Next Article in Journal
Forest Fragmentation and Forest Mortality—An In-Depth Systematic Review
Previous Article in Journal
Forest Height and Volume Mapping in Northern Spain with Multi-Source Earth Observation Data: Method and Data Comparison
 
 
Font Type:
Arial Georgia Verdana
Font Size:
Aa Aa Aa
Line Spacing:
Column Width:
Background:
Article

Using Unoccupied Aerial Systems (UAS) and Structure-from-Motion (SfM) to Measure Forest Canopy Cover and Individual Tree Height Metrics in Northern California Forests

1
School of the Environment, San Francisco State University, San Francisco, CA 94132, USA
2
Department of Biology, Sonoma State University, Rohnert Park, CA 94928, USA
*
Author to whom correspondence should be addressed.
Forests 2025, 16(4), 564; https://doi.org/10.3390/f16040564
Submission received: 21 February 2025 / Revised: 17 March 2025 / Accepted: 19 March 2025 / Published: 24 March 2025
(This article belongs to the Section Forest Inventory, Modeling and Remote Sensing)

Abstract

:
Quantifying forest structure to assess changing wildfire risk factors is critical as vulnerable areas require mitigation, management, and resource allocation strategies. Remote sensing offers the opportunity to accurately measure forest attributes without time-intensive field inventory campaigns. Here, we quantified forest canopy cover and individual tree metrics across 44 plots (20 m × 20 m) in oak woodlands and mixed-conifer forests in Northern California using structure-from-motion (SfM) 3D point clouds derived from unoccupied aerial systems (UAS) multispectral imagery. In addition, we compared UAS–SfM estimates with those derived using similar methods applied to Airborne Laser Scanning (ALS) 3D point clouds as well as traditional ground-based measurements. Canopy cover estimates were similar across remote sensing (ALS, UAS-SfM) and ground-based approaches (r2 = 0.79, RMSE = 16.49%). Compared to ground-based approaches, UAS-SfM point clouds allowed for correct detection of 68% of trees and estimated tree heights were significantly correlated (r2 = 0.69, RMSE = 5.1 m). UAS-SfM was not able to estimate canopy base height due to its inability to penetrate dense canopies in these forests. Since canopy cover and individual tree heights were accurately estimated at the plot-scale in this unique bioregion with diverse topography and complex species composition, we recommend UAS-SfM as a viable approach and affordable solution to estimate these critical forest parameters for predictive wildfire modeling.

1. Introduction

The North Coast Region of California, USA, has diverse topography, strong Medi-terranean climatic gradients of temperature and precipitation, and complex species composition. About 1.3 million acres of the region is made up of oak woodlands [1,2,3], which are areas that sustain a high level of biodiversity and serve multiple ecological functions, such as erosion control, landslide prevention, and watershed protection [4]. As a result of a significant increase in wildfire frequency and intensity in this unique bioregion [3], landowners and managers are in the process of implementing vegetation and landscape-specific treatments across the state, including thinning in excessively dense stands, timber harvesting, mechanical fuel reduction, prescribed fire, grazing, and re-forestation [5]. As such, there is a critical need to accurately quantify fine-resolution forest structure data over large areas [6] to assess changing wildfire risk factors as California moves beyond the initial stages of short-term disaster recovery and begins to develop mitigation, land management, and resource allocation strategies. The ability to obtain cost-effective and accurate forest metrics at high spatial resolutions is crucial for informing these efforts, particularly in fire-prone landscapes where vegetation structure influences fire behavior and spread.
Remote sensing offers valuable methods for understanding forest structure and lessening the need for costly field inventory campaigns [7]. Passive optical remote sensing techniques, such as aerial and satellite-based multispectral imagery, rely on reflected sunlight to generate spectral data that can be analyzed for forest structure and vegetation indices, but their ability to retrieve structural attributes is inherently limited by the vertical complexity of forest canopies [7,8,9,10]. However, while these methods can be used for individual tree crown segmentation, canopy cover estimation, and species identification [7,8,9,10], they do not directly capture three-dimensional (3D) forest attributes because they lack active signal emissions, which is necessary for retrieving detailed structural information such as tree height, diameter at breast height (DBH), and canopy base height. Consequently, structural inaccuracies may arise in dense forests where overlapping crowns obscure underlying vegetation, leading to over- or under-estimation of key metrics [11,12].
Rather, these measurements are obtained via light detection and ranging (LiDAR) which uses active remote sensors to yield three-dimension point clouds. LiDAR sensors emit laser pulses and measure their return time to construct detailed forest structure models [11,13,14,15]. ALS-derived tree height estimates generally achieve high accuracy, with root mean square error (RMSE) values ranging from 0.5 m to 2.0 m, whereas UAS-SfM typically exhibits RMSE values between 2.5 m and 5.0 m [11,12]. These discrepancies arise primarily due to UAS-SfM’s reduced ability to penetrate the canopy and limited georeferencing precision compared to ALS. ALS-derived data are typically collected using piloted aircraft but can also be acquired via UAS-mounted LiDAR sensors for localized, high-resolution studies [11,14,15,16,17]. Despite its accuracy, ALS remains cost-prohibitive for frequent monitoring applications, as large-scale ALS missions can cost between $500 and $1200 per km2, making them impractical for routine land management and wildfire risk assessment [11].
Recent technological advancements have led to unoccupied aerial system (UAS) platforms (i.e., drones) offering a higher spatial resolution and more cost-effective options than traditional piloted ALS missions for capturing spectral data that can be analyzed for accurate species identification and canopy cover [7,8,9,10,12]. Structure-from-Motion (SfM) is a photogrammetric method that can create three-dimensional models of a feature of interest (i.e., buildings, vegetation, etc.) by using overlapping two-dimensional images taken from a wide range of angles and perspectives [18]. When UAS images are acquired with high overlap, SfM can be used to produce 3D point clouds similar to LiDAR. SfM is based conceptually on similar principles as stereoscopic photogrammetry where objects generate estimates of the camera position and orientations to return the 3D coordinates of the features [11,19,20,21]. Cost-effectiveness makes UAS-SfM a viable alternative for regional-scale forest monitoring efforts in areas where ALS data are unavailable or outdated.
Traditional forest inventory methods, such as manual field surveys using tape measures and hypsometers, are time-consuming, labor-intensive, and costly, particularly when covering large or remote forested landscapes [11]. ALS missions are expensive and infrequent, often requiring years before public acquisitions become available [12]. In contrast, UAS-SfM offers a more accessible alternative for researchers, landowners, and land managers by providing timely, high-resolution data at a fraction of the cost of traditional ALS surveys. Furthermore, UAS-SfM surveys can be scheduled on demand using fewer personnel, avoiding the long wait times associated with ALS data acquisition, which often occurs at multi-year intervals depending on government funding cycles [12].
UAS-based SfM (UAS-SfM) is an efficient method for gathering forest data because researchers can plan flights over desired areas and do not carry the burden of paying for a more costly piloted ALS mission or waiting for a public ALS acquisition, which can be expensive, infrequent, and often take years from acquisition to publish. UAS-SfM is also cost-effective in that sufficient 3D forest models can be produced using standard consumer-grade cameras [22]. Additionally, higher-grade multispectral cameras can capture calibrated spectral information across the spectrum, allowing the calculation of vegetation indices such as the Normalized Difference Vegetation Index (NDVI) that may be useful for species identification, monitoring canopy stress or phenology, and estimating canopy cover [23]. Also, applications of SfM algorithms for detecting individual trees and deriving tree-level metrics are widely available. For example, Lim et al. (2015) used a camera mounted on a drone to generate a high-resolution ortho-image and normalized digital surface model (nDSM) to extract individual trees through a process of segmentation [24]. Similarly, Li et al. (2012) developed a point-based segmentation algorithm to identify and extract individual trees from point clouds [13]. However, UAS-SfM accuracy can vary based on flight parameters, ground control point (GCP) density, and image overlap, making standardization of methods critical for consistent results across forest types [12].
Given the utility and potential of an UAS-SfM approach for forest structure quantification, our goal was to determine the effectiveness of UAS-SfM for estimating canopy parameters relevant for predictive fire modeling in Northern California oak woodlands. Accurate canopy metrics are critical for understanding wildfire risk, as tools like the Monitoring Trends in Burn Severity (MTBS) program rely on precise structural data to assess fuel loads and fire behavior dynamics [12,25]. As wildfire frequency increases across the western United States, cost-effective forest monitoring tools are needed to improve predictive modeling of fire behavior and support targeted fuel reduction treatments. While there have been previous studies examining the use of UAS-SfM, few studies occur in the North Coast Region of California and include validation data at different spatial scales. Here, by comparing parameter estimates from UAS-SfM with ground-based and ALS data, we evaluated the accuracy of UAS-SfM to (1) measure canopy cover at landscape (1684-ha) and plot (400 m2) scales, and (2) delineate individual tree-tops at a plot scale (400 m2) to calculate tree density (trees/plot), individual tree total height (m) and canopy base height (m). These findings will contribute to the broader understanding of UAS-SfM’s applicability in forest management, particularly in fire-adapted ecosystems where timely and accurate canopy metrics are necessary for mitigation efforts.

2. Materials and Methods

2.1. Study Sites

This study focused on oak woodland/mixed-conifer forest sites on two nearby properties in the Mayacamas Mountains of Sonoma County, California, USA: (1) Pepperwood Preserve (38.57 N, 122.68 W), a 261-ha nature preserve (61–475 m.a.s.l.) managed by Pepperwood Foundation, and (2) Saddle Mountain Open Space Preserve (38.5 N, 122.63 W), a 388-ha nature preserve (233–549 m.a.s.l.) managed by the Sonoma County Agricultural Preservation and Open Space District (Figure 1). At both study sites, the dominant overstory trees are Douglas fir (Pseudotsuga mensizii), coast live oak (Quercus agrifolia), Oregon white oak (Quercus garryana), black oak (Quercus kelloggii), and California laurel (Umbellularia californica).
Both sites have a typical Mediterranean climate: cool, wet winters and hot, dry summers, and they receive most of their annual precipitation between October and April. The North Coast Range has strong climatic gradients of temperature and precipitation, controlled primarily by distance to the coast and elevation and secondarily by complex topography [3]. In addition, this region experiences Diablo winds, which are offshore, dry, northeasterly, downslope winds that frequently occur in October when live fuel moisture is at its driest [26]. The Diablo winds have proven to increase the severity of fire danger in the region. Notably, in October 2017, the Tubbs fire, driven by strong Diablo winds, moved through Pepperwood Preserve, burning 95% of the preserve.
This study focused on 44 woodland/forest plots selected to capture variation in forest structure, dominant vegetation composition, and burn severity. These include:
  • Twenty-two 20 m × 20 m long-term plots at Pepperwood Preserve, stratified based on topographic gradients, ecological heterogeneity, and post-fire burn severity;
  • Twenty-two 11.3 m radius plots at Saddle Mountain, selected to represent varying forest structures and dominant species across the preserve’s topographic gradient.
While no forest management treatments (e.g., thinning, prescribed burns) occurred within our study plots, both preserves have active management strategies aimed at reducing wildfire risk due to their proximity to residential structures classified as Moderate to Very High Fire Hazard Severity Zones (FHSZ) by CAL FIRE’s Fire and Resource Assessment Program [27]. Pepperwood Preserve management efforts include prescribed burns, creating fuel breaks, and removing/thinning fuels [28]. Similarly, Saddle Mountain Open Space Preserve included prescribed burns, fuel breaks, and removing/thinning fuels in a 2019 Adaptive Management Plan [29].

2.2. Ground-Based Field Data Collection

At each site, ground-based field measurements at the plot and individual tree level were taken concurrently with UAS flights (September/October 2019 at Pepperwood and August 2020 at Saddle Mountain) to validate UAS-derived forest structure estimates. Plot-level canopy cover was measured using a spherical densiometer, with readings recorded at the plot center and each cardinal direction before averaging the percent canopy cover in each plot [30,31].
For individual trees, total height and lowest live crown height were measured for each tree ≥ 10 cm diameter at breast height (DBH) in each plot. Lowest live crown was defined as the height above the ground of the lowest branch supporting live vegetation within 2 m of the next height branch and was measured using either a measuring tape (for lower base canopy) or a Laser Technology Impulse 200 Laser Rangefinder (for higher base canopy and tree heights) [12]. Measurements with the rangefinder followed manufacturer guidelines, using a two-point sine method to estimate tree height. Field technicians were positioned at a distance of 10–15 m from the tree base for an optimal viewing angle, ensuring minimal obstructions from dense vegetation (Laser Technology, Inc., Centennial, CO, USA). To mitigate challenges associated with dense stands and occlusions, field technicians assisted by holding bright panels to delineate tree bases and trunk locations, allowing the rangefinder operator to ensure clear line-of-sight measurements. These procedures helped reduce the likelihood of misestimation in multi-layered stands but do not eliminate all sources of error. As with Forbes et al. [25], we defined CBH as the average distance between the bottom of the canopy and ground. Then, to determine canopy base height (CBH), the lowest live crown height for each tree within each plot was averaged.
In addition, a multi-scan approach using a terrestrial laser scanner (TLS; Riegl VZ-400i; RIEGL Laser Measurement Systems GmbH, Horn, Austria) was employed to scan each plot concurrently with UAS and field measurements. Each plot had nine TLS scan positions spaced 10 m apart in a systematic grid design to ensure uniform point density and reduce occlusions throughout the plot and canopy. These scans were merged and georeferenced using RiSCAN Pro software (Riegl Laser Measurement Systems GmbH, version 2.8.2) for each plot before the final point cloud was aligned to ALS data (Section 2.3). TLS-derived tree locations (X, Y coordinates) were used as reference data for UAS-derived tree detections (Section 2.6). For more details on TLS methods, see Forbes et al. [25].

2.3. Unoccupied Aerial System (UAS) Structure from Motion (SfM) Multispectral Data Collection and Processing

UAS data were collected using two different platforms: (1) a SenseFly eBee X fixed-wing platform and (2) DJI Matrice 200 quadcopter platform. Platform selection depended on terrain and site conditions, with the fixed-wing UAS (SenseFly eBee X) used for large, open areas due to its extended flight endurance, while the quadcopter UAS (DJI Matrice 200) was deployed in densely forested regions where limited open space prevented fixed-wing takeoff and landing [12,32].
The UAS platforms were equipped with a MicaSense RedEdge-MX sensor and a MicaSense Downwelling Light Sensor (DLS) 2. The MicaSense RedEdge-MX sensor collects five spectral bands: blue (475 nm center, 20 nm bandwidth), green (560 nm center, 20 nm bandwidth), red (668 nm center, 10 nm bandwidth), red edge (717 nm center, 10 nm bandwidth), and near-IR (840 nm center, 40 nm bandwidth). The MicaSense DLS 2 measures ambient sunlight and sun angle, while radiance data were calibrated using reflectance of a MicaSense calibration panel (RP04-1926247-OB).
Custom flights were planned using SenseFly’s eMotion 3 software (version 3.7) for the eBee X and MicaSense’s Atlas Flight software (version 1.9) for the Matrice 200. Flight surveys were conducted in zones with flight lines producing a 75% forward and side-to-side image overlap, at a maximum flight altitude of 122 m above ALS-derived digital surface model [32], producing images with a pixel resolution of 8 cm. UAS surveys were con-ducted between 27 September 2019 and 8 October 2019 (Pepperwood Preserve; 10 flight zones), with the eBee X, and 4 August 2020 and 5 August 2020 (Saddle Mountain Open Space Preserve; 5 flight zones), mostly with the eBee X but with the M200 for one small zone to the northwest. All flights were under leaf-on and clear-sky conditions. Surveys were conducted within a 45 degrees solar incidence angle to minimize shadowing. For more UAS flight details, see Reilly et al. [32].
Images were SfM-processed using Pix4Dmapper software (Pix4D, Lausanne, Switzerland, version 4.4.12). Pix4Dmapper processing included: multispectral reflectance calibration, generation of multispectral reflectance orthomosaics with an 8 cm pixel size, and generation of 3D points clouds. The UAS-SfM point clouds for the flight areas had an average point density of 180 points/m2. Following data preparation, the orthomosaic images and 3D point clouds underwent additional pre-processing steps. UAS pre-processing consisted of (1) ground finding/digital terrain models (DTM; i.e., bare earth) generation, (2) iterative closest point (ICP) registration to ALS point clouds based on ground points, (3) height normalization using ALS as the ground reference, (4) calculation of pixel-level Normalized Difference Vegetation Index (NDVI), and (5) clipping data to flight zones and study site plots [32]. Further details of this processing pipeline are described in Reilly et al. [27]. UAS-SfM DTM rasters were used to height normalize digital surface model (DSM) rasters, producing a normalized digital surface model (nDSM) raster for canopy cover estimations. Point clouds used to identify individual trees in this study were height normalized using ALS DTMs [32].

2.4. UAS-Derived Canopy Cover Estimates

Area-based canopy cover was estimated from UAS-SfM data at the zone- and plot-level using eCognition software (Trimble Geospatial, Sunnyvale, California, USA, version 10.0) employing an adapted eCognition workflow [33]. The workflow relies on Pix4Dmapper orthomosaic images and nDSM (Section 2.3). Additionally, NDVI was calculated from the orthomosaic image with the equation:
NDVI = (NIR − Red)/(NIR + Red)
Forest cover pixels were classified based on nDSM and NDVI thresholds (Figure 2). The nDSM range of 0.05–10 m was used to exclude ground cover and very low vegetation (below 0.05 m) while ensuring inclusion of all tree canopies, with 10 m providing a buffer for potential outliers or unusually tall vegetation in the study area. The NDVI range of 0.2 to 0.45 was selected based on iterative testing, excluding bare ground and other non-canopy features while focusing on moderate photosynthetic levels common in oak woodlands. The resulting canopy raster was exported from eCognition to ArcGIS Pro (Esri 2021, version 2.8.1, Redlands, CA, USA) for statistical analyses. For UAS zone-level canopy cover estimates, the exported canopy raster was resampled from 8 cm to 20 cm pixels and summarized in 20 m cells to coincide with the scale of plot measurements. Total canopy cover was estimated within 20 m grid cells (Figure 3).
Two datasets were used to validate UAS-derived canopy cover estimates: ALS data and ground-measured canopy cover (see Section 2.2). ALS point clouds and associated multispectral imagery (red and NIR bands) from the 2013 Sonoma County Vegetation Mapping and LiDAR Program (www.sonomavegmap.org, accessed on 30 September 2019) were processed using a workflow similar to UAS processing in eCognition to generate canopy cover estimates. Both ALS- and UAS-derived canopy cover estimates were summarized within the same 20 m grid cells to ensure consistency in spatial analysis. For ground-based validation, UAS-derived canopy cover estimates were extracted using an overlay with the corresponding plot polygons to align directly with field-measured canopy cover. Linear regression analyses were performed to compare UAS-derived canopy cover estimates to ALS data at the 20 m cell level and to ground-measured canopy cover values at the individual plot level (N = 44 plots) in R Statistical Software (v4.0.4; R Core Team 2021).
The UAS imagery was collected 6–7 years (2019 and 2020) after the ALS data was collected (2013), and this temporal difference could impact our analysis of canopy cover due to changes in vegetation, including growth, consumption, scorch, and epicormic sprouting after the 2017 Tubbs Fire in Pepperwood Preserve (Figure 1). To assess the impact of fire on our analysis, each 20 m × 20 m grid cell in Pepperwood Preserve was categorized based on burn severity derived from the Monitoring Trends in Burn Severity (MTBS) mosaic [34,35]. MTBS data, originally at a resolution of 30 m × 30 m, was resampled to 20 m × 20 m using nearest-neighbor interpolation to align with the spatial resolution of UAS and ALS datasets. While resampling introduces potential uncertainty, this approach was necessary to ensure spatial consistency across datasets for analysis. Nearest-neighbor interpolation was chosen to preserve the categorical nature of burn severity data, as this method does not alter the original class values. Future studies may explore finer-resolution burn severity datasets or more advanced interpolation techniques to further improve alignment with UAS-derived data. Burn severity categories included ‘None’, ‘Low’, ‘Medium’, and ‘High’, based on the differenced Normalized Burn Ratio (dNBR) [34].
This categorization was used to analyze potential differences in UAS- and ALS-derived canopy cover estimates across burn severity levels. While this study does not account for spatial autocorrelation, future analyses could adopt statistical methods such as Kruskal–Wallis tests or other non-parametric approaches, as demonstrated by Reilly et al. (2021), to assess group differences robustly [32]. Between the UAS and ALS data collection periods, the 2017 Tubbs Fire caused 1,320 cells to sustain ‘Low’ levels of burn severity, 303 cells to sustain ‘Medium’ levels of burn severity, and 243 cells to sustain ‘High’ levels of burn severity, while 4,564 cells in Pepperwood Preserve were unaffected by the fire.

2.5. Individual Tree Detection (ITD)

Individual tree detection (ITD) includes automated techniques to identify individual tree positions and delineate their crowns in 3D point clouds and/or imagery (i.e., segmentation). By segmenting crowns, analyses can focus on tree-level metrics (Section 2.7). Roussel et al. [36] highlighted multiple ITD methods designed for ALS data, noting that algorithms can use raw 3D point clouds or raster canopy height models (e.g., nDSM). However, ITD can be focused on high resolution imagery, such as using deep learning models applied to UAS multispectral imagery [37,38], and hybrid 3D and image-based approaches are only recently being merged [39,40].
The open-source, point cloud-oriented R package, lidR, includes unsupervised implementation options for both nDSM- and point cloud-based individual tree detection (ITD). These methods require no training or hand annotations and offer flexibility for user-defined metrics. This study used the Li et al. [13] algorithm implemented in the lidR package (version 3.2.3) in R (version 4.0.4) to perform ITD and segmentation within UAS-SfM point clouds. This algorithm, specifically designed for ITD with point cloud data, uses relative spacing between trees with object-oriented classification rules that can be tailored to forest structure and tree characteristics.
Key parameters were optimized to improve tree detection and segmentation accuracy. These included the distance threshold (DT), search radius (R), minimum height (hmin), and NDVI threshold. Initial testing included a broad range of parameter combinations (e.g., DT: 1–10 m, R: 1–5 m, hmin: 1–5 m, NDVI: 0.01–0.5) to evaluate the algorithm’s performance under diverse conditions. However, many combinations yielded inaccurate tree detections due to the structural complexity of the study area, leading to over- or under-segmentation. To address this, parameter ranges were refined based on observed forest structure and tree density in the field, as well as the algorithm’s sensitivity to these parameters. The final ranges reflect values that provided the most reliable results across the diverse forest conditions in our study area. Final iterations tested DT values of 1, 2, and 5 m; R values of 1–3 m; hmin values of 1, 2, and 3 m; and NDVI thresholds of 0.02–0.4. The default Z parameter (15) and maximum crown radius (10) were used, as these did not impact results. Preliminary testing also revealed that using two DT values did not improve performance, so only one DT value was tested per iteration. This approach returned a dataset of tree-top locations and segmented tree crowns.

2.6. Accuracy Assessment of Tree Detections

The accuracy of automated UAS-derived tree detection with associated parameter selection was assessed by comparing UAS-derived tree-top locations (Section 2.5) to TLS-derived stem locations (Section 2.1). We found that variations in tree spacing affected detection accuracy, prompting us to separate analyses by plots grouped into five tree density classes, determined by TLS-derived stem counts within each plot. Lower-density plots often exhibited greater gaps or broader crown widths, while higher-density plots had overlapping or tightly spaced crowns, which affected segmentation performance. The tree density classes were defined as follows:
  • Density class one: Plots with 3, 5, 7, and 8 trees (7 plots)
  • Density class two: Plots with 9, 12, 13, and 14 trees (7 plots)
  • Density class three: Plots with 16, 17, 18, 19, and 20 trees (12 plots)
  • Density class four: Plots with 21, 23, 24, 26, 29, 30, and 31 trees (10 plots)
  • Density class five: Plots with 32, 34, 37, 38, 39, 44, and 58 trees (8 plots)
TLS-derived tree stem locations were paired with diameter at breast height (DBH) measurements and identified using Lidar360 [25,41]. Crown widths for each tree with DBH > 10 cm were calculated using a region-specific equation (Equation (2)) maintained by the Forest Vegetation Simulation (FVS) program. FVS uses the equation:
CW = a1 + a2 × DBH
where CW = crown width, a1 and a2 are species-specific coefficients, and DBH = diameter at breast height.
To assess accuracy, UAS-derived tree-top points were spatially joined with TLS-derived crown width polygons. Points falling within the estimated crowns were classified as true positives, while TLS crowns without corresponding UAS points were classified as false negatives (omissions). UAS-derived points located outside any TLS crown were classified as false positives (commissions). Additionally, when multiple UAS-derived tree-top points fell within the same TLS crown, the closest point to the TLS stem location was designated as the true positive, and the additional points within the crown were classified as false positives, reflecting over-segmentation.
UAS-derived tree location data from each iteration of the two methods were compared to TLS-derived tree locations to assess the accuracy of each method, with subsequent evaluation of true positives, false negatives, and false positives. Validation metrics used to assess each tree identification method and iteration included recall (r), precision (p), and F1-score (F1), which were calculated using the following equations:
r = True Positives/(True Positives + False Negatives)
p = True Positives/(True Positives + False Positives)
F1 = 2 × (r × p)/(r + p))

2.7. Derivation of Individual Tree Metrics

Following individual tree detection, UAS-SfM individual tree metrics were derived from the points falling within those segmented trees (Section 2.5) that were accurately detected (Section 2.6). UAS-SfM total tree heights were derived from the maximum Z value for points in each segmented tree. UAS-SfM canopy base height (CBH) was estimated by determining the fifth percentile of height for points within each segmented tree. This percentile was used because it was the closest to a 1:1 relationship, while all other percentiles had very little significance. The UAS-SfM tree heights and CBH for each detected tree were compared to field-based individual tree measurements using linear regression models in R Statistical software (v4.0.4; R Core Team 2021).

3. Results

3.1. Canopy Cover

UAS-SfM percent canopy cover estimates were significantly correlated with ground-based densiometer estimates (r2 = 0.89; RMSE = 5.56%), yet were overestimated for our range of data, which included percent cover of about 40% and greater (Figure 4). The UAS-SfM canopy cover estimates were also strongly correlated with the ALS-derived canopy cover estimates and in several areas, the ALS-derived estimated canopy cover was greater than the UAS-derived estimate (adjusted r2 = 0.79, p < 0.0001; Figure 5).
In Pepperwood Preserve, the absolute error among the UAS- and ALS-derived canopy cover estimates was more significant in cells affected by the Tubbs fire; 55.1% of ‘Low’ severity cells, 69% of ‘Medium’ severity cells, and 50.6% of ‘High’ severity cells with absolute errors over 10% (Figure 6).
These differences may stem from fire-induced mortality, epicormic sprouting, and subsequent canopy regrowth occurring in the interval between ALS (2013) and UAS (2019/2020) acquisitions. The impact of burn severity on canopy cover estimation suggests that fire history must be considered when comparing datasets collected over different time frames.

3.2. Individual Tree Detection

The accuracy of UAS-SfM tree detection varied across tree density classes, with higher tree densities leading to increased omission errors due to overlapping crowns. Density class one, the iteration with the smallest DT, R, and hmin parameters, had the highest detection of 93%. It also produced the highest commission error, with 273 additional trees detected. The highest DT value had a lower detection rate (30 of 41 trees) and the lowest commission error (16 trees), yielding the highest F1 score of 0.73 (Table 1A). Density class two performed best with a higher minimum height value and NDVI threshold of 0.2. The optimized iteration for density class two detected 70% of the trees and had a commission error of 36 additional trees, resulting in an F1 score of 0.65 (Table 1B). The optimal parameters for density class three consisted of a DT value of 1, R of 3, hmin of 2, and NDVI threshold of 0.04. This resulted in 76% of trees being detected along with 94 additional trees, generating an F1 score of 0.69 (Table 1C). Similar to density class two, density class four performed better with higher values for R, hmin, and NDVI threshold, but with a DT value of 1. The optimized iteration for density class four produced a F1 score of 0.7 (Table 1D). Lastly, density class five was optimized using the lowest values for all parameters. With the low value parameters, 72% of trees in density class five were identified, 147 additional trees were detected, and an F1 score of 0.65 was produced (Table 1E). In general, the higher DT value was optimal for plots with the lowest density of trees (density class one; 3–8 trees/plot), while the lowest values for all parameters was the optimal iteration for plots with a higher density of trees (density class 5; 32–58 trees/plot).

3.3. Individual Tree Metrics

For those UAS-SfM trees that were correctly identified using optimized iterations (N = 561; Section 3.2), tree heights had a significant linear relationship with field-measured heights (adjusted r2 = 0.70, RMSE = 5.01 m, p < 0.0001; Figure 7A).
Tree height varied by species type:
  • Conifer trees had a more significant linear relationship with field measurements (r2 = 0.69, p < 0.0001) and had a higher RMSE (6.17 m; Figure 7C)
  • Broadleaf trees had a weaker correlation (r2 = 0.37, RMSE = 3.88 m, p < 0.0001; Figure 7B).
UAS-SfM canopy base height (CBH) estimates did not exhibit a significant correlation with field-based measurements (r2 = 0.32, RMSE = 3.98 m, p < 0.0001).

4. Discussion

4.1. Feasibility of UAS-SfM for Forest Structure Estimation

This study aimed to determine the feasibility of using UAS with a multispectral sensor to estimate forest canopy cover, individual tree heights, and CBH in Northern California oak woodlands and mixed-conifer forests. Our approach was of interest to land managers since the UAS provided high-resolution multispectral imagery (8 cm pixel) that has not been previously available for the study area. Furthermore, UAS flights can be conducted at a lower cost and higher temporal frequency than piloted ALS surveys, which are expensive and often require years for data acquisition and public release.

4.2. Canopy Cover Estimations

Perhaps due to the fine resolution of the UAS data, canopy cover estimated via UAS-SfM 3D and image orthomosiac NDVI in eCognition was greater than canopy cover estimated via both a ground-level densiometer and airborne laser scanner NDVI data. The difference between field-based estimates and UAS-derived canopy cover could be due to an inherent error in densiometer estimates, differences in observers, and a limited view of the entire canopy, thus yielding a coarse resolution of data [42,43]. In a similar study, McIntosh et al. [44] found that coarse resolution (1:40,000) aerial photos had a weak correlation with field-measured canopy cover. Future research might consider more high-resolution validation datasets to estimate canopy cover accuracy, such as hemispherical photography and LiCOR gap fractions [45,46].

4.3. Individual Tree Detection and Segmentation

Even though the individual tree detection (ITD) and segmentation method used for this study was originally intended for airborne lidar point cloud data [13], the UAS-SfM individual tree detection successfully identified the presence of trees above 10 cm DBH (Figure S1). However, species diversity and forest structure complexity impacted segmentation accuracy.
Due to the structural complexity and species diversity of the study area, the parameterization and resultant model accuracy of tree-point detection was separated into five tree density classes. Within the density classes, the varying levels of tree-top detection, omission, and commission were dependent on algorithm parameters, which must be adjusted to fit the structure of a specific study area and the user’s needs. Ideally, a user would have some field survey data to optimize parameters for a range of forest types in a study area, as we did.
Our model with the highest F1 score (0.73) was similar to a SfM-derived ITD study in a deciduous broadleaf study area in Iran (F1 score = 0.75) using a region growing algorithm [6]. In the future, more flight missions should be considered as seasonal (i.e., leaf-on versus leaf-off) and temporal differences likely affect optimal parameter values (e.g., NDVI thresholds) and the resultant accuracy. Additionally, the variation in results across tree density classes suggests that the method may require further refinement to improve detection accuracy in higher-density areas.

4.4. Tree Height and Canopy Base Height Estimations

The UAS-SfM maximum height associated with segmented crowns accurately estimated field-based tree heights (r2 = 0.69, RMSE = 5.1 m). Other studies [7,17] have shown strong results between UAS-derived tree heights and field-based measurements (r2 = 0.71, RMSE = 1.83 m and r2 = 0.74, RMSE = 2.03, respectively). We were, however, unable to detect and estimate tree metrics for every tree in our study area.
Compared to other studies conducted in areas with primarily straight, single-stem conifer trees (e.g., Ponderosa pine, Norway spruce, European larch, etc.), our study sites were heterogeneous stands with a mix of oak and conifer species (Figure S1). With their broad, spread-out crowns, oaks are more difficult to establish local maxima because they do not have one characteristic peak [47]. This tends to lead to the over-detection of trees. Additionally, due to the high canopy cover and the lack of penetration with UAS-SfM, understory trees are often difficult to identify.
Indeed, with the high density of canopy cover, it was difficult to identify the bottoms and edges of tree canopies to calculate crown base height. While this is a limitation of using UAS-SfM data in high canopy cover areas, ALS data can be used to identify plot-level canopy base heights [14,15]. Our results were consistent with findings from a LANDFIRE database study [48], which reported difficulties in accurately characterizing vegetation structure in heterogeneous forest stands using remote sensing techniques (r2 < 0.45 for 11 of the 12 study sites). This similarity highlights the challenges posed by diverse forest compositions and the limitations of remote sensing methods, including UAS-SfM, in capturing complex canopy structures.

4.5. Future Recommendations

In the future, alternative methods to the point-based tree detection and segmentation algorithm we used here, such as raster-based canopy height models [9], 3D deep learning [34,35] and additional use of multispectral information [37,38], might be useful when applied to UAS SfM products. Alternatively, UAS LiDAR is a viable yet more costly option for this application. In a recent study, Ramalho de Oliveira et al. [46] were able to detect 98% of 2199 trees in a pine stand employing an automated tree detection algorithm using UAS LiDAR. In a more structurally complex mixed species eucalypt forest, Jaskierniak et al. [16] used UAS LiDAR data and a novel bottom-up approach using kernel densities to stratify the vegetation profile to detect 85% of individual trees and delineate tree crowns. While this developed bottom-up approach proved to be successful in a complex forest, it still might not be as successful with UAS-SfM data due to the limitation of canopy penetration [12,32].

5. Conclusions

Overall, results from this study show that UAS imagery and SfM point clouds can be used to estimate canopy cover, canopy height, and detect larger individual trees in areas of low canopy cover. UAS-SfM-derived canopy cover estimates had a strong linear relationship with ground-based measurements (r2 = 0.89, RMSE = 5.56%, p < 0.0001) and ALS-derived canopy cover (adjusted r2 = 0.79, p < 0.0001), confirming its reliability for estimating canopy structure at multiple spatial scales. Individual tree detection using UAS-SfM successfully identified trees above 10 cm DBH, with the highest-performing model achieving an F1-score of 0.73, comparable to other studies applying SfM-based ITD methods. UAS-SfM tree height estimates also significantly correlated with field-measured heights (r2 = 0.69, RMSE = 5.1 m), demonstrating its potential for capturing forest structure. However, challenges remain in detecting understory trees and defining canopy base height, where UAS-SfM-derived CBH showed lower correlation with field measurements (r2 = 0.32, RMSE = 3.98 m, p < 0.0001). Temporal differences between UAS and ALS data, caused by vegetation growth and fire effects, further underscore the need for careful consideration of data collection timing. Despite these limitations, compared to ALS missions ($500–$1200/km2), UAS-SfM remains a cost-effective and flexible tool for on-demand and/or repeated monitoring ($8500–$12,500 for one-time purchase) that can provide valuable insights for land management and fire risk assessment over smaller spatial extents.

Supplementary Materials

The following supporting information can be downloaded at: https://www.mdpi.com/article/10.3390/f16040564/s1, Figure S1: Diameter at breast height (DBH) distribution (above) and mean (below) of individual trees at both study sites (Pepperwood Preserve and Saddle Mountain Preserve).

Author Contributions

Conceptualization, A.K., J.D.D. and L.B.; methodology, A.K.; software, A.K.; validation, A.K.; formal analysis, A.K.; investigation, A.K., L.B. and L.P.B.; resources, L.B. and L.P.B.; data curation, A.K.; writing—original draft preparation, A.K.; writing—review and editing, L.B., L.P.B. and J.D.D.; visualization, A.K.; supervision, L.B.; project administration, L.P.B.; funding acquisition, L.B. and L.P.B. All authors have read and agreed to the published version of the manuscript.

Funding

This research was funded by a CAL FIRE Forest Health and Forest Legacy Program grant (8GG18806) and an NSF CAREER Grant [2145728] to LPB and a California State University Agricultural Research Institute (20-01-106) grant to LPB and LB.

Data Availability Statement

The raw data supporting the conclusions of this article will be made available by the authors on request.

Acknowledgments

This project would not have been possible without the hard work and dedication of Matthew Clark, Corbin Matley, and Elise Piazza for UAS data collection and processing. Sean Reilly, Brieanne Forbes, Paris Krause, and Catherine Seel also assisted with field data collection and processing. We are grateful to Matthew Clark for his comments on the manuscript and to Sean Reilly and Elliott Smeds for their assistance with figures. We additionally thank the Sonoma County Agricultural Preservation and Open Space District, Pepperwood Preserve, and David Ackerly for granting us access to their sites and facilitating our research data collection.

Conflicts of Interest

The authors declare no conflicts of interest. The funders had no role in the design of the study, in the collection, analyses, or interpretation of data, in the writing of the manuscript, or in the decision to publish the results.

References

  1. Gaman, T.; Firman, J. Oaks 2040: The Status and Future of Oaks in California; California Oak Foundation: Oakland, CA, USA, 2006; p. 14. [Google Scholar]
  2. Holmes, K.A.; Veblen, K.E.; Young, T.P.; Berry, A.M. California Oaks and Fire: A Review and Case Study. In Proceedings of the Sixth California Oak Symposium: Today’s Challenges, Tomorrow’s Opportunities, Rohnert Park, CA, USA, 9–12 October 2006. [Google Scholar]
  3. Stuart, J.D.; Stephens, S.L.; Agee, J.K. North Coast Bioregion. In Fire in California’s Ecosystems; Sugihara, N.G., Van Wagtendonk, J.W., Shaffer, K.E., Fites-Kaufman, J., Thode, A.E., Eds.; University of California Press: Berkeley, CA, USA, 2006; pp. 147–169. ISBN 978-0-520-24605-8. [Google Scholar]
  4. Bernhardt, E.A.; Swiecki, T.J. Restoring Oak Woodlands in California: Theory and Practice. Available online: http://phytosphere.com/restoringoakwoodlands/oakrestoration.htm (accessed on 22 May 2022).
  5. Forest Management Task Force. California’s Wildfire and Forest Resilience Action Plan; State of California: Sacramento, CA, USA, 2021. [Google Scholar]
  6. Miraki, M.; Sohrabi, H.; Fatehi, P.; Kneubuehler, M. Individual Tree Crown Delineation from High-Resolution UAV Images in Broadleaf Forest. Ecol. Inform. 2021, 61, 101207. [Google Scholar] [CrossRef]
  7. Shin, P.; Sankey, T.; Moore, M.; Thode, A. Evaluating Unmanned Aerial Vehicle Images for Estimating Forest Canopy Fuels in a Ponderosa Pine Stand. Remote Sens. 2018, 10, 1266. [Google Scholar] [CrossRef]
  8. Ke, Y.; Quackenbush, L.J. A Review of Methods for Automatic Individual Tree-Crown Detection and Delineation from Passive Remote Sensing. Int. J. Remote Sens. 2011, 32, 4725–4747. [Google Scholar] [CrossRef]
  9. Aubry-Kientz, M.; Dutrieux, R.; Ferraz, A.; Saatchi, S.; Hamraz, H.; Williams, J.; Coomes, D.; Piboule, A.; Vincent, G. A Comparative Assessment of the Performance of Individual Tree Crowns Delineation Algorithms from ALS Data in Tropical Forests. Remote Sens. 2019, 11, 1086. [Google Scholar] [CrossRef]
  10. Weinstein, B.G.; Marconi, S.; Bohlman, S.A.; Zare, A.; White, E.P. Cross-Site Learning in Deep Learning RGB Tree Crown Detection. Ecol. Inform. 2020, 56, 101061. [Google Scholar] [CrossRef]
  11. Wallace, L.; Lucieer, A.; Malenovský, Z.; Turner, D.; Vopěnka, P. Assessment of Forest Structure Using Two UAV Techniques: A Comparison of Airborne Laser Scanning and Structure from Motion (SfM) Point Clouds. Forests 2016, 7, 62. [Google Scholar] [CrossRef]
  12. Reilly, S.; Clark, M.L.; Loechler, L.; Spillane, J.; Kozanitas, M.; Krause, P.; Ackerly, D.; Bentley, L.P.; Menor, I.O. Unoccupied Aerial System (UAS) Structure-from-Motion Canopy Fuel Parameters: Multisite Area-Based Modelling across Forests in California, USA. Remote Sens. Environ. 2024, 312, 114310. [Google Scholar] [CrossRef]
  13. Li, W.; Guo, Q.; Jakubowski, M.K.; Kelly, M. A New Method for Segmenting Individual Trees from the Lidar Point Cloud. Photogramm. Eng. Remote Sens. 2012, 78, 75–84. [Google Scholar] [CrossRef]
  14. Andersen, H.-E.; McGaughey, R.J.; Reutebuch, S.E. Estimating Forest Canopy Fuel Parameters Using LIDAR Data. Remote Sens. Environ. 2005, 94, 441–449. [Google Scholar] [CrossRef]
  15. Erdody, T.L.; Moskal, L.M. Fusion of LiDAR and Imagery for Estimating Forest Canopy Fuels. Remote Sens. Environ. 2010, 114, 725–737. [Google Scholar] [CrossRef]
  16. Jaskierniak, D.; Lucieer, A.; Kuczera, G.; Turner, D.; Lane, P.N.J.; Benyon, R.G.; Haydon, S. Individual Tree Detection and Crown Delineation from Unmanned Aircraft System (UAS) LiDAR in Structurally Complex Mixed Species Eucalypt Forests. ISPRS J. Photogramm. Remote Sens. 2021, 171, 171–187. [Google Scholar] [CrossRef]
  17. Panagiotidis, D.; Abdollahnejad, A.; Surový, P.; Chiteculo, V. Determining Tree Height and Crown Diameter from High-Resolution UAV Imagery. Int. J. Remote Sens. 2017, 38, 2392–2410. [Google Scholar] [CrossRef]
  18. Shin, P. Unmanned Aerial Vehicles for Estimating Forest Canopy Fuels in Ponderosa Pine Forest. Master’s Thesis, Northern Arizona University, Flagstaff, AZ, USA, 2018. [Google Scholar]
  19. Westoby, M.J.; Brasington, J.; Glasser, N.F.; Hambrey, M.J.; Reynolds, J.M. ‘Structure-from-Motion’ Photogrammetry: A Low-Cost, Effective Tool for Geoscience Applications. Geomorphology 2012, 179, 300–314. [Google Scholar] [CrossRef]
  20. Dandois, J.P.; Ellis, E.C. High Spatial Resolution Three-Dimensional Mapping of Vegetation Spectral Dynamics Using Computer Vision. Remote Sens. Environ. 2013, 136, 259–276. [Google Scholar] [CrossRef]
  21. Morgenroth, J.; Gomez, C. Assessment of Tree Structure Using a 3D Image Analysis Technique—A Proof of Concept. Urban For. Urban Green. 2014, 13, 198–203. [Google Scholar] [CrossRef]
  22. Frankhauser, P.; Tannier, C.; Vuidel, G.; Houot, H. An Integrated Multifractal Modelling to Urban and Regional Planning. Comput. Environ. Urban Syst. 2018, 67, 132–146. [Google Scholar] [CrossRef]
  23. Liu, L.; Lim, S.; Shen, X.; Yebra, M. A Hybrid Method for Segmenting Individual Trees from Airborne Lidar Data. Comput. Electron. Agric. 2019, 163, 104871. [Google Scholar] [CrossRef]
  24. Lim, Y.S.; La, P.H.; Park, J.S.; Lee, M.H.; Pyeon, M.W.; Kim, J.-I. Calculation of Tree Height and Canopy Crown from Drone Images Using Segmentation. J. Korean Soc. Surv. Geod. Photogramm. Cartogr. 2015, 33, 605–614. [Google Scholar] [CrossRef]
  25. Forbes, B.; Reilly, S.; Clark, M.; Ferrell, R.; Kelly, A.; Krause, P.; Matley, C.; O’Neil, M.; Villasenor, M.; Disney, M.; et al. Comparing Remote Sensing and Field-Based Approaches to Estimate Ladder Fuels and Predict Wildfire Burn Severity. Front. For. Glob. Change 2022, 5, 818713. [Google Scholar] [CrossRef]
  26. Bowers, C.L. The Diablo Winds of Northern California: Climatology and Numerical Simulations. Master’s Thesis, San Jose State University, San Jose, CA, USA, 2018. [Google Scholar]
  27. CAL FIRE Fire Hazard Severity Zones (FHSZ). Available online: https://osfm.fire.ca.gov/divisions/community-wildfire-preparedness-and-mitigation/wildfire-preparedness/fire-hazard-severity-zones/ (accessed on 22 May 2022).
  28. Gillogly, M.; Dodge, C.; Halbur, M.; Micheli, L.; McKay, C.; Heller, N.; Benson, B. Adaptive Management Plan for Pepperwood Preserve; Dwight Center for Conservation Science: Pepperwood, Santa Rosa, CA, 2017. [Google Scholar]
  29. Ag + Open Space. Saddle Mountain Open Space Preserve Management Plan; Ag + Open Space: Santa Rosa, CA, USA, 2019. [Google Scholar]
  30. Lemmon, P.E. A Spherical Densiometer for Estimating Forest Overstory Density. For. Sci. 1956, 2, 314–320. [Google Scholar]
  31. Lemmon, P.E. A New Instrument for Measuring Forest Overstory Density. J. For. 1957, 55, 667–668. [Google Scholar]
  32. Reilly, S.; Clark, M.L.; Bentley, L.P.; Matley, C.; Piazza, E.; Oliveras Menor, I. The Potential of Multispectral Imagery and 3D Point Clouds from Unoccupied Aerial Systems (UAS) for Monitoring Forest Structure and the Impacts of Wildfire in Mediterranean-Climate Forests. Remote Sens. 2021, 13, 3810. [Google Scholar] [CrossRef]
  33. Trimble Support Tree Crown Delineation. Available online: https://support.ecognition.com/hc/en-us/articles/360016224679-Tree-Crown-Delineation (accessed on 11 March 2022).
  34. Eidenshink, J.; Schwind, B.; Brewer, K.; Zhu, Z.-L.; Quayle, B.; Howard, S. A Project for Monitoring Trends in Burn Severity. Fire Ecol. 2007, 3, 3–21. [Google Scholar] [CrossRef]
  35. Picotte, J.J.; Bhattarai, K.; Howard, D.; Lecker, J.; Epting, J.; Quayle, B.; Benson, N.; Nelson, K. Changes to the Monitoring Trends in Burn Severity Program Mapping Production Procedures and Data Products. Fire Ecol. 2020, 16, 16. [Google Scholar] [CrossRef]
  36. Roussel, J.-R.; Auty, D.; Coops, N.C.; Tompalski, P.; Goodbody, T.R.H.; Meador, A.S.; Bourdon, J.-F.; de Boissieu, F.; Achim, A. lidR: An R Package for Analysis of Airborne Laser Scanning (ALS) Data. Remote Sens. Environ. 2020, 251, 112061. [Google Scholar] [CrossRef]
  37. Zhang, C.; Zhou, J.; Wang, H.; Tan, T.; Cui, M.; Huang, Z.; Wang, P.; Zhang, L. Multi-Species Individual Tree Segmentation and Identification Based on Improved Mask R-CNN and UAV Imagery in Mixed Forests. Remote Sens. 2022, 14, 874. [Google Scholar] [CrossRef]
  38. Schiefer, F.; Kattenborn, T.; Frick, A.; Frey, J.; Schall, P.; Koch, B.; Schmidtlein, S. Mapping Forest Tree Species in High Resolution UAV-Based RGB-Imagery by Means of Convolutional Neural Networks. ISPRS J. Photogramm. Remote Sens. 2020, 170, 205–215. [Google Scholar] [CrossRef]
  39. Qin, H.; Zhou, W.; Yao, Y.; Wang, W. Individual Tree Segmentation and Tree Species Classification in Subtropical Broadleaf Forests Using UAV-Based LiDAR, Hyperspectral, and Ultrahigh-Resolution RGB Data. Remote Sens. Environ. 2022, 280, 113143. [Google Scholar] [CrossRef]
  40. Windrim, L.; Bryson, M. Detection, Segmentation, and Model Fitting of Individual Tree Stems from Airborne Laser Scanning of Forests Using Deep Learning. Remote Sens. 2020, 12, 1469. [Google Scholar] [CrossRef]
  41. Krause, P.; Forbes, B.; Barajas-Ritchie, A.; Clark, M.; Disney, M.; Wilkes, P.; Bentley, L.P. Using Terrestrial Laser Scanning to Evaluate Non-Destructive Aboveground Biomass Allometries in Diverse Northern California Forests. Front. Remote Sens. 2023, 4, 1132208. [Google Scholar]
  42. Jennings, S. Assessing Forest Canopies and Understorey Illumination: Canopy Closure, Canopy Cover and Other Measures. Forestry 1999, 72, 59–74. [Google Scholar] [CrossRef]
  43. Englund, S.R.; O’Brien, J.J.; Clark, D.B. Evaluation of Digital and Film Hemispherical Photography and Spherical Densiometry for Measuring Forest Light Environments. Can. J. For. Res. 2000, 30, 1999–2005. [Google Scholar] [CrossRef]
  44. McIntosh, A.C.S.; Gray, A.N.; Garman, S.L. Estimating Canopy Cover from Standard Forest Inventory Measurements in Western Oregon. For. Sci. 2012, 58, 154–167. [Google Scholar] [CrossRef]
  45. Paletto, A.; Tosi, V. Forest Canopy Cover and Canopy Closure: Comparison of Assessment Techniques. Eur. J. For. Res. 2009, 128, 265–272. [Google Scholar] [CrossRef]
  46. Ramalho de Oliveira, L.F.; Lassiter, H.A.; Wilkinson, B.; Whitley, T.; Ifju, P.; Logan, S.R.; Peter, G.F.; Vogel, J.G.; Martin, T.A. Moving to Automated Tree Inventory: Comparison of UAS-Derived Lidar and Photogrammetric Data with Manual Ground Estimates. Remote Sens. 2020, 13, 72. [Google Scholar] [CrossRef]
  47. Zaforemska, A.; Xiao, W.; Gaulton, R. Individual Tree Detection from UAV LiDAR Data in a Mixed Species Woodland. Int. Arch. Photogramm. Remote Sens. Spat. Inf. Sci. 2019, 42, 657–663. [Google Scholar] [CrossRef]
  48. Reeves, M.C.; Ryan, K.C.; Rollins, M.G.; Thompson, T.G. Spatial Fuel Data Products of the LANDFIRE Project. Int. J. Wildland Fire 2009, 18, 250. [Google Scholar] [CrossRef]
Figure 1. (a) Location of the study region within California, USA; (b) study areas within Sonoma County, California, showing the boundaries of Pepperwood Preserve and Saddle Mountain Open Space Preserve; (c) UAS flight zones (outlined in yellow) overlaid on RGB UAS imagery of Pepperwood Preserve with field plots marked in red; (d) UAS flight zones (outlined in yellow) overlaid on RGB UAS imagery of Saddle Mountain Open Space Preserve with field plots marked in red.
Figure 1. (a) Location of the study region within California, USA; (b) study areas within Sonoma County, California, showing the boundaries of Pepperwood Preserve and Saddle Mountain Open Space Preserve; (c) UAS flight zones (outlined in yellow) overlaid on RGB UAS imagery of Pepperwood Preserve with field plots marked in red; (d) UAS flight zones (outlined in yellow) overlaid on RGB UAS imagery of Saddle Mountain Open Space Preserve with field plots marked in red.
Forests 16 00564 g001
Figure 2. Workflow for canopy cover estimates employing Pix4Dmapper, eCognition, and ArcGIS Pro.
Figure 2. Workflow for canopy cover estimates employing Pix4Dmapper, eCognition, and ArcGIS Pro.
Forests 16 00564 g002
Figure 3. Canopy cover output raster (20 cm resolution) (Left); 20 m × 20 m grid overlaid on canopy cover raster (Right).
Figure 3. Canopy cover output raster (20 cm resolution) (Left); 20 m × 20 m grid overlaid on canopy cover raster (Right).
Forests 16 00564 g003
Figure 4. The linear relationship between UAS-SfM and densiometer estimated percent canopy cover across all plots at both sites.
Figure 4. The linear relationship between UAS-SfM and densiometer estimated percent canopy cover across all plots at both sites.
Forests 16 00564 g004
Figure 5. Comparison of UAS-SfM and ALS-derived canopy cover estimates (%) at 20 m resolution.
Figure 5. Comparison of UAS-SfM and ALS-derived canopy cover estimates (%) at 20 m resolution.
Forests 16 00564 g005
Figure 6. The absolute error between the UAS-SfM and ALS-derived percent canopy cover estimates in Pepperwood Preserve categorized by 2017 Tubbs Fire MTBS burn severity. The numbers above the bars indicate the percentage of cells within each MTBS burn severity category.
Figure 6. The absolute error between the UAS-SfM and ALS-derived percent canopy cover estimates in Pepperwood Preserve categorized by 2017 Tubbs Fire MTBS burn severity. The numbers above the bars indicate the percentage of cells within each MTBS burn severity category.
Forests 16 00564 g006
Figure 7. Relationship between UAS-SfM estimated and field-measured tree heights for 561 trees across both sites, (A) with results separated by broadleaf (B) and conifer (C) trees.
Figure 7. Relationship between UAS-SfM estimated and field-measured tree heights for 561 trees across both sites, (A) with results separated by broadleaf (B) and conifer (C) trees.
Forests 16 00564 g007
Table 1. UAS-SfM tree detection accuracy assessment using five density classes (A–E). Optimal parameters within each density class include DT, R, hmin, and NDVI threshold (Section 2.5). The optimal model per density class, as chosen by highest F1 score, is indicated in bold.
Table 1. UAS-SfM tree detection accuracy assessment using five density classes (A–E). Optimal parameters within each density class include DT, R, hmin, and NDVI threshold (Section 2.5). The optimal model per density class, as chosen by highest F1 score, is indicated in bold.
A.Density Class One (3–8 Trees/Plot)
Plots: 1, 2, 1303, 1304, 1309, 1349, 1852
Total Trees: 41
DT ValueSearch
Radius
Minimum HeightNDVI ThresholdDetected Trees (%)Commission Error (%)Omitted Trees (%)rpF-1
1110.049366670.930.120.22
1320.0478278220.780.220.34
5110.047327270.730.730.73
1330.278268220.780.230.35
2330.261156390.610.280.38
B.Density Class Two (9–14 Trees/Plot)
Plots: 12, 1306, 1307, 1310, 1319, 1340, 1851
Total Trees: 79
DT ValueSearch
Radius
Minimum HeightNDVI ThresholdDetected Trees (%)Commission Error (%)Omitted Trees (%)rpF-1
1110.049128690.910.240.38
1320.048296180.820.460.59
5110.0437261720.180.590.27
1330.28194190.810.460.59
2330.27046300.700.600.65
C.Density Class Three (16–20 Trees/Plot)
Plots: 11, 16, 18, 19, 1301, 1325, 1329, 1335, 1336, 1337, 1342, 1854
Total Trees: 215
DT ValueSearch
Radius
Minimum HeightNDVI ThresholdDetected Trees (%)Commission Error (%)Omitted Trees (%)rpF-1
1110.0489172110.890.340.49
1320.047644240.760.640.69
5110.04348660.340.810.48
1330.27341270.730.640.68
2330.25621440.560.720.63
D.Density Class Four (21–31 Trees/Plot)
Plots: 3, 5, 15, 17, 20, 24, 25, 26, 1315, 1853
Total Trees: 270
DT ValueSearch
Radius
Minimum HeightNDVI ThresholdDetected Trees (%)Commission Error (%)Omitted Trees (%)rpF-1
1110.048199190.810.450.58
1320.046117390.610.780.68
5110.04210790.211.000.35
1330.26012400.600.830.70
2330.2456550.450.890.60
E.Density Class Five (32–58 Trees/Plot)
Plots: 4, 6, 7, 8, 13, 14, 23, 1332
Total Trees: 299
DT ValueSearch
Radius
Minimum HeightNDVI ThresholdDetected Trees (%)Commission Error (%)Omitted Trees (%)rpF-1
1110.047249280.720.590.65
1320.043917610.390.700.50
5110.04131870.130.950.22
1330.24616540.460.740.57
2330.2307700.300.820.44
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Share and Cite

MDPI and ACS Style

Kelly, A.; Blesius, L.; Davis, J.D.; Bentley, L.P. Using Unoccupied Aerial Systems (UAS) and Structure-from-Motion (SfM) to Measure Forest Canopy Cover and Individual Tree Height Metrics in Northern California Forests. Forests 2025, 16, 564. https://doi.org/10.3390/f16040564

AMA Style

Kelly A, Blesius L, Davis JD, Bentley LP. Using Unoccupied Aerial Systems (UAS) and Structure-from-Motion (SfM) to Measure Forest Canopy Cover and Individual Tree Height Metrics in Northern California Forests. Forests. 2025; 16(4):564. https://doi.org/10.3390/f16040564

Chicago/Turabian Style

Kelly, Allison, Leonhard Blesius, Jerry D. Davis, and Lisa Patrick Bentley. 2025. "Using Unoccupied Aerial Systems (UAS) and Structure-from-Motion (SfM) to Measure Forest Canopy Cover and Individual Tree Height Metrics in Northern California Forests" Forests 16, no. 4: 564. https://doi.org/10.3390/f16040564

APA Style

Kelly, A., Blesius, L., Davis, J. D., & Bentley, L. P. (2025). Using Unoccupied Aerial Systems (UAS) and Structure-from-Motion (SfM) to Measure Forest Canopy Cover and Individual Tree Height Metrics in Northern California Forests. Forests, 16(4), 564. https://doi.org/10.3390/f16040564

Note that from the first issue of 2016, this journal uses article numbers instead of page numbers. See further details here.

Article Metrics

Back to TopTop