1. Introduction
Tropical forests are often located in remote and difficult-to-access areas. Therefore, field data collection costs are high, which forces compromises in the measurements collected or the number of locations sampled. As a consequence, high-quality forest structure information is difficult and expensive to obtain by traditional ground surveys in these areas. Airborne Light Detection And Ranging (LiDAR) data have been widely used to produce structural parameter estimates of both temperate and tropical forests and to monitor native and commercial forests [
1,
2,
3,
4,
5].
This technology provides a quick and complete assessment of forest structure, which allows the calculation of metrics such as canopy height, wood volume, biomass, and carbon stocks [
4,
6]. In addition, LiDAR also has multiple applications in the planning and monitoring of activities related to forest management through the assessment of digital elevation and surface models with sub-meter accuracy [
7,
8,
9,
10,
11], enabling surveying of areas difficult to access at relatively low cost. Forest monitoring, particularly of areas undergoing forest management, require repeat LiDAR estimates of damages produced by logging, changes in the canopy cover [
12], and biomass stock dynamics [
5,
13]. The benefits for scientists and forest companies of LiDAR data are immense. However, despite the great usefulness of this technology, its acquisition is still expensive [
14] and limited to large, contiguous areas. Challenges in obtaining data are considerable for regions that are furthest away from population centers where the companies providing these services are usually located (e.g., [
7]), which limits the use of LiDAR surveys—notably those that require repeated flights over the same area or that do not have large budgets for data acquisition.
In the last few years, we have observed an increase in the use of unmanned aerial vehicles (UAVs) for forest use [
15,
16,
17] and as a complementary tool to aircraft-borne LiDAR for forest studies [
18]. For example, UAV-borne visual sensors have provided 3D products through photogrammetric analyses and high-resolution orthomosaics, creating a revolution in landscape mapping through its combined low-cost hardware and high-resolution outputs [
17,
19,
20]. This approach is limited, however, to mapping areas visible from multiple perspectives, and as such, visibility is typically unable to penetrate most forest canopies to the ground and therefore has significantly limited products [
19,
21]; for example, it is unable to produce digital elevation models, which require ground points, and therefore tree height models—which are critical for most studies of aboveground biomass.
In the search for alternatives, LiDAR sensors have very recently become sufficiently small to be mounted on UAVs and have been used to generate models similar to those produced by standard aircraft-borne LiDAR systems, substantially extending the usefulness of UAVs [
21,
22,
23,
24]. UAV-borne LiDAR—which usually employs a flight above ground level (AGL), below 100 m, at a low (10–40 km/h) speed, with wider scanning angles, and high pulse frequency—are capable of producing very high-density point clouds, which largely exceed the ones produced by aircraft-borne LiDAR. The development of allometric models to estimate aboveground dry biomass stored in dominant and co-dominant individual trees in tropical forests is feasible with the typical 5–10 pts/m
2 LiDAR data obtained from aircraft-borne LiDAR sensors [
25]. However, the much higher point-density clouds produced by UAV-borne LiDAR enable more extensive and varied uses, including assessing interior forest structure with higher precision and accuracy [
21,
26], digital terrain model with very high resolution [
27], direct diameter-at-breast height (DBH) estimates [
24], and individual tree detection and detailed crown segmentation [
28].
The objective of this study was to compare LiDAR data and its products obtained from two different platforms: aircraft and UAV. Specifically, we compared: (a) LiDAR point clouds and metrics; (b) digital terrain, surface and canopy models, and (c) aboveground biomass (AGB) models for a group of forest inventory plots located in an Amazonian tropical forest in the Chico Mendes Extractive Reserve in Acre, Brazil.
4. Discussion
Our study demonstrates the potential of UAV-borne LiDAR to assess the structure and biomass of tropical forests in the Amazon. To our knowledge, this is the first study to compare aircraft- and UAV-borne LiDAR AGB predictions for tropical forests. The plots did not suffer significant natural or anthropogenic disturbances during the two-year interval (2015–2017) between the aircraft and GatorEye flights. As, during this time, only limited natural changes in the forest canopy occurred, we were able to robustly compare vegetation metrics and models produced by the two systems, similar to as if they had happened synchronously [
37].
The GatorEye sensor pulse frequency was only twice the Harrier 68i Trimble, but, in part due to having 16 lasers versus the one used in the Harrier system, it produced a point cloud almost 35 times denser. The main differences between the point clouds generated by these sensors were due to the GatorEye lower flight height and speed (60 m AGL at 12 m/s) when compared with the aircraft (600 m at 55 m/s), and due to the number of lasers. Low-altitude flight at relatively slow speeds can produce point densities much higher than traditional aircraft LiDAR approaches [
38]. The dense point cloud generated by GatorEye in this study was similar to or higher than that of other UAV-borne LiDAR systems [
28,
39], which allowed the identification of individual tree and branch structure with results similar, in the overstory, to terrestrial laser scanners. That allows a broader use of these data for applications such as individual canopy tree detection, stem segmentation, tree height, and estimation of understory leaf area index [
21,
24]. The differences in point cloud density were possibly responsible for the main differences observed among the extracted LiDAR metrics. Even when the density of the points is not as high as in this study, they can significantly improve CHM-based metrics derived from vegetation in both temperate and tropical forests [
40].
Some UAV-borne LiDAR sensors are capable of producing multi-return per pulse (e.g., RIEGL VUX series); however, the GatorEye’s Velodyne VLP-16 sensor returns the strongest and last returns. When we compared the ground-filtered clouds of both sensors, we observed that even when flying at a higher altitude and speed, the Harrier 68i was capable of producing up to five useful returns per pulse. The second and particularly third pulse echoes mostly exist beneath trees [
22]; thus analyzing the Harrier 68i ground filtered point cloud, we observe more second and third returns than first. The cloud density of the ground returns of the GatorEye’s cloud was still higher than that from the aircraft system but had a higher standard deviation in the number of ground returns per area, indicating the GatorEye ground points were not as uniformly distributed across the ground as the returns produced by the aircraft system. We attribute this difference to the GatorEye acquiring very dense point in areas of lower leaf area index (LAI), with more limited returns in areas of dense coverage, while the aircraft-system more fully penetrated across the entire forest floor.
The differences highlighted above did not affect the correlation of the high-resolution (1 × 1 m) digital terrain models (DTM) produced by both systems within the PSPs. Typically, at high resolutions, the relatively low-density point clouds of both systems would force estimation of ground location and forest structure through statistical interpolation approaches [
14]. However, we found an almost total agreement between the DTMs from both systems, indicating that the ground returns from both systems were sufficient to fully and accurately represent the forest ground surface—a critical aspect for post-process work such as AGB calculation.
Different from the DTM, although highly correlated, the much higher GatorEye point cloud density, and the time interval between the flights, did produce some noticeable differences between the DSM and CHM models. Changes observed in the forest structure, through natural tree falls and growth canopy trees as well as within previously existing gap areas in the PSP, produced height differences. Thus, some negative height differences (CHM 2017–CHM2015) can be attributed to tree- and branch-fall gap creation and positive height differences made by pioneer species rapidly growing in the previously existing forest gaps. However, due to the large dispersion observed between CHM models, other factors such as i. crown delineation inconsistencies created by individual tree growth or absence of returns to computed tree height in a cell (in particular for the lower density aircraft data); ii. differences in flight and LiDAR sensors specifications (e.g., with a much higher number of returns per area, the GatorEye data resulted in a higher number of heights in the cells, generating a better chance to identify the highest part of the treetop); iii. phenological phases differences (e.g., flowering or leaves fall); and iv. crown movement by wind [
14] also contributed to the observed differences between models.
The AGB model produced by the aircraft system presented a higher R2 value and lower RMSE. In similar studies, despite technical differences in the LiDAR data acquisition, the AGB models produced by UAV-borne LiDAR (e.g., [
41]) had similar accuracy to those produced by aircraft systems (e.g., [
8,
12]). Other factors, such as the quality of fieldwork on the establishment, measurement, and geolocation of the ground plots, also can strongly affect the accuracy of the models. Still, in this study, the models were developed from the same PSP data and geolocation.
All AGB models produced presented high accuracy, and the uncertainty at 1 ha resolution was lower or comparable with values previously reported in other tropical forests for both aircraft and UAV-based LiDAR systems [
5,
13,
41]. The inclusion of more LiDAR independent variables produced models with higher R2 statistics. The use of more than two independent variables to compose a LiDAR AGB model is considered acceptable (e.g., [
42]); however, to avoid highly collinear LiDAR metrics that would limit the model’s predictive usefulness over the range of forest structure condition in the study site [
7], we only accepted models with VIF below 5 and Pearson correlation test below 0.7. In our study, the best model to the aircraft was a combination of two LiDAR metrics representing the forest higher stratum (Elev P95) and the elevation mode (Elev mode) [
43] and to GatorEye an univariate model based only in the 90th percentile above ground height. Correlation coefficients above 0.8 are not typical in tropical forests, but we found a study [
37] where, in similar conditions, the authors found a very high accuracy in a model built with three independent variables. In our study, we built models for both GatorEye and Aircraft with Adj. R
2 > 0.9 with acceptable VIF (below 5), but we discarded these models because they present Pearson correlation > 0.7.
The AGB models produced by both systems were highly correlated; however, there was no 1:1 statistical correspondence. Therefore, both systems proved to be equally efficient for estimating ABG, but the models of both systems were not statistically equal. Although the average biomass estimated at the landscape scale was very similar, the reduced statistical correspondence was likely due to the canopy structure changing during the two-year difference between LiDAR data collections. We believe that studies using data collected in the same period may achieve the statistical correspondence match between the models from the two systems.
One current use of the AGB models produced by LiDAR is through the use of satellite images to upscale AGB estimates to a regional scale. Usually, when large LiDAR samples are available, a first upscale is performed from the field plot scale to LiDAR sampled area [
44,
45]. Here, we tested the possibility to upscale the AGB model produced by GatorEye. AGB LiDAR models can be, when functioning correctly, generalized [
46] or applied in different regions [
42]. In our extrapolation of both models to the entire area covered by the 2015 LiDAR flight, the AGB maps produced presented a surprisingly high correlation, as both models used the same LiDAR metrics as independent variables, and these metrics were unaffected by the difference in cloud point density (see
Figure 3). When the same procedure was performed with the GatorEye univariate regression model, although the model correlation was still high, there was a tendency for the model to slightly underestimate AGB values.