How Much Can We See from a UAV-Mounted Regular Camera? Remote Sensing-Based Estimation of Forest Attributes in South American Native Forests

: Data collection from large areas of native forests poses a challenge. The present study aims at assessing the use of UAV for forest inventory on native forests in Southern Chile, and seeks to retrieve both stand and tree level attributes from forest canopy data. Data were collected from 14 plots (45 × 45 m) established at four locations representing unmanaged Chilean temperate forests: seven plots on secondary forests and seven plots on old-growth forests, including a total of 17 different native species. The imagery was captured using a ﬁxed-wing airframe equipped with a regular RGB camera. We used the structure from motion and digital aerial photogrammetry techniques for data processing and combined machine learning methods based on boosted regression trees and mixed models. In total, 2136 trees were measured on the ground, from which 858 trees were visualized from the UAV imagery of the canopy, ranging from 26% to 88% of the measured trees in the ﬁeld (mean = 45.7%, SD = 17.3), which represented between 70.6% and 96% of the total basal area of the plots (mean = 80.28%, SD = 7.7). Individual-tree diameter models based on remote sensing data were constructed with R 2 = 0.85 and R 2 = 0.66 based on BRT and mixed models, respectively. We found a strong relationship between canopy and ground data; however, we suggest that the best alternative was combining the use of both ﬁeld-based and remotely sensed methods to achieve high accuracy estimations, particularly in complex structure forests (e.g., old-growth forests). Field inventories and UAV surveys provide accurate information at local scales and allow validation of large-scale applications of satellite imagery. Finally, in the future, increasing the accuracy of aerial surveys and monitoring is necessary to advance the development of local and regional allometric crown and DBH equations at the species level.


Introduction
Native forests cover a large percentage of land in South America [1], with vast areas of old-growth and secondary forests [2] which are playing an essential ecological role

UAV Data Acquisition and Image Processing
The image was captured using a Bormatec Maja fixed-wing airframe equipped with an ArduPilot flight controller APM2 from 3DR company and GPS data logger, pressure and temperature sensor, airspeed sensor, triple-axis gyro, and accelerometer [34]. The GPS used by APM2 is a navigation system with an expected position error of a few meters that was also used for geotagging the image. The precise position of the mosaic was made by using 4 ground control points (using 1 m 2 flags on plots), which served to match the relative position of each tree. We used a regular RGB photographic camera Canon S100, with a built-in GPS. The camera was within the payload capacity of the MAJA system, and the time intervals for photograph retrieval could be easily changed. The sensor size was 1/1.7" (~7.53 × 5.64 mm), the pixel count was 12.1 MP, the pixel pitch was 1.87 µm, and the focal length was 50 mm. The plan was programmed in the software Mission Planner and took place in the summer of 2015. We followed an orthophoto image flight plan, flying to 150 m above the forest canopy in order to have a good pixel resolution. The aim was to have at least an 80% overlap sideward at terrain level, which resulted in images of very high spatial resolution (range of 5.6-6.6 cm) for each plot. The point cloud, digital surface model, and image mosaicking was performed using the software Pix4Dmapper (v 1.4.28). Each tree crown was identified on the ground using the georeferenced mosaic for each plot and the position of each tree. Finally, tree crown segmentation was manually performed and the ground data was linked ( Figure 1).

Statistical Analysis
The data analysis was performed at the stand and tree levels. For the stand level, the response variables were tree density (RD, trees ha −1 ), quadratic mean diameter (QMD, cm), and basal area (BA, m 2 ). The canopy structure variables used as predictors were the total forest canopy area, mean tree crown area, and standard deviation of the trees' crown area. The effect of the environmental variables on the response variables was assessed using a linear regression model.
For the tree level, the response was the individual-tree DBH, and the predictors were the crown planar area (m 2 ), crown curve area (m 2 ) and crown volume (m 3 ) estimated by the point cloud and digital surface model for each identified tree from the images (Figure 1). At this point, two scenarios were considered, i.e., one assuming information available concerning the tree species and another using crown-derived information irrespective of the species. The exploratory approach included the use of boosted regression trees (BRT), combining machine learning, and statistical techniques, and consisted of adding regression trees fitted in a forward stagewise process in order to improve the model accuracy [38,39]. This approach starts with the creation of a single tree minimizing the loss function, using the residuals to add a new tree, and progressively, the first estimates are updated to include the information of the new tree, and the resulting residuals are recursively used to fit a new additional tree. The final model reflects the contribution of all trees and provides an estimate of the contribution of each variable to the model. The model construction requires the following four parameters: the learning rate (i.e., the contribution of each new tree to the growth model, in this case, tested around the value 0.01); the bag fraction (i.e., the information on which fraction of the whole data should be drawn randomly to fit the new tree to avoid over-fitting, see [35], in this case fixed to be 0.80); the tree complexity (i.e., the number of interactions among variables, in this case tested iteratively from 1 to 4, aiming for a parsimonious model); and the number of trees required for optimal prediction (which was optimized based on the three previous parameters, see [36]). The statistics on predictive performance were estimated from the subsets of data excluded from model fitting, using cross-validation (RMSE, R 2 ). The statistical analyses were developed in the R version 3.5.3 [40] and the BRT models were based on the "dismo" extension of the "gbm" package [41].
On the basis of the above analyses, the variables showing the largest contribution were further studied in a mixed model approach, using the species as the random grouping factor. As in the BRT case, the species is a variable that cannot be easily identified from the remote sensing data but is otherwise used to confirm the between-species variability in the predictions. In this case, the model parameters were fitted based on restricted maximum likelihood by using the "nlme" package [42].

Assessment of Forest Attributes Observed from a UAV in Old-Growth and Secondary Forests
There were 17 tree species represented in the sample plots; A. punctatum, P. lingue, N. obliqua, and E. cordifolia were the most common species. The measurements describe the secondary forest sampled (Table 1, Figure 2) as dominated by N. obliqua presenting a median quadratic diameter between 24.5 and 42.1 cm, a BA around 65.2 m 2 (SD = 8.7), and a density of around 916 trees ha 1 (SD = 280). The old-growth forests were mostly dominated by A. punctatum and A. Araucana, with a median quadratic diameter between 24.5 and 74.2 cm, a mean BA around 74.3 m 2 (SD = 14.5), and a lower density of around 590 trees ha −1 (SD = 303).
In total, 2136 trees were measured on the ground, from which 858 trees were visualized from the UAV imagery of the canopy, with a range from 26% (Plot d) to 88% (plot b) of the measured trees (mean = 45.7%, SD = 17.3) (Figure 3). There were large differences concerning tree density and the spatial distribution of the tree diameters, which precluded the detection of certain trees ( Figure 4). All in all, trees included in the UAV image mosaic represented between 70.6% (plot g) and 96% (plot b) of the total BA of the plots (mean = 80.28%, SD = 7.7); on highly diverse plots and secondary forests, the ratios were lower, ranging from 18.2% (plot v) to 88.9% (mean = 59.1%, SD = 23.5). Concerning tree species richness, when there were fewer than five tree species, all of them were represented in the UAV processed data. In total, 2136 trees were measured on the ground, from which 858 trees were visualized from the UAV imagery of the canopy, with a range from 26% (Plot d) to 88% (plot b) of the measured trees (mean = 45.7%, SD = 17.3) (Figure 3). There were large differences concerning tree density and the spatial distribution of the tree diameters, which precluded the detection of certain trees ( Figure 4). All in all, trees included in the UAV image mosaic represented between 70.6% (plot g) and 96% (plot b) of the total BA of the plots (mean = 80.28%, SD = 7.7); on highly diverse plots and secondary forests, the ratios were lower, ranging from 18.2% (plot v) to 88.9% (mean = 59.1%, SD = 23.5). Concerning tree species richness, when there were fewer than five tree species, all of them were represented in the UAV processed data.    . Spatial distribution of the tree diameters (DBH, ground data) in four of the studied plots, representing old-growth (corresponding to plots b and g) and secondary forests (plots y and x).
The plots selected denote the highest and lowest tree density in each category. The size of the circle represents tree diameters and green circles represent trees effectively detected using unmanned aerial vehicle.  plots y and x). The plots selected denote the highest and lowest tree density in each category. The size of the circle represents tree diameters and green circles represent trees effectively detected using unmanned aerial vehicle.

Diameter Classes Represented in the Remote Sensing Data
Some plots with A. araucana were nearly monospecific (90% of BA) and were excluded from the calculations concerning the analysis of forest variables by diameter class. For the rest, the diameter classes were defined every 10 cm. The BA effectively retrieved from the UAV data was 83.6% and 83.3% on old-growth and young forests, respectively, in particular, it exceeded 70% on diameter classes over 35 and 25 cm, respectively. About 37.9% and 48.5% of individual trees were detected on old-growth and young forests, respectively, rising to over 50% for diameters over 35 and 25 cm, respectively ( Figure 5).

Stand Structure Variables
Considering the individual-tree crown data, there was a relationship between the measured QMD and the UAV-retrieved BA and density. On the one hand, for QMD, it was a positive correlation; therefore, for a higher QMD there were more trees detected (R 2 = 0.71) and retrieved BA (R 2 = 0.47). Density, on the other hand, presented a negative correlation; therefore, the sparser the forest, the higher the proportion of trees (R 2 = 0.57) and BA (R 2 = 0.37) detected by the UAV images ( Figure 6). excluded from the calculations concerning the analysis of forest variables by diameter class. For the rest, the diameter classes were defined every 10 cm. The BA effectively retrieved from the UAV data was 83.6% and 83.3% on old-growth and young forests, respectively, in particular, it exceeded 70% on diameter classes over 35 and 25 cm, respectively. About 37.9% and 48.5% of individual trees were detected on old-growth and young forests, respectively, rising to over 50% for diameters over 35 and 25 cm, respectively ( Figure 5).   Table 3.
Significant relationships among the stand variables and the canopy structure estimates from the UAV imagery were also found. The relative forest density was related to the total canopy surface and inversely related to the standard deviation of the tree crown's surface (R 2 = 0.73, p-value < 0.05). A similar pattern was observed regarding the QMD (R 2 = 0.73, p-value < 0.001). There were no observed linear relationships between the BA and the canopy structures retrieved ( Table 2).

Individual-Tree DBH
For the individual-tree DBH, the BRT approach estimated the between-species variability to be ca. 17% of the total, and the main variable explaining tree diameter was the tree crown area, accounting for 68.9%. Other variables tested were below 5% and were not considered at this stage. The BRT model had a high predictive power (R 2 = 0.78, based on CV, R 2 = 0.69). These results were further studied on a mixed model, using the tree crown area as a predictor in the fixed part of the model, and the species as the random factor (Figure 7). The predictive power of the model was R 2 = 0.55 for the fixed part of the model and R 2 = 0.66 for the whole model, including between-species variability ( Table 3). The between-species variability was expressed in the intercept using a random parameter (Table 4); these values were related to their corresponding marginal effects on the BRT's estimates.  In the mixed model, the between-species standard deviation was 6.81 for an estimated residual deviation of 12.06. The main contributions for the BRT model were crown area (68.9%), species (16.9%), and crown volume (6.97%), the others being below 5%.

Discussion
In this study, we use a UAV-DAP approach to measure basic forest attributes in secondary and old-growth native forests, including individual-tree DBH based on these attributes. Despite the complexity of native Chilean temperate forests, the results demonstrate the potential of this approach, and confirm that the approach is feasible for retrieving reliable estimates to be used in forest assessment, management, and planning. At the same time, the study identified and addressed some important limitations. In general terms, old-growth and secondary forests differ in species composition, stand age, trees basal area, density, DBH distribution and range, presence of snags, and the heterogeneity of canopy structure [43,44]. The findings reflect that, in young dense forests, the results underestimate tree density, although they reach a reasonable accuracy concerning stand basal area. This is a direct consequence of using a UAV, which obtains better estimates of the canopy visible from above, when applied to these particular stand dynamics. With increasing QMD, the first cohorts of pioneer trees are suppressed by larger trees, concentrating on most of the BA and the canopy [43,45].
At the stand level, the results show that the total forest canopy area is related to the relative density and QMD but not to the BA (p-value < 0.001). We also found a relationship between the SD of the tree crown area and relative density (p-value < 0.05), which, in this case, could help differentiate a secondary forest from an old-growth forest. On the one hand, higher forest canopy area and less SD in tree crowns are more likely attributed to a secondary forest canopy structure due to higher competition for light in younger forests. On the other hand, old-growth or mature forests are more related to a lower canopy cover and high SD of tree crowns due to gap dynamics and uneven-aged characteristics (high variability of crown sizes) [46,47]. A positive relationship between the total forest canopy area and basal area was to be expected, although it was not significant, possibly due to a high accumulation of biomass in large trees, species-based canopy closure strategies, stand development stage, structure, composition, and environmental conditions [48]. This shows that similar forest canopy areas can result in largely different stand BAs, stressing the need to better understand alternative stand characteristics related to BA in order to develop robust models. In addition, the poor response of crown volume to diameter in the models can also be explained by the fact that this variable is notoriously difficult to establish from DAP in dense forests, as the underside of the canopy is usually not well represented in a photogrammetric point cloud (as opposed to ALS). As expected, the results suggest that the use of a UAV-DAP approach presents more limitations on forests with larger diversity of tree species and a complex structure, which are, at the same time, areas subject to special conservation objectives. Moreover, for habitat or ecosystem mapping, direct mapping of individual plants and species in diverse forests is both necessary and complex [49].
The UAV-DAP approach can provide large amounts of data concerning tree canopies, through deriving 3D photogrammetric point clouds from very high overlapping UAV images (>85% overlap); however, some information concerning tree species diversity, forest structure, and understory vegetation is not easily retrieved, particularly in dense multistory tropical forests and diverse dense old-growth forests. The results quantify to a certain extent this loss, which is larger in a forest with a greater complexity in terms of tree diversity and structure [50,51], and show that canopy-based estimates achieve an average of 45% of the trees representing over 80% of the total basal area. Furthermore, we observed high variability in the species dominating the canopy versus the total species richness, with a tendency to underestimate the species, particularly on highly diverse forests. Old-growth forests under this broad definition often have a complex structure and a heterogeneous spatial arrangement that varies depending on the forest type [52,53]. As forests become largely logged and maintained in young successional stages in many areas, there is a renewed interest in the unique structural and functional characteristics of old-growth forests because of their diminishing land cover, high relevance for the conservation of regional biodiversity, and their value for global carbon storage [54].
These unique services justify efforts for an early warning detection of forest degradation [55,56]. The results confirm that changes in canopy attributes can be easily retrieved and could be used as early warnings of forest degradation detection, as several approaches defining forest degradation are linked to the loss of the main forest attributes, i.e., composition, structure, and function [57]. Forest degradation is also associated with an impact on functional forest processes, including variables such as shoot growth in woody plants, seed availability in soil seed banks, abundance of seedlings, and the age structure of common species (as compared with that expected based on the successional state of the forest and the regeneration strategy of the various species) [58]. In this context, the use of a UAV represents an alternative for detecting forest degradation processes, especially in old-growth forests. Whereas some studies suggest that forest degradation has several dimensions that are difficult to represent in a single measure, other studies have proposed grouping indicators according to the forest attributes that have been affected by degradation [57,59]. These indicators include loss of canopy cover and reduction in structural complexity, potential gaps in some diameter classes over DBH 60 cm, and a low frequency of individuals in the intermediate diameter classes [57,59].
Finally, the combination of the modeling approaches presented offers both the advantages of machine learning in exploratory analysis, variable selection, and model structure, as well as the replicability and applicability of the more conventional regression methods [60,61]; the results of both approaches were consistent and confirm the use of a UAV as a promising alternative for inventory in native forests, providing an accurate estimation of aboveground biomass concentrated in larger trees (as those are the main economic interest and often largely determine local forest management). Despite the loss of accuracy in forests with large diversity and structure [56,62], the predictions of DBH at the tree level are enough to be the basis for quantifications of total biomass and tree fractions, which has important applications in carbon assessments and forest management.

Conclusions
The use of a UAV to assess and monitor forest management in Chilean temperate forests in terms of both tree species diversity and forest structure is needed at the local and regional levels to understand the impacts of interventions and to develop effective management strategies for their conservation. In this study, we analyzed the use of a UAV and field-based data methods with the potential to achieve these objectives. For most studies, the availability of resources (time, money, and manpower) is the major constraint. We suggest combining the use of both field-based and remotely sensed methods to achieve that goal as these methods can be complementary, in particular in old-growth forests. Remote-sensing data should be used to predict and map the tree species diversity and stand structure at regional scales, while field inventories provide accurate information at local scales and allow validation of remotely sensed data. In the future, fields such as forest dynamics, forest species dominance in stands, mapping, and assessing forest disturbances will grow considerably with the benefits of unmanned aircraft technology. Our results suggest that drone-derived canopy variables contributed substantially to explaining patterns of structure and biodiversity in these temperate forest plots. This study provides immediate guidance for the application of these tools in forest inventory in Chilean temperate forests and describes a methodology that can support the management of their forests according to their structure. Future studies aimed at higher accuracy estimation must advance the following three objectives: (i) better tree level segmentation, (ii) precise metric functions between crown and trunk attributes at a specific forest type level, and (iii) tree-level identification of forest species. In general, the results show that, when the focus is biomass with economic value (larger trees), accurate tree-level DBH can be obtained, especially when some level of species differentiation is applied.

Data Availability Statement:
The data presented in this study are available on request from the corresponding authors.