ALS; therefore, the former obtains a higher point density for a given flight altitude than the latter [
26]. Alternatively, SPL achieves acceptable point densities while flying at a higher altitude, thereby permitting greater coverage [
27]. SPL systems use a laser that is split into a 10 × 10 array of “beamlets” with the return energy being acquired by a 10 × 10 array of single-photon sensitive detectors [
28,
29]. The intensity value for each SPL return pulse is not well defined and is derived differently from that of linear systems. For example, the data provider for the SPL over-flight that was used in this study uses the pulse width of the returned energy as an analogue of linear lidar intensity. In addition, the return distribution, such as the ratio of first to second returns, is different in the SPL case when compared to linear ALS systems.
The short recovery time of the detector is a crucial element of SPL technology, as it enables multiple close-by photon measurements along the beam’s path for each laser pulse that is emitted. The high sensitivity that is required of the pulse detector to detect single photon returns from the surface also makes it susceptible to background noise; the most important noise source is solar illumination reflecting off said surface [
30]. This background noise is proportional to the instrument Field-Of-View (FOV) and to the receiver telescope aperture, both of which are reduced in the type of sensor that is used for this paper. Noise filtering algorithms, such as the Differential Cell Count method, are used to further reduce interference from background solar illumination [
31].
Several ABAs that have been developed under linear ALS systems were adapted for use with SPL data to map forest attributes. ABA metrics that were derived under multispectral ALS and SPL systems were comparable [
32,
33]. The SPL data resulted in slightly better estimates for all canopy structural variables compared to multispectral linear ALS, except for basal area. Since SPL covers 590 km
2 h
−1 compared to 50 km
2 h
−1 for multispectral linear ALS at equivalent point density, SPL sensors clearly provided a productivity advantage over linear ALS systems for methods using ABA [
34]. However, the classification performance of SPL for tree species identification has yet to be ascertained since the SPL point cloud exhibits both a different vertical distribution as well as differences in the ratio between first and second returns compared to linear mode systems.
The main objective of this study is to compare the tree species identification capabilities from three datasets that were acquired respectively with linear monospectral ALS, linear multispectral ALS, and an SPL system. To our knowledge, this is the first study to compare these three types of ALS systems when used for species classification at the individual crown level. In particular, we wish to verify whether the methods that were developed for linear ALS data perform as well with SPL data. Species identification methods were tested at three classification levels: broad species types (hardwood, HW vs. softwood, SW), narrow species groups (e.g., pines, spruces), and specific tree species. A secondary objective was to determine whether an increased number of species identification features that were derived from multispectral lidar or the higher point density of SPL provides greater classification accuracy compared to the standard mono-spectral linear ALS baseline. Finally, additional specific questions were addressed: Are the most relevant features the same for the three sensor types, or do they differ significantly? Does feature selection affect classification accuracy in the same manner for these three datasets?
4. Results
The RF classification accuracies were compared for four different species groupings, three ALS systems (ASL12, MSL16, SPL18), and four broad feature groupings: 3D only; intensity (I) only; all the features of a given ALS system; and all the features of all the ALS systems pooled (
Table 6). This comparison was performed following an initial variable selection (based upon MDA, inter-correlation, and VSURF). The best accuracies were achieved for the first level of classification, i.e., the type distinction between hardwood and softwood species, while the lowest accuracies occur at the 12 species level. At the finest classification level, there was a noticeable difference in accuracy between most hardwood (in the grey background of
Table 7) and softwood species (in the white background of
Table 7) for the best model (all sensors, all features) with eastern larch (
Larix laricina) being a notable exception to this pattern. This result was not necessarily surprisingly, given that larch is a deciduous softwood.
Multispectral ALS (3D + intensity features) produced the best results in all species groups, and all feature subsets, while SPL ALS displayed a systematically lower accuracy compared to the two other types of sensors. The decrease in performance was almost always greater going from standard ALS to SPL, highlighting the different nature of SPL compared to the two linear ALS systems. However, both linear ALS systems (standard and MSL) generally produced comparable results, with a small advantage being shown by the MSL sensor in most cases.
The relative information contents of the 3D and intensity features varied across systems. Unsurprisingly, the three-wavelength intensity features of MSL provided greater species identification performance than did its 3D features. The reverse was true in the case of the two other systems. In most cases, the contrast between the discrimination power of the 3D and the intensity features was greater in SPL, with the 3D features performing much better than the intensity features. The SPL models displayed two fewer intensity (I_) features than the standard ALS, given that they were more strongly correlated and were removed in the feature selection process. It must be reiterated that SPL intensity is an ill-defined quantity and care must be taken in the interpretation of results that are derived from it. In all cases, the single intensity channel of standard ALS provided greater accuracy than that of the SPL system, while the accuracy that was provided by the 3D features of SPL was similar to that of the other two sensors, or slightly lower.
For each ALS system and each species grouping, the greatest accuracy was attained when the 3D and intensity features were combined. For the simplest classification level (hardwood vs. softwood), the pooled 3D and intensity variables did not feature substantially greater accuracy compared to that of the best subset (intensity-only or 3D-only, depending upon the case). For the most complex level (12 species), the contrast was greater, particularly in the case of standard ALS, where the numbers rose from 38.9% (3D-only) to 50.7% (all).
Combining all the features from all the systems improved the accuracy in all cases but one (type discrimination using all available features). This improvement, in general, was about 5% compared to using MSL only, except for the classification of tree type.
Figure 4 shows the feature rankings for the 12 species-all sensors-all features model that were ordered by
Mean Decrease Gini and which were produced with the
varImpPlot function of the Random Forest package in R. The Mean Decrease Gini (unitless) is the mean of a feature’s total decrease in node impurity, weighted by the proportion of samples reaching that node in each individual decision tree in the Random Forest. It is a measure of how important a feature is for classification accuracy across all the trees in the Random Forest. The relative ranking of the features is of interest in these Figures. The suffix following the variable name of each feature refers to the dataset from which the feature was calculated. Slope features figured amongst the most important, as did both green channel-based multi-spectral vegetation indices. The first return intensity dispersion coefficient from standard ALS is the most significant feature for the 12 species classification.
Figure 5 and
Figure 6 break down in order of importance the features by 3D and intensity, respectively.
In most cases, more parsimonious classification models, i.e., using only the best 25 or 15 features, displayed only a slight decrease of accuracy compared to using all of the pre-selected features. This decrease was very small (≤0.5%, except for SPL) for the type level, and more apparent, while rarely exceeding 2% for the other classification systems. Our results indicate that the more complex sensors (MSL and SPL) did not substantially improve the performance of our models, with the SPL models being the least accurate in all cases.
6. Conclusions
This paper compared the performance in tree species identification achieved by three different lidar systems, including multispectral and single-photon instruments, at the individual tree crown level, using the same training crowns and methodology across the three datasets.
MSL provided the greatest species identification accuracy across all the groupings, while SPL displayed the lowest. In the case of the combined dataset, MSL provided more intensity features, while ALS and SPL provided mostly 3D features. When the results were broken down by feature type (3D vs. I), we found that geometric features performed better than intensity features for the monospectral linear and single-photon instruments. As expected, the enhanced intensity features of linear multispectral lidar performed better than the geometric features, even with the enhanced point density that had been acquired by the three laser beams in that particular instrument. In all cases, the combined geometric and intensity features performed the best. Single-photon lidar intensity features performed the poorest across all datasets. Interpreting this result is made difficult by the fact that the derivation and meaning of the SPL intensity measurements is still not well described in published research.
In dense mixed forests such as PRF, hardwoods remain a classification challenge at the 12 species classification level, while softwoods are classified more accurately. Hardwoods are more challenging to delineate accurately and are more prone to identification error when selecting training crowns in the field. The low number of exemplars in certain species classes lowered the effectiveness of the Random Forest classifier, since all classes would have their training data limited by the class with the lowest count.
The fact that training crown polygons were segmented and field-sampled in one year (2012) and used in subsequent lidar over-flights (2016 and 2018) is encouraging, as it means that fieldwork does not have to be duplicated to use a more recent acquisition. A novel combination of all three dataset features in a single classification model, which improved accuracy by an additional 5% in most cases, was performed as well. The success of this combination suggests that multi-temporal species differences between features arising from multiple lidar acquisitions would not necessarily have to originate from three different types of sensors, as were used in this study, but these differences in features could contribute to accuracy improvement, which merits further investigation.