Keywords: rangeland, fine fuels, medusahead, SfM, biomass

Introduction

Semi-arid ecosystems act as carbon sinks and are thought to play a major role in global interannual carbon variations (Ahlström et al. 2015). In addition to soil carbon sequestration, semi-arid ecosystems at regional scales provide important ecosystem services such as wildlife habitat and mesic refugia. Semi-arid ecosystems, such as in the western U.S. are water-limited and pressures from climate change and population growth make these lands vulnerable to degradation. Specifically, the sagebrush-steppe ecosystem is currently under threat from the invasion of exotic annual grasses, such as cheatgrass (Bromus tectorum L.) and medusahead (Taeniatherum caput-medusae (L.) Nevski), which are decreasing biodiversity (Knapp, 1996), altering the fire cycle (Bradley et al., 2018), and reducing carbon storage (Bradley et al., 2006).
Quantifying vegetation biomass is critical for assessing ecosystem structure, tracking vegetation growth, and quantifying carbon storage (Houghton et al., 2009). Above ground biomass (AGB) in semi-arid ecosystems helps scientists and land managers better understand the contribution of these ecosystems to the global carbon flux, and the impacts of ecosystem shifts towards desertification (Chambers et al., 2014). In particular, AGB is defined as the dried weight of vegetation above the ground including both alive and dead components, is difficult to accurately measure in semi-arid ecosystems because of the heterogeneity and fine-scale structure of vegetation (Fern et al., 2018; Wijesingha et al., 2019). This is especially so for invasive annual grasses such as medusahead and cheatgrass, where previous growth can become a mat of dry litter, decreasing native vegetation growth and further promoting the invasive annual grass species (Evans & Young 1970). As a result, an increase in the continuity of fine fuel loads occurs. This build-up of fine fuels, coupled with senescence early in the growing season promote ignition and increased fire spread. This feedback loop, ‘invasive grass-fire cycle’ further results in degradation of semi-arid ecosystems (Fernández-Guisuraga et al. 2022).
Management approaches to the invasive grass-fire cycle are varied across techniques and spatial and temporal scales. Approaches such as re-seeding, grazing, and fuel breaks, aim to effect and observe changes in a plot, pasture, or regional scales (100s-1000s km2). Where the sites may be relatively small in extent (e.g. grazing exclosure scales, or m to km), there is a benefit of reducing the need to destructively harvest samples to observe changes while reducing impact on the vegetation. Monitoring data at high spatial resolutions and with low uncertainty are needed to provide baseline information to both of these scales to be able to extrapolate measurements with greater certainty, and where changes over time and in response to treatments and climate are valuable to observe.
How AGB is Measured in the Field, Destructively and Non-Destructively
Current methods for collecting AGB can be described as a combination of site-specific and extrapolated measurements. An example of site-specific measurements is the destructive harvesting of biomass. Often quantification of biomass at the plot scale is extrapolated to larger scales (Clark et al., 2008). Further, in the field plot-level data collection of plants are destructively harvested, dried, and weighed to obtain a biomass metric, or point frame data is used to relate to biomass. The manual process itself has uncertainty in the collection from human error. Often these plots are small, and the biomass can vary significantly across spatial scales, even within 1 m or less. The act of removing plants itself alters the landscape and impacts future studies of those plots. Advances in remote sensing systems like lidar, unmanned aerial systems (UAS), and structure from motion (SfM) software have led to advances in quantifying biomass in dryland ecosystems (Anderson et al., 2018; Cunliffe et al., 2016). Fine (spatial/structural) resolution remote sensing has potential to help solve the above rangeland vegetation challenges. SfM is photogrammetry, the method of using 2D stereoscopic images and detecting common points, resulting in a 3D reconstruction in the form of a point cloud. SfM provides a similar data product to lidar: a point cloud in which vegetation structure can be interpreted. However since SfM is based on optical passive imagery it does not penetrate the canopy (Salamí et al., 2014). SfM and lidar are similar in their point cloud reconstruction, however because lidar is active, it has an intensity associated with each point. Additionally, lidar can be collected as full-waveform or in discrete returns which can characterize the understory and ground surface (Wallace et al. 2017).
In contrast, SfM is derived from passive remote sensing relying on optical imagery to create point clouds where each point has an associated color value for the spectra of the optical image. This value is rarely radiometrically calibrated. SfM point clouds are not discretized from a waveform but developed based on the overlap of imagery taken. Vegetation parameters such as volume, can be derived from point clouds to develop allometric relationships between vegetation and destructively harvested biomass. SfM offers a low-cost, time efficient, and in some cases, more accurate method compared to terrestrial laser scanning (TLS) for estimating vegetation structure in dryland ecosystems (Cooper et al. 2017, Olsoy et al., 2018; Wallace et al., 2017).
Extremely close-range SfM, defined by the use of hand-held instruments to collect the imagery for the SfM method, has shown success comparable to TLS in deriving parameters such as height and volume from the reconstruction of individual trees (Miller et al., 2015). In grasslands, extremely close-range SfM was shown to outperform TLS and traditional height measurements in developing allometric equations, in part because SfM can capture finer details compared to TLS, dependent upon the experimental setup (Cooper et al., 2017). TLS at plot scale (Anderson et al. 2017) and TLS and hand-held camera photos at 0.5 m x 0.5 m (0.46*0.46, Cooper et al. 2017) had high correlations between SfM and TLS and AGB. SfM has been shown to successfully capture grasses and reconstruct fine features such as leaves and stems (Cooper et al. 2017, Kröhnert et al., 2018). These previous studies demonstrate that SfM shows promise in providing accurate information for allometric equations used to extrapolate biomass of rangeland grass to a landscape-scale.
Field work for capturing extremely close-range SfM imagery requires less training for field crews compared to TLS and UAS data collection and eliminates the need for specialized equipment. Developing an allometric relationship at a plot scale (cm) will inform future SfM studies that utilize UAS imagery (Gillan et al., 2019). Furthermore, close range SfM can be complementary and used as a precursor for extrapolating to scales that are amenable to commonly available UAS (e.g., m to km). However, vegetation and soil classification in such a fine resolution dataset remains challenging. Spectral reflectance in the visible spectrum and near-infrared may be able to better separate vegetation and dry biomass litter from one-another, or from the mineral soil or ground surface. Although differences in the soil composition can cause different types of confusion with vegetation and litter, possibly reducing transferability of methods between study sites (Huang et al., 2007), in areas where the ground surface can be largely obscured may benefit from other approaches such as structural analyses from either SfM or active remote sensing technologies may be more successful (Calders et al., 2015).
A common technique for estimation of AGB relies on Digital Terrain Model (DTM) and Digital Surface Model (DSM) computation and the retrieved Canopy Height Model (CHM) and Leaf Area Index (LAI) from them (Ota et al., 2015, Zhang et al., 2018, Xu et al., 2019). The biomass from lidar and UAS-derived point clouds CHM is generally classified into ground and vegetation by setting a height threshold (Cunliffe et al., 2016, Viljanen et al., 2018, Näsi et al., 2018, Navarro et al., 2020). However, the accuracy of CHM mostly exceeds the height of short shrubs and grasses, and thus CHM models are less suitable for aboveground biomass estimation in dryland ecosystems (Zahawi et al., 2015). Schulze‐Brüninghoff et al. (2021) suggested a method combining lidar-derived metrics (sum of voxels, canopy height model, and canopy surface structure) and vegetation spectral properties to estimate fresh and dry biomass in a machine learning approach. In addition, by combining ground camera and UAS, Taugourdeau et al. (2022) extracted vegetation metrics from RGB bands and height to estimate aboveground biomass in a random forest model. Therefore, the uncertainty of biomass estimation is limited to the uncertainty in CHM, data resolution, and the type of land treatments (grazed or ungrazed). In this regard, the discrimination between vegetation and bare ground structural properties (smoothness of the surface, randomness of vegetation foliage, and curvature) could improve AGB estimation from high-resolution ground camera and UAS SfM-derived point clouds.
The objectives of this work are to explore the potential for close-range SfM to quantify fine-fuel AGB, with the thought of replacing destructively harvested sampling with the methodology, or to utilize the close-range SfM volumetric estimations to inform potential UAS data collection. To accomplish this, we first developed a method to acquire and process field data from photos to point clouds. Second, we explored how to separate and classify the ground surface through the thick litter layer found in the study area. Third, we examined the relationship between the measurement unit of volumetric calculation and the biomass of the study area’s vegetation. Lastly, we discuss some observations on how the SfM point frame might be used, and how this method may inform UAS data seeking to estimate biomass or vegetation volume.

Methods

  1. Study Area
Our study area is the Three Fingers allotment located in southeastern Oregon (Fig. 1), approximately 80 km west of Boise, Idaho. The Vale District U.S. Bureau of Land Management (BLM) manages the allotment, which has a history of disturbance from grazing, recreation, and fire. The area has experienced significant departure from native vegetation communities, with few remaining native shrubs and an extensive cover of cheatgrass and medusahead, and other annual grasses (>80% in all experimental exclosures).
This study was part of a larger study focused on evaluating the effect of dormant season grazing on fine fuels (Arispe et al. 2022). As part of the project, data were collected across two years and on three different pastures on the Three Fingers allotment—approximately 55,000 hectares. The three pastures within the project included, McIntyre (MCI; 3100 ha), South Camp Kettle (SCK; 2500 ha), and Saddle Butte (SB; 3800 ha). Within each pasture a northern and southern exclosure was randomly placed, for a total of six treatment areas. Each exclosure contained four paddocks 150 m x 150 m in size, for a total of 24 paddocks.