Introduction

Encroachment of woody plant species on rangelands is of critical concern for rangeland health and function (Ansley et al., 2001; Carey et al., 2012; Shackelford et al., 2013). Woody plant encroachment often occurs with livestock overgrazing, fire suppression, and/or mechanical disturbance (Archer and Stokes 2000; Archer 2010; Ku et al., 2012; Adam et al., 2017; Teleki et al., 2019). Woody plant encroachment can negatively affect ecosystem services by altering soil nutrient cycling, hydrologic cycling, and reduction in herbaceous growth (Schlesinger et al., 2000; Archer et al., 2001; Heaton et al., 2003; Huxman et al., 2005; Wilcox 2007; Mirik and Ansley 2012). The increase of woody plant encroachment has been shown to facilitate decreases in biodiversity because of other species being outcompeted for resources (Ratajczak et al., 2012; Teleki et al., 2019).

Honey mesquite (Prosopis glandulosa; hereafter “mesquite”) is one of the most invasive native woody species within the southwestern United States (Dahl 1982; Ansley et al., 2001, 2018; Ku et al., 2012; Adam et al., 2017). Its native range expands from northern Mexico and throughout the southwestern United States in 11 states including California, Arizona, New Mexico, Texas, and Louisiana (Ansley et al., 1997; NRCS 2020). The encroachment of mesquite has resulted in loss of native rangelands across the southwest with negative effects on ecosystem functions such as reductions in water yield (Ansley et al., 1997; Ball and Taylor 2003; Nie et al., 2012). Although mesquite has been a part of the Texas landscape since before European settlements, it has become problematic to many ranchers and landowners within its range, which has been attributed to its adaptability and rate of spread into areas where it was not prevalent beforehand (Hennessy et al., 1983; Ansley et al., 1997; Anderson and Roderick, 2003). Mohamed et al., (2011) showed that even under proper grazing management, mesquite encroachment causes adverse effects on perennial grass production, which decreases available forage for livestock production. Because of mesquite's ability to reestablish and spread throughout the landscape, it is essential for rangeland managers to be able to monitor mesquite presence at the landscape level (Adam et al., 2017; Ansley et al., 2018) over time and space. Remote sensing approaches can be a viable solution to monitor mesquite presence across landscapes.

Remote sensing has been used to map and quantify the distribution of mesquite (Musick 1984; Fang et al., 2005; Yilmaz et al., 2008, Mirik and Ansley 2012; Adam et al., 2017). Methods used for monitoring mesquite include Light Detection and Ranging (LiDAR) (Ku et al., 2012), aerial imagery (Ansley et al., 2001), and satellite imagery at multiple resolutions (2–60 m) (Musick 1984; Mohamed et al., 2011; Adam et al., 2017). Ku et al., (2012) were able to use terrestrial LiDAR to estimate mesquite biomass at the plot level (5 × 20 m). However, LiDAR still remains a costly operation (Iglhaut et al., 2019; Jiang et al., 2020). Ansley et al., (2001) were able to successfully assess increases in mesquite canopy cover from 1976 to 1995 using low-altitude aerial images with high-resolution aerial photography (0.1 m) from a traditional piloted aircraft in North Texas but were limited by temporal flight availability and cost. The use of Digital Orthophoto Quadrangle (DOQQ) imagery from the National Aerial Imagery Program (NAIP) is limited to a revisit time of 2 yr, and spatial and spectral resolution can make species differentiation difficult to achieve when multiple woody species are present. The use of LANDSAT satellite imagery for detecting mesquite species has proven to be effective because of the high temporal resolution (revisit times every 16 d), but its use can be restricted because of the coarse spatial resolution (30 m) and the amount of cloud cover present when data are collected by satellites (Musick 1984; Fang et al., 2005; Yilmaz et al., 2008). These methods have proven to be successful at monitoring mesquite; however, low spatial resolution, high costs for high-resolution imagery, and/or low revisit times may limit continuous monitoring of mesquite (Ansley et al., 2001; Yilmaz et al., 2008; Cunliffe et al., 2016; Gillan et al., 2019). Recent developments of remote sensing platforms such as PlanetScope imagery (Dos Reis et al., 2020) and Sentinel-2 (Drusch et al., 2010) with a higher spatial resolution (3 m and 10–60 m, respectively), spectral resolution (4 and 13 bands, respectively), and revisit times (1–5 d) than LANDSAT (spatial resolution: 30 m; spectral resolution 9 bands; revisit time: 16 d) offer new opportunities to improve the detection, classification, and monitoring of temporal trends of invasive species such as mesquite.

The Sentinel-2 satellite mission from the European Space Agency provides open-access imagery with 13 spectral bands every 5 d (ESA 2021). This data source has grown in popularity in research since its launch in 2015 because of increased spatial and spectral resolutions (e.g., Puliti et al., 2018; Kattenborn et al., 2019; Jensen et al., 2020). Sentinel-2 imagery has been used for vegetation mapping (Addabbo et al., 2016), invasive species monitoring (Jensen et al., 2020), and woody cover spatial distribution (Ng et al., 2017; Kattenborn et al., 2019). Given this higher spatial and spectral resolution, Sentinel-2 has the potential to detect mesquite trees that were not detectable with LANDSAT imagery. However, Sentinel-2 imagery is still too coarse to differentiate small trees (< 1 m² cover) in the landscape, which can result in reduced accuracy (Kattenborn et al., 2019; Daryaei et al., 2020). Within the past decade, unpiloted aerial vehicles (UAVs) have become more widely used to assess the volume, amount, and spatial distribution of woody vegetation (Laliberte and Rango 2011; Mayr et al., 2018; Jackson et al., 2020). UAVs may offer opportunities to bridge the gap between satellite imagery and field data collection.

Unpiloted aerial vehicles and higher-resolution satellite imagery have emerged as a possible solution to quantify the distribution of mesquite (Laliberte and Rango 2011; Jackson et al., 2020). The use of conventional red-green-blue (RGB) sensors mounted on a UAV can provide 2-dimensional (2D) and 3-dimensional (3D) products that can be used to better assess woody cover such as measurements of canopy structures, distinguishing different vegetation types, and amount of vegetation cover (Zainuddin et al., 2016; Mayr et al., 2018). LiDAR can provide accurate measurements of canopy volume and height for trees and shrubs, but it still remains relatively costly compared with the data derived from structure from motion (also known as digital aerial photogrammetry) (Iglhaut et al., 2019; Jiang et al., 2020). The functionality of UAVs allows for the generation of imagery at very high spatial resolutions and temporal coverage to map and monitor small patches of vegetation within rangelands that often cannot be captured by satellite imagery (Gonzalez-Dugo et al., 2013; Sankey et al., 2018). Deriving tree heights and land cover area are key components in rangeland monitoring and surveying methods (Wang et al., 2017). Recording tree heights and diameter allows managers the ability to monitor growth of the plant species, calculate the volume of the plant, and calculate forage biomass, and it can aid in determining available habitat for wildlife species (Ivosevic et al., 2015; Zainuddin et al., 2016; Guo et al., 2017; Manfreda et al., 2018; Zhang et al., 2018). Laliberte et al., (2010) were able to perform rangeland inventory assessments to monitor rangelands in Southern Idaho using UAVs with a hierarchical classification process to distinguish different features and vegetation types within the landscape. Rossi et al., (2018) were able to more accurately delineate degraded forests in northern Argentina using a combination of UAV imagery and satellite imagery.

New remote sensing approaches allow the combination of UAV imagery as training datasets for satellite imagery (Puliti et al., 2018; Kattenborn et al., 2019). The information acquired by UAVs can be used to inform lower-resolution imagery (e.g., Sentinel-2) to improve image classification and combine high spatial resolution imagery with high spectral resolution images to map and quantify the distribution of woody cover species (Rossi et al., 2018; Kattenborn et al., 2019). Puliti et al., (2018) used a combination of UAV and Sentinel-2 imagery to estimate forest growing stock volume in southeastern Norway using hierarchical model-based inference. Images acquired from UAVs have been used as training data for Sentinel satellite imagery to identify three invasive woody cover species (Kattenborn et al., 2019) and detect riparian vegetation with an emphasis on riparian forest habitats (Daryaei et al., 2020). These new technological advances can help improve our ability to assess the spatial distribution of woody encroachment in rangelands and provide additional tools in the natural resource manager's toolbox for planning and adaptive management (Laliberte et al., 2010; Manfreda et al., 2018; Al-Najjar et al., 2019).

The goal of this project was to assess the presence and spatial distribution of mesquite derived from two different remote sensing sources. The specific objectives were 1) evaluate the accuracy of using UAV-based approaches to estimate mesquite height and 2) classify mesquite vegetation presence using a combination of height metrics and spectral information by integrating UAV training data and Sentinel-2 remote sensing data.

Methods

Data collection

We collected UAV data at two altitudes: 50 m and 100 m above ground level (AGL). All UAV flights took place on sunny, clear days, with winds < 16 km • h^–1. We used a DJI Phantom IV UAV (SZ DJI Technology Co. Shenzhen, China) for Hood County and a DJI Phantom IV RTK UAV (SZ DJI Technology Co. Shenzhen, China) for Jim Hogg County and South Pasture. The DJI Phantom IV has an integrated GPS/GLONASS system (Global Positioning System/GLObal NAvigation Satellite System), allowing for faster, more precise satellite acquisition during flights, and the DJI Phantom IV RTK connects to a mobile base station that allows images to be geocoded to a higher georeferenced accuracy (1.2 cm relative horizontal accuracy) during flight. Both UAV units were equipped with a 2.5-cm, 20-megapixel camera mounted to a gimbal that stabilizes the camera with the movement of the UAV (pitch, roll, and yaw). Image data collection in Hood County was conducted on June 27, 2019 and consisted of 7 flight missions having a total flight area of 70 ha, 17 ha of which were used for model training and height analysis. Image acquisition at the Jim Hogg County ranch consisted of 2 flight missions on November 6, 2020, encompassing a flight area of 142 ha, with 46 ha used for model training and height analysis. The UAV flights at South Pasture occurred on October 14, 2020 and consisted of 2 flight missions, with a total flight area of 144 ha, 42 ha of which were used for model training and height analysis.

Natural color UAV images (red, green, and blue bands) were acquired with a double-grid flight pattern, at a speed of 3–5 km • h^–1, at 70° camera angle, and with 80% image overlap (DiMaggio et al., 2020). We chose these parameters to improve the 3D model by having oblique images to capture the vertical structure of mesquite. We used the Pix4D capture application (Pix4D Inc., San Francisco, CA) for Android for the DJI Phantom IV and the DJI Pilot application (SZ DJI Technology Co. Shenzhen, China) for the DJI Phantom IV RTK. We did not use ground control points (GCPs) in our flight missions, which provided us the opportunity to compare the two UAVs georeferenced accuracies without additional input from georeferenced points.

For each study site, we located pastures that had mesquite trees present and used ArcGIS Pro v. 2.5 (ESRI, The Redlands, CA; hereafter “ArcGIS Pro”) to create 10 random points that fell within each pasture. In the field, we navigated to each of these points and created a 22.86-m buffer area. Within the buffer area, we measured every mesquite tree's canopy height and its location (Klimas et al., 2004) using a Trimble R1 GPS unit (50-cm accuracy). The buffer area size was chosen on the basis of current Natural Resource Conservation Service practices for quantifying mesquite in rangelands (D. Daniels, unpublished data). To measure the canopy heights, we used a telescoping measuring rod and recorded each height and unique ID number for each mesquite/GPS point location. We then co-located the GPS points with the UAV imagery to identify mesquite trees and delineate mesquite presence and heights.

We acquired Sentinel-2 imagery for each study site (Hood County, Jim Hogg County, and South Pasture) from the US Geological Survey's Earth Explorer (USGS, Reston, VA;  https://earthexplorer.usgs.gov). Sentinel-2 imagery captures data in 13 bands: red, green, blue, and 1 near infrared bands at 10-m resolution; 3 red edge bands, 1 narrow near infrared band, and 2 shortwave infrared bands at 20-m resolution; and coastal aerosol, water vapor, and shortwave infrared/cirrus bands at 60-m resolution (ESA 2021). We selected cloud-free imagery collected nearest the date of UAV data acquisition. We acquired one scene on June 29, 2019 for Hood County; one scene on October 31, 2020 for Jim Hogg County; and one scene on November 17, 2020 for South Pasture.

Fig. 1.

Study sites locations: Hood Co. is in the Cross Timbers ecoregion (1); Jim Hogg Co. is in the Coastal Sand Plain ecoregion (2); and South Pasture is in South Texas Plains ecoregion (3) (Gould 1962; Fulbright et al., 1990).

Results

Pixel resolutions ranged from 1.71 cm² to 2.4 cm² for the 50-m AGL flights and 3.39 cm² to 4.76 cm² for the 100-m AGL flights (Table 1). Geolocation RMSE values were higher in the imagery acquired from the DJI Phantom IV when compared with the DJI Phantom IV RTK (see Table 1). For both the VMT and the LAS height estimations, the 50 m AGL Hood County site had the highest r² values out of all the sites with (r² = 0.95, RMSE = 24.9 and 26.6 cm) (Fig. 2). The 100-m AGL Hood County site also had the highest r² values out of the 100-m AGL flights (r² = 0.91, RMSE = 32.3 cm for the VMT; r² = 0.92, RMSE = 29.6 cm for the LAS). The estimated mesquite heights from the 50-m AGL flight corresponded more closely to the field-measured heights than those estimates from the 100-m AGL flight for both height estimation methods (r² = 0.884 vs. r² = 0.758, respectively, for the VMT; r² = 0.796 vs. r² = 0.764, respectively, for the LAS; see Fig. 2). The Jim Hogg County site had higher r² values for the VMT at 50-m AGL than 100-m AGL (r² = 0.851 and r² = 0.587, respectively, and lower r² values with LAS at 50-m AGL than 100-m AGL (r² = 0.827 and r² = 0.857, respectively). We found that the Hood County site had regression RMSE values that ranged from 24.9 cm to 32.3 cm, and the other two sites had average errors that ranged from 54.7 cm to 93 cm (Table 2). Differences between r² values and prediction r² values for the different height analysis and sites ranged from 0.0022 to 0.0452, and PRESS statistics were 5–18% higher than corresponding original error sum of squares values for the height analysis (see Table 2).

Overall accuracy assessment of mesquite presence derived from Sentinel-2 satellite imagery (Fig. 3) was > 80% for all three study sites (Table 3). The highest accuracy was seen in the South Pasture property with an accuracy of 87.5%, followed by Hood County with 84% and Jim Hogg with 80.5%. The highest values for producer and user accuracy were in the South Pasture site (Producer accuracy = 92%; User accuracy = 87%). The Jim Hogg County site yielded lower producer accuracies than the Hood County site (Producer accuracy = 70%; User accuracy = 89%) but higher user accuracies (Producer accuracy = 73%; User accuracy = 60%). The highest Kappa values were observed at South Pasture (K = 0.74), followed by the Hood and Jim Hogg sites (K = 0.61 and K = 0.56).

Table 1

Unpiloted aerial vehicle (UAV) processing accuracy separated by flight mission; geolocation root mean square error (RMSE) triangulation values in cm from Pix4D photogrammetry software processing.

Table 2

Statistical regression comparisons between unpiloted aerial vehicles (UAV)-derived height methods to field collected canopy height data; root mean square error of regressions (RMSE) in cm; PRESS statistics and predicted r² values were computed for model validation; PRESS statistics % difference is the percentage higher the predicted error sum of squares is compared with the original error sum of squares.

Fig. 2.

Canopy height comparison regressions. Rows indicate the three study sites; Y-axis, the field-recorded canopy height measurement; Columns, the combination of unpiloted aerial vehicle (UAV)-derived outputs and altitude differences; X-axis, the UAV-derived canopy height with each of the UAV methods (vertical measurement tool and LAS point cloud raster.

Fig. 3.

Image classification of mesquite. From top to bottom: UAV imagery; Sentinel-2 imagery; classified imagery using random forest prediction.

Discussion

The analysis of UAV-derived canopy height methods provided reliable measures of woody cover vegetation (Mesquite) in Texas rangelands. Our analyses of UAV point cloud data provided reliable measures of mesquite height. Our highest r² was 0.95 for both the VMT and LAS height metric (Hood County; 50 m AGL), and our lowest r² was 0.587 for VMT (Jim Hogg County 100 m AGL) and 0.764 for the LAS height metric (South Pasture; 100-m AGL). For the Hood County site, prediction errors ranged from ±24 to 32 cm. Errors were larger at the Jim Hogg and South Pasture sites with errors ranging from ±54 to 92 cm. Trees were taller at the Jim Hogg and South Pasture site, which may have increased the prediction error. Our UAV-derived canopy height results were comparable with other studies (Zarco-Tejada et al., 2014; Wallace et al., 2016; Mayr et al., 2018). Mayr et al., (2018) used a UAV at 70-m AGL to record canopy heights of Acacia mellifera, Dichrostachys cinerea, and Terminalia sericea by subtracting digital terrain models from digital surface models and compared outputs with field-recorded canopy heights, which yielded r² values ranging from 0.7 to 0.72. Zarco-Tejada et al., (2014) had r² values that ranged between 0.81 and 0.86 when comparing field-measured olive tree heights to digital surface model UAV outputs from a fixed-wing UAV flown at 200-m AGL. Wallace et al., (2016) compared aerial LiDAR- and UAV-derived 3D point cloud heights acquired at 30-m AGL and reported r² values of 0.84 (LiDAR) and 0.68 (point cloud). Our results from three different ecoregions in Texas indicate that UAVs could reliably measure mesquite canopy heights to assist in assessing canopy structure.

Table 3

Accuracy assessments using confusion matrix results of the three study sites comparing mesquite and other landscape classes; with Overall Accuracy (OA), User Accuracy (UA), Producer Accuracy (PA), and Kappa statistics.

We found that UAV flight altitude was not critical for the estimation of mesquite tree canopy heights. Prediction errors were similar for each of the flight altitudes evaluated, regardless of location. Our results at 50-m AGL (see Table 2) and 100-m AGL (see Table 2) were similar to Torres-Sánchez et al., (2015), who reported r² values of 0.90 at 50-m AGL and 0.84 at 100-m AGL. Our model verification results for tree height estimation (see Table 2) showed that our original models were robust with r² differences < 0.05 and PRESS statistic error sum of squares ranging between 5% and 18% higher than the error sum of squares in the original model. Models were more robust for the 50-m AGL flights than the 100-m AGL flights for all study sites; however, the differences in accuracy were relatively minor. Flying UAVs at higher altitude would allow users to cover larger areas for woody vegetation monitoring. Our results indicate that the differences in accuracy between the flight altitudes would not compromise the quality monitoring of mesquite heights.

We found that both UAV models proved to be able to capture accurate representations of the landscape. The DJI Phantom IV RTK provided higher horizontal accuracy (see Table 1) and provided the ability to create single large flight missions rather than multiple small missions. We observed higher r² values with the DJI Phantom IV missions at Hood County than at the other two sites. However, the Hood County flights had the higher geolocation RMSE values (see Table 1). The RMSE value, however, was not critical to the analysis because we were still able to identify individual trees that we measured in the field. We observed that mature mesquite trees often had thin branches as the highest point in the canopy, which may not be dense enough to pick up in the SfM analysis, thus resulting in reduced accuracy. Meyer et al., (1971) showed that as mesquite trees grow, most of the new stems grow on the upper branches, making them the smallest and thinnest part of the tree. With taller trees in Jim Hogg County and South Pasture (1 m to > 8 m) compared with Hood County (1–6 m), lower r² values could be explained by the presence of these thin branches, which were likely difficult to capture in the SfM analysis.

Our results indicate that UAV imagery can be used as training data for using Sentinel-2 satellite imagery and machine learning (random forest) techniques were useful for developing mesquite presence maps over large areas. Overall accuracy values for the models exceeded 80%, similar to other comparable studies. Ng et al., (2017) reported a user accuracy of 72% and a Producer accuracy of 73% specific to Prosopis in Sentinel-2 imagery, which was similar to the results reported in this study (User accuracy = 60–86%; Producer accuracy = 70–92%). Discrepancies with image classification in satellite imagery because of its coarse spatial resolution where single pixels encompass multiple landscape features are a common problem (Ng et al., 2017; Daryaei et al., 2020). To limit model confusion, we classified the presence of mesquite without trying to classify multiple landscape classes. The strength of agreement of our sites were moderate to substantial, with our Kappa values ranging between 0.56 and 0.74 (Landis and Koch 1977). The approaches presented in this study could be extrapolated to other woody plant species on rangelands. Canopy height has been shown to be one of the most important measures to determine woody cover structure and ecological value in rangelands (Zarco-Tejada et al., 2014; Wallace et al., 2016), along with providing insight into wildlife habitat (Hyde et al., 2006; Martinuzzi et al., 2009). Our results could be easily used to quantify the presence and spatial distribution of other woody plant species to quantify the amount of cover in the landscape. These methods, used in combination with time-series data, could provide useful baselines for monitoring change and spatial structure of woody cover in rangelands (Rhodes et al., 2021).

Combining high spatial resolution of UAV imagery with Sentinel-2 satellite imagery could assist with future monitoring of mesquite in rangelands. Past remote sensing methods have spatial, spectral, and/or temporal resolution limitations (Mata et al., 2018). Recent advances in satellite and UAV imagery are providing new opportunities to meet these needs. Adam et al., (2017) used a different approach than our study to classify mesquite and other landscape classes by performing random forest classification and support vector machine on World View-2 satellite imagery (1.5–2 m spatial resolution). Although they produced higher user and producer accuracies with mesquite than our results (User accuracy = 96%; Producer accuracy = 100%), this could be attributed to the higher spatial resolution (2 m vs. 10 m) or number of accuracy assessment points to study size area (164 points for 4 795 ha vs. 200 points for < 150 ha; Shao et al., 2019). Their overall accuracy (overall accuracy = 86%), however, is comparable with our results (see Table 3). Further research needs to address point density for accuracy assessment and area coverage to validate UAV classification with lower-resolution imagery.

The analysis of UAV-derived canopy height methods provided reliable measures of mesquite height in Texas. We found that the UAV flight altitude was not critical between 50 m and 100 m AGL for the estimation of mesquite tree canopy heights. We also found that using machine learning techniques, UAV imagery can be used as training data for Sentinel-2 satellite imagery to produce landscape maps of mesquite presence. Being able to combine the high spatial resolution of the UAV imagery with the high temporal resolution of the Sentinel-2 satellite imagery could improve our ability to use UAVs to monitor mesquite in rangelands. Our findings show the potential use of UAV to effectively link information on the ground, UAV-derived products, and satellite imagery for a wider variety of applications. Similar approaches can be used to determine the presence of other woody cover species in different ecosystems. The development of new multispectral cameras can provide opportunities to identify woody cover species and use fractional cover analysis to determine the cover of different species across the landscape by upscaling the information to satellite imagery. Additional spectral, temporal, and spatial resolution (e.g., next-generation Planetscope imagery) combined with UAV products present new opportunities to efficiently quantify and monitor land cover and land cover changes at scales that were not possible a few years ago.

Implications

Our study showed that UAVs can be used to efficiently assess mesquite height in rangelands in different ecoregions of Texas. These heights can provide valuable information for management decisions for woody plant encroachment. Having height metrics can also assist with determining habitat for wildlife species. Our regions of interest were limited by UAV capability to collect data (i.e., battery time), so we tested our approach in three different study areas, which provided more generalizable approaches to assess the amount mesquite presence. These assessments can be used to quantify the presence and spatial distribution of mesquite and potentially other woody plant species to help develop strategies for woody plant management in rangelands. The methods described in this study also show possibilities for managers to assess wildlife habitat, conduct rangeland inventories, and establish baselines to prioritize ecological management strategies.

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.