1. Introduction
Invasive species are a leading cause of biodiversity loss [
1] and a threat to many ecosystem services [
2]. These species are also referred to as non-native, alien, or exotic, and can cause harm to the environment by outcompeting native species for food and other resources, causing direct impacts on resources and indirect impacts on the growth and vitality of other native species [
3]. In arid regions such as the American southwest, controlling certain introduced and invasive species that threaten water resources has become an issue of great concern during recent decades [
4,
5,
6].
Accurately mapping species level vegetation is an essential step in managing invasion risk [
7], guiding remediation efforts and intervention strategies [
8], monitoring outcomes of management actions [
9], and ultimately understanding what processes are facilitating growth or expansion [
10]. These activities are particularly important for justifying and sustaining public support of management programs [
11], especially on public lands. However, mapping invasive vegetation at the species level using traditional platforms, such as Landsat or MODIS, is difficult due to the coarse spatial resolution of the imagery (i.e., 30–500 m). Many plant species are much smaller than a Landsat pixel, making discrimination difficult [
12]. A host of methods have been developed to spectrally unmix signals from platforms such as Landsat [
5,
13], but these methods are unable to map the spatial distribution of different species at the subpixel scale [
14]. Imagery captured by unmanned aircraft systems (UAS, or drones) at high spatial resolutions can support targeted management efforts [
15,
16,
17], but pipelines for capturing, processing, and analyzing these data do not always leverage the full range of possible products, and they can be expensive and time consuming, reducing their effectiveness [
18].
Structural characteristics, such as canopy height and terrain, have generally been omitted from remote sensing classification workflows, but they can provide key information on hydrology and geomorphology. UAS provide the ability to capture images with sufficient overlap to apply modern photogrammetric structure from motion (SfM; [
19]) workflows to develop high resolution digital elevation and surface models [
20]. Several studies have begun integrating 3D structural layers derived from UAS images using an SfM workflow with spectral orthomosaics in the classification process [
21,
22,
23,
24]. The benefits of an integrated spectral–structural approach are that the spatial and spectral information from the high resolution UAS images can be leveraged directly for species-level discrimination while also incorporating structural landscape characteristics that may not manifest in spectral signatures but might impact the spatial distribution of species, such as the geomorphology. Additionally, while spectral signatures may change throughout the year based on phenology and environmental changes, structural characteristics will remain more temporally stable and can thus potentially be used to predict which areas may be prone to invasion, overcoming a limitation of a purely spectral approach.
This paper develops a spectral–structural workflow for mapping several invasive species in an arid region that is prone to flooding. The workflow combines drone images with SfM and a machine learning classification approach to map vegetation species. Drone images were captured and processed into a spectral orthomosaic and structural models including a digital terrain model (DTM) and canopy height model (CHM). The spectral orthomosaic was then used to derive vegetation indices, while the DTM was used to derive a hydrology-based flow accumulation model of the site. The spectral and structural data layers were then used as inputs into a spectral-only and spectral–structural random forest classification schemes to map vegetation species. The area of focus is the Lower Salt River area in central Arizona in the Tonto National Forest, which has been the site of both fire and flooding during recent years. The workflow described can be applied regionally or in other similar environments to effectively monitor and manage restoration efforts.
5. Discussion
Past studies using UAS for terrestrial investigations have tended to focus either on the development of 3D models (i.e., DTM, DSM, etc.) for terrain, geomorphic, or other similar analyses, or they have focused on the development of accurate, high spatial resolution multispectral orthomosaics for creating classifications. However, structural information from the 3D models also has value for classification, particularly vegetation discrimination. In this study, we compared a spectral-only model, which included only spectral layers, to a spectral–structural model that included both spectral and elevation layers to differentiate invasive species in an arid study region prone to fire and flooding. The spectral–structural model outperformed the spectral-only model, with overall accuracy increasing to 93% compared to 80% for the spectral-only model. The vegetation in the study area, particularly non-native species, is known for opportunistically establishing and spreading after flood events. Since flooding and water flow patterns are closely related to the terrain elevation, the importance of elevation uncovered in the classification here is not unexpected. However, the degree to which the accuracy of the classification model improved with the incorporation of the DTM and CHM structural layers suggests they may be more important than previously considered, particularly for classifying invasive vegetation in the study area.
The CHM and DTM combined to account for more than 30% of the variable importance in the spectral–structural model (
Figure 10). According to the SEINet data Portal for southwestern biodiversity, there is considerable variation in the height of the vegetation species in the study area (
Table 3) [
60], which may explain why the CHM layer was highly important in the combined spectral–structural model. In short, vegetation species in the region may be more differentiable based on plant height compared to their spectral differences. These differences likely led to the CHM being more important than any of the spectral layers for classifying vegetation. More broadly, the importance of the structural variables in the classification model is important for future studies because the structural characteristics of a landscape often do not change as frequently as the spectral characteristics. While spectral signatures are prone to seasonal and phenological changes as well as water and nutrient inputs [
61,
62], structural signatures of the plants themselves such as canopy height and the elevation of the ground where they are growing do not change as dynamically. The findings from this study suggest there is great potential to use structural information such as terrain elevation to understand which areas may be at risk for future invasion. For example, if saltcedar distribution is predicted well through a digital terrain model, similar elevations can be proactively treated following a disturbance event (e.g., flood) to prevent establishment without having to wait for the plants to physically establish themselves in order to detect a spectral signal. Another benefit of this finding for land management is that digital terrain models already exist for much of the world (although not necessarily at the high spatial resolution used here). Depending on their resolutions, pre-existing digital terrain models may provide initial insights into which areas are at risk for future invasion. Global canopy height and other 3D structure data has also recently become available through the Global Ecosystem Dynamics Investigation (GEDI) high resolution laser deployed on the International Space Station [
63].
Given the importance of the DTM layer in the classification, and the wide availability of DTM datasets worldwide, a logical next study would be to test how the spatial resolution of the DTM affects classification accuracy. An investigation into the optimal spatial scale could aid in determining the coarsest resolution at which structural information can provide key information for species discrimination. The spatial resolution used in this study was 0.16 m, which is moderate for a drone study. Selecting an appropriate minimum mapping unit (MMU) has long been recognized in remote sensing studies of land cover [
64], where MMU is the area of the smallest entity to be mapped. For land cover studies, some scholars have suggested that MMU should be 2–5 times smaller than the smallest object of interest [
65], but similar thresholds have not yet been determined for classifications involving a terrain model. Elevation values vary continuously, whereas land cover can transition more discretely (e.g., from plant in one pixel to water in the next). Determining an appropriate MMU for structural layers such as DTMs can help users balance accuracy needs with data volume and processing costs [
40]. Even if UAS-acquired DTMs are required, coarser MMUs can minimally translate into higher flying altitudes with less time and fewer images needed to cover the study area, saving time, money, and resources.
Interestingly, while the CHM and DTM were found to be important to the classification model, flow accumulation only marginally contributed. FLOW was derived directly from the DTM, and is also correlated with DTM. Therefore, it is possible the FLOW layer simply did not contribute any new information to the model. If DTM was removed from the model, the importance of FLOW may increase. The BLUE band was found to be the most important spectral layer in the spectral-only and the spectral–structural classification. Blue light is scattered considerably by atmospheric constituents, which has made it challenging to use this information from satellite imagery because the band can be noisy. When flying UAS at low altitudes, atmospheric scattering effects are often reduced [
40]. Recent research using other close-range remote sensing methods found that blue and even ultra-blue wavelengths hold potential for vegetation discrimination and estimating biophysical (e.g., chlorophyll) components [
66]. Our findings suggest that sensors specifically designed for close-range UAS that capture reflectance in the blue wavelength regions where chlorophyll a and b are absorbed may aid vegetation discrimination.
Lastly, in terms of public land management, saltcedar is generally absent from the area that has been chemically treated for giant reed, and these areas also have a more heterogeneous mixture of vegetation (
Figure 11). In the areas that have been mechanically treated for giant reed, saltcedar appears to be thriving, and in the areas that have not been treated at all for giant reed, saltcedar appears in dense stands (
Figure 11). Saltcedar is notoriously difficult to remove mechanically [
67] and can reproduce adventitiously, making follow-on treatments necessary. These classification results along with the spatial overlays of treatment areas suggest that treatments should be targeted based on the type of vegetation and location within the study area, particularly since mechanical treatments can be difficult due to the need for heavy machinery [
67].