COMPARISON OF LOW-COST COMMERCIAL UNPILOTED DIGITAL AERIAL PHOTOGRAMMETRY TO AIRBORNE LASER SCANNING ACROSS MULTIPLE FOREST TYPES IN CALIFORNIA

Science-based forest management requires quantitative information about forest attributes traditionally collected via sampled field plots in a forest inventory program. Remote sensing tools, such as active three-dimensional (3D) Light Detection and Ranging (lidar), are increasingly utilized to supplement and even replace field-based forest inventories. However, lidar remains cost prohibitive for smaller areas and repeat measurement, often limiting its use to single acquisitions of large contiguous areas. Recent advancements in unpiloted aerial systems (UAS), digital aerial photogrammetry (DAP) and high precision global positioning systems (HPGPS) have the potential to provide low-cost time and place flexible 3D data to support forest inventory and monitoring. The primary objective of this research was to assess the ability of low-cost commercial off the shelf UAS DAP and HPGPS to create accurate 3D data and predictions of key forest attributes, as compared to both lidar and field observations, in a wide range of forest conditions in California, USA. A secondary objective was to assess the accuracy of nadir vs. off-nadir UAS DAP, to determine if oblique imagery provides more accurate 3D data and forest attribute predictions. UAS DAP digital terrain models were comparable to lidar across sites and nadir vs. off-nadir imagery collection, although

programs that provide timely and verifiable information on forest conditions (i.e. canopy cover, stand height, biomass, etc.). Traditionally, forest inventory and monitoring programs use field plots with detailed measurements of forest composition and structure, from which sample-based estimates are calculated (Bechtold and Patterson, 2005;Gillis et al., 2005;Tomppo et al., 2010). However, incomplete spatial coverage and lengthy re-measurement intervals can limit the effectiveness of field plots in quantifying forest change and providing timely estimates of forest conditions, especially for remote unmanaged regions and small areas, both of which often lack adequate plot sampling to support traditional sample-based estimation (Rao, 2017;Wulder et al., 2004).
For large-scale regional and national objectives, sample-based field inventories are often integrated with remotely sensed data such as multispectral satellite imagery to generate spatially complete estimates of forest conditions (Ohmann and Gregory, 2002;Tomppo et al., 2008;Wilson et al., 2013). Imagery from the Landsat and Sentinel 2 missions are especially attractive for integration with forest inventory programs, due to their spectral and spatial compatibility with many vegetation attributes, open imagery archives, global coverage, and frequent repeat cycle (Drusch et al., 2012;Kennedy et al., 2014;Wulder et al., 2012a). However, passive optical sensors have known saturation and sensitivity limitations (Lu, 2006;Turner et al., 1999), posing problems for predicting attributes such as biomass, stand density and vertical forest structure (Eskelson et al., 2012;Pierce et al., 2009;Zald et al., 2014).
Compared to passive optical sensors, light detection and ranging (lidar) is well suited to characterize the three-dimensional structure of forests (Dubayah and Drake, 2000;Lefsky et al., 2002;Reutebuch et al., 2005). Lidar is increasingly integrated with samplebased forest inventory plots to generate spatially complete estimates of forest conditions, as well as a sampling tool for large-area estimation (Andersen et al., 2012;Wulder et al., 2012b). Despite declining costs, airborne lidar is still only cost effective for large continuous areas or strip sampling, limiting its applicability when frequent repeat data is required, or for small forest parcels and landowners for which lidar acquisition is cost prohibitive.
An emerging alternative to lidar is three-dimensional (3D) data derived from digital aerial photogrammetry (DAP). 3D DAP (also known as Structure from Motion (SfM), and colloquially as "phodar") uses overlapping images from a passive optical sensor to calculate a point's position in space (Ota et al., 2015;Shin et al., 2018a;Swetnam et al., 2018). 3D DAP applications include large-scale integration with sample-based inventory plots (Strunk et al., 2019) as well as small-scale prediction using 3D DAP collected from unpiloted aerial systems (UAS) (Iizuka et al., 2018;Puliti et al., 2015;Swetnam et al., 2018). Due to its ability to acquire highly flexible user-defined acquisition locations and frequency, UAS DAP is especially attractive for small landowners, photo plots in a samplebased inventory, and frequent remeasurement.
Despite the potential of UAS DAP, there remain multiple issues to address for it to become a broadly useful tool for forest inventory and monitoring. The majority of studies using UAS DAP have relied on expensive survey grade UAS platforms carrying fixed highresolution cameras and associated high precision global positioning systems (HPGPS) (Alonzo et al., 2018;Iizuka et al., 2018;Shin et al., 2018b). HPGPS has been a prerequisite for UAS DAP to ensure accurate georeferencing of imagery and co-registration with field plots and other geospatial data sources. Low-cost commercial-grade UAS with highresolution optical sensors and integrated GPS systems are now available in small and affordable all-in-one solutions capable of conducting aerial surveys out of the box.
However, DAP models that are both correctly scaled and spatially accurate require the addition of individual photo locations (Bryson et al., 2010). Utilization of HPGPS systems, either within a survey grade UAS that creates accurately geotagged photos, or in the acquisition of ground control points (GCPs) placed throughout the study area, can greatly increase the spatial accuracy of DAP models (Sanz-Ablanedo et al., 2018). Traditionally, HPGPS systems can be both expensive to purchase and technical in their use, leaving them out of reach for small scale projects, however, the recent introduction of significantly less expensive HPGPS, that have both a user-friendly interface and a growing online support community, has increased the access to this technology. The use of these more-affordable HPGPS systems in UAS DAP studies are needed to show the benefits of their use while also helping understand their potential limitations. Specifically, integration of HPGPS GCPs with low-cost UAS could provide easy to acquire 3D DAP for a fraction of the cost of survey grade systems.
Additionally, many studies using UAS DAP for obtaining forest data have focused on individual forest types (Fankhauser et al., 2018;Navarro et al., 2020;Puliti et al., 2015;Shin et al., 2018b;Wallace et al., 2016). The absence of an assessment of DAP across a wide range of forest types and conditions has hindered the development of best practices and the widespread application of UAS DAP in forest inventory and monitoring.
Furthermore, standard practices of UAS DAP data acquisition typically include only collecting images at nadir, yet multi-angle DAP has potential in improving characterization of vertical forest structure (Fankhauser et al., 2018). By including off-nadir imagery in a DAP dataset, the image sensor has an increased view of the sides of the forest canopy, allowing the photogrammetry algorithm the ability to create a more "complete" model of the whole canopy than with nadir imagery alone. An example of this can be seen in Figure   1, where the 3D DAP model generated with multi-angle imagery includes more of the lower tree canopy than nadir imagery alone. In combination, low-cost commercial-grade UAS, low-cost and user friendly HPGPS, and multi-angle (on and off-nadir) imagery have the potential for UAS DAP to be an affordable alternative to lidar for forest inventory but have been largely unexplored.  that UAS DAP will show less accuracy in more dense forest canopies, where the terrain beneath the canopy has increased occlusion from the image sensor of the UAS. It is also expected that the addition of off-nadir imagery would consistently increase the accuracy of UAS DAP due to its potential to better more completely characterize forest canopy structure.

Study Area
Study sites were selected based on availability of recent lidar in the State of California, desire to assess UAS DAP across a wide range of forest types and structural conditions, and access of property for research activities. Study sites required available lidar data collected within two years prior to UAS imagery acquisition, restricting the wide of range locations to those with available lidar data collected during 2017-2019.
Sites also had to cover a range of forest conditions, including conifer and hardwood dominated sites, stand ages, and varying levels of canopy structural complexity. Lastly, sites had to be accessible, with landowners giving permission for research activities that    GPS. This GPS data is utilized to assist in the image alignment process, but due to the low accuracy of the internal GPS, UAS images were georectified using ground control points (GCPs, 12-inch-wide black and white tiles) placed throughout the flight area, with their coordinates precisely measured with an HPGPS. Once the initial image alignment was completed, the UAS GPS data was no longer referenced and only the GCPs were used to georeference the generated models. A dense point cloud was then generated using the "High" accuracy setting allowing the use of each image's full resolution when locating matching points between photos. This resulted in a high-density point cloud of approximately 250 pts/m 2 . Due to high computational requirements associated with such high point cloud densities, point clouds were filtered to a voxel spacing of 0.05 m 3 between points, resulting in densities of approximately 50 pts/m 2 without degrading structural characteristics in the point clouds.

Lidar and DAP Point Cloud Processing
Lidar and DAP derived point clouds were processed using the lidR package (Roussel et al., 2020) in R (R Development Core Team 2020). All point clouds were clipped to the boundaries of the sites plus a 20 m buffer to avoid edge effects during processing. Ground points were classified using the cloth simulation filter (csf) algorithm (Zhang et al., 2016). Digital terrain models (DTMs) were generated from classified ground points using a Delaunay triangulation algorithm (Kim and Cho, 2019). Digital surface model (DSMs) were generated using the pitree algorithm (Khosravipour et al. 2014). Point clouds were then normalized using the generated DTM's made for the given model at each site. Canopy height models (CHMs) were then generated from the normalized point clouds using the same pitree algorithm as for the DSMs. All DSMs, DTMs, and CHMs were regenerated as rasters with a 1 m resolution.
For predicting forest attributes, the normalized point clouds were clipped to the same 0.05 hectare fixed-area plots previously described in Field Data section above.

Statistical Analyses
All statistical analyses were conducted in R (R Development Core Team 2020).
For each gridded surface model (DTM, DSM, and CHM) we compared both UAS DAPderived (nadir and off-nadir) pixel values for the model against the lidar-derived version for each site separately. In this study, the use of the term accuracy is utilized when describing how close models and predictions from UAS DAP compare to predictions and observations made by the collection methods most commonly used in forest inventory.
For 3D data products, such as digital surface models, lidar is the most common source whereas ground collected field data from fixed or variable radius plots are used in the collection of forest structural attributes, such as tree heights and basal area. To determine how accurately the UAS DAP predictions and models were to these standard method of collecting similar data types this study followed many of the accuracy assessment protocols for continuous variables as described by Riemann et al., 2010. R-squared, root mean square deviation (RMSD), normalized RMSD (nRMSD), agreement coefficient (AC), systematic agreement coefficient (AC sys ) and unsystematic agreement coefficient (AC uns ) were calculated between UAS DAP and lidar derived surface models.
For prediction of forest attributes, plot values of AGB, BAH, TPH, QMD, and LHT were predicted using the structural metrics from lidar and DAP models for each plot. This was accomplished using a linear regression for each forest attribute for all plots across sites. The forest attribute variables were checked for linearity and normality using histograms and Q-Q plots, resulting in ABG data being cube root transformed, while BAH, TPH, QMD, and LHT were square root transforms. The leaps package in R (Lumley and Miller, 2020) was used to determine the best subsets of predictor variables for regression models for each forest attribute. Following the rule of thumb of no more than one predictor variable per 20 sample units, the maximum number of predictor variables in a model was set to five, and the best model from candidate models was determined by adjusted R 2 and BIC values. These models were then used to predict plotlevel AGB, TPH, QMD, and LHT using the derived structural metrics from lidar and DAP models for each plot at each site using leave one out cross validation (LOOCV) with the caret package (Kuhn, 2020). Cross-validated predictions of plot-level forest attributes made by each best fit model were compared to that of observed forest attributes.

Accuracy of UAS DAP Surface Models
Digital terrain models (DTMs) derived from DAP displayed high correlation

Accuracy of UAS Forest Attribute Predictions
Best-fit plot-level regression models show that both DAP and lidar point cloud metrics can be used to accurately predict forest structural metrics (Table 4 and Figure 6).
DAP models had similar prediction accuracy to lidar-based predictions, with the exception of the off-nadir DAP models prediction of QMD (R 2 = 0.45). The strongest predictor variables used when estimating AGB, TPH, BAH and LHT tended to be zmean, zsd, and pzabove2 while models predicting QMD tended to rely on quartile metrics. All model predictions showed moderate to high correlation (R 2 = 0.53 -0.84) to observed values of AGB, THP, BAH, and LHT. Overall, regression models using lidar derived predictor variables were more accurate than models of the same response variables using DAP derived predictors, with the exception of the model of QMD using nadir DAP predictor variables (R 2 = 0.70). DAP models containing off-nadir images (both off-nadir and multi-angled models) tended to have marginally higher correlation values and marginally lower RMSD and nRMSD values, compared to nadir based models.
When comparing predicted forest attributes for all sites between the different remote sensing models (ie. the different DAP models against lidar), all DAP predictions were highly correlated with lidar predictions (Table 5). DAP models containing only offnadir images were shown to have the highest correlations with lidar-based predictions (r 2 = 0.75 -0.85), with the exception of QMD predictions. Multi-angle models containing both nadir and off-nadir images had slightly poorer performance when compared to lidar, with r 2 values ranging from 0.66 to 0.83.

UAS Surface Models
In previous studies, UAS DAP models were aided by supplementing previously collected elevation datasets, such as lidar generated terrain models, to normalize the DAP point clouds and create CHMs (Dandois et al., 2015;Fankhauser et al., 2018;Iglhaut et al., 2019;Jayathunga et al., 2018;Puliti, 2017;Strunk et al., 2019). This, however, means that the low-cost DAP still required a much higher cost lidar acquisition, leaving this method out of reach for smaller studies sites and small landowners who have never had the ability to acquire lidar data. In this study, surface models generated from UAS DAP were found to have high levels of agreement when compared to those generated from lidar, with decreased levels of accuracy in DSMs, DTMs and CHMs at sites DF and HC, whose forests had few canopy gaps, resulting in the occlusion of the terrain from the UAS passive optical sensor, leaving larger data gaps with increasing canopy density above the modeled terrain (an important element when generating accurate CHMs). This is supported by visualizations of the UAS DAP and lidar point clouds in Figure 7, where there were fewer classified ground points in locations with dense canopies, consistent with observations made in previous studies (Belmonte et al., 2020;Dandois and Ellis, 2010;Wallace et al., 2016). This became most problematic with off-nadir DAP models with dense canopies as can be seen at the DF site, where due to heavy occlusion, there were not enough matched ground points to accurately model the terrain, resulted in an R 2 value of 0.17 when compared to the terrain model generated from lidar ( Figure 4).

Limitations
The agreement between UAS DAP and lidar CHMs and DSMs showed poorer results (Figure 3 and Figure 5). This may be caused by slight shifts in modeled vegetation location rather than missing or erroneous values. While coupling our UAS DAP models with low-cost HPGPS to increase spatial accuracy of our models, high canopy cover and density increased the amount of HPGPS error when averaging the location of the ground control points. This, and the distances between the control points themselves, may have shifted the UAS DAP coordinates, lowering the overall agreement between 1 m pixels.
Another source of potential error in CHMs and DSMs can be seen in UAS DAP modeling of vegetation in sites with open canopy (OW, MC and MCtb). It was expected that these sites would perform well given better performance of passive optical sensors with reduced canopy cover and terrain occlusion. However, DAP models of these sites had missing canopy structural data (Figure 1 and Figure 6). One cause for this could be the use of aggressive depth filtering in photogrammetric processing to remove outlier point observations resulting from poor imagery or bad alignment issues. Recent research suggests the use of aggressive depth filtering in the generation of UAS DAP point clouds may lead to the filtering of segments of the forest canopy as noise and that lower depth filtering settings should be used when modeling forest canopies, allowing the point clouds to contain more detail (Tinkham and Swayze, 2021).
In this study, lidar outperformed UAS DAP when predicting plot-level forest attributes. Lidar is the preferred remotely sensed data for characterizing ground terrain and vertical forest structure in support of forest inventory and monitoring. However, airborne lidar data is cost prohibitive for small areas and frequent data collection and based on the results of this study, low-cost UAS DAP can generate similar data products to lidar in a less expensive, flexible, and rapidly deployable manner. Land managers can utilize UAS DAP in forest inventory and monitoring to generate high resolution imagery, 3D models, and forest attribute prediction without the need for previously collected DTMs from lidar. UAS DAP has also been shown to have limitations due to its use of passive RGB imagery. DTMs generated from UAS DAP in dense, closed canopy forest conditions, such as the DF and HC sites found in this study, show lower levels of agreement than in sites with more open conditions. Although we show that DAP can make accurate predictions of forest attributes, the spatial variation and bias in DAP surface models at all sites suggests that UAS DAP should not be used when doing pixel-to-pixel level change detection from repeated measurements. It also suggests that utilizing lidar generated DTMs when normalizing UAS DAP DSMs in dense canopy conditions might still be necessary to create an accurate CHM. Also, current methods of using lidar and UAS DAP point clouds in forest inventory lack the ability to determine species level data.
Photogrammetric point clouds can, however, integrate spectral information from the image sensors into the generated point clouds. This added spectral information could allow for additional predictive power in UAS DAP by allowing it to make predictions on species and forest health as well as forest structural attributes, but more research is needed in the use of multi and hyper-spectral UAS DAP in forest inventory (Iglhaut et al., 2019).

CONCLUSION
This study demonstrates that low-cost, commercial-grade, UAS DAP coupled with new-to-market, low-cost HPGPS can generate comparable data products and predictions to lidar and in-field observations of forest attributes across a wide range of forest sites and conditions. The addition of off-nadir imagery into UAS DAP models only marginally affects the accuracy of surface model and forest attribute predictions.
Comparisons of UAS DAP versus lidar based surface models indicates that the need for previously acquired lidar terrain models may not be necessary to achieve accurate CHMs from photogrammetry models, and for all forest types UAS DAP generates predictions of forest attributes comparable to lidar.
This study shows that UAS DAP can be both an affordable and accurate remote sensing tool in forest inventorying and monitoring, and that forest managers should consider the structural characteristics of the forest of interest when determining whether to include off-nadir images in their UAS data acquisition. For use in continuous forest inventory and monitoring programs, UAS DAP can make accurate predictions of forest stand metrics, however, it may not have the spatial accuracy to make direct comparisons of generated surface models between data collection periods depending on forest canopy type. The research presented here shows that low-cost UAS, when combined with lowcost HPGPS, can be an accurate and affordable alternative to lidar in forest inventories, increasing access to high quality spatial information that can lead to cheaper and more informed management decisions.