Grasslands are among the largest ecosystems in the world, providing important ecologic and economic services [1
]; however, they face multiple threats from climate change and human activity (e.g., conversion to cropland, biodiversity loss, expansion of invasive species), which can lead to their degradation [2
]. Woody plant encroachment (WPE) has become an important issue for grasslands in recent years. It is related to the expansion of native and non-native trees and shrubs into grasslands [3
], and has been connected to changes in primary productivity, nutrient cycling, energy flow, the structure and function of the ecosystem [3
]; these all lead to issues in rangeland management and livestock production. There exist various definitions of woody plant encroachment in the literature; except for the term “woody plant encroachment”, the terms “woody plant invasion” [4
], “woody thicketization” [5
], “woody plant expansion” [6
], “invasion of woody weed” [7
], “xerification” [8
], and “invasion of shrubs” [9
] are also used. This is because WPE is a global phenomenon, and definitions depend on the precipitation gradient of the region. In particular, WPE occurs in the grasslands of the south-central and southwestern United States (mesquite and creosote brush) [10
], North America (juniper) [11
], South America (honey locust) [12
], Southern Africa (Acacia
], Australia [14
], Mongolia [15
], Europe [16
], and the Arctic (willow and Alnus
WPE also takes place in the Canadian prairies, where tree encroachment (e.g., aspen, willow) has received more attention in the literature [18
]. For instance, trembling aspen (Populus tremuloides
) is the dominant tree species encroaching on grasslands and pastures within the aspen parkland ecoregion in western Canada [24
]. Other species, such as willow (Salix
spp.) and Balsam poplar (Populus balsamifera
) are also encroachers, but to a lesser extent. Thorny buffaloberry is an encroaching species in Alberta [25
]. The most common encroachers that occur throughout several Canadian prairie ecoregions (i.e., aspen parkland, moist mixed grassland, mixed grassland) are western snowberry (Symphoricarpos occidentalis
] and wolfwillow (Elaeagnus commutate
). Therefore, these two shrub species will be the main focus of this research, since they have been less studied. Moreover, the province of Saskatchewan will be our study area, since it includes the three previously mentioned ecoregions. An example of an encroaching shrub species in the rangelands of southern Saskatchewan is western snowberry, found in the commercial rangelands and provincial pastures of the Grand Coteau region and Weyburn. One can also find western snowberry and wolf willow in Burstall rangelands, the Northeast Swale of Saskatoon, Meewasin Valley, Kernen Prairie, and most of Saskatchewan’s southern provincial parks (pers. comm. Mr. Merek Wigness, Dr. Eric Lamb, Dr. Thuan Chu, and pers. observ.). It is understood that shrub encroachment is either already an issue or might become an issue in most of southern Saskatchewan’s rangelands. Nevertheless, the cover of these species within the prairies is currently unknown.
It is clear that maintaining grassland health is crucial, especially when food scarcity is estimated to rise, and sustainable management solutions are needed [27
]. This fits within the United Nations Sustainable Development Goal 15.3 on “Land degradation neutrality”. Remote sensing can be used with success to fulfill this aim by mapping the spatiotemporal distribution of various encroaching species with the use of different methods and datasets [28
]: for instance, to detect two Acacia
species from hyperspectral imagery with the use of differences in their phenology in Namibia [30
], to classify Prosopis
spp. with an object-based approach in Kenya [31
], to detect redberry juniper and honey mesquite in north central Texas with spectral contrast of a three-band aerial image [32
], to classify three woody invasive species with spectral, textural, and structural features in Chile [33
], and to detect six types of woody species with multispectral aerial imagery and LiDAR derived heights in the Netherlands [34
]. Overall, for species-specific detection, high spatial resolution is necessary. However, the use of high spectral and temporal resolution could compensate for the lack of spatial resolution, and is more preferable for regional and landscape scale mapping. Furthermore, when thinking about the phenological behavior of each woody species of interest, it might be necessary to define the optimal detection timeframe within the growing season for each one. We therefore focus our study on a seasonal spectral approach. Hyperspectral data have been used to detect WPE species due to their wide band range, which allow for the detection of finer spectral differences. In addition, field-based hyperspectral measurements offer the opportunity to fine-tune spaceborne and airborne sensors for larger-scale shrub species mapping by selecting appropriate spectral bands and regions with spectral separability metrics and statistics (e.g., InStability Index, Transformed Divergence, etc.). Afterwards, one can define remote sensing indices that use these bands and apply a broader land cover classification.
To our knowledge, no study has looked at seasonal hyperspectral and multispectral differences between western snowberry and wolfwillow for their potential detection with remotely sensed data, which can facilitate WPE management in the Canadian prairies. Therefore, the main purpose of this study is to derive the seasonal sensitive spectral regions for separation between western snowberry and wolfwillow shrubs in grasslands. Our main objectives are (1) to identify the optimal season for detection of the two shrub cover types, (2) to determine the sensitive wavelengths and bands that allow for their separation, and (3) to investigate differences in separability potential between a hyperspectral and broadband multispectral approach.
4. Discussion and Conclusions
Our results from the hyperspectral metrics, broadband metrics and two-sampled t
-tests show that the summer season is the optimal one for the spectral separation of western snowberry and wolfwillow, as it has the highest number of significantly different spectral regions and bands. This is reasonable, since the summer is the peak of the growing season with the highest photosynthetic activity, during which differences between shrub species could become more obvious. For this reason, the summer season has been frequently selected for data acquisition when separating shrub species due to the higher vigor of vegetation in that season [31
]. Summer months have also shown better discrimination abilities compared to other months—even for separating an evergreen and a deciduous species [29
]. As for the optimal wavelength regions and bands, both blue and red are important, and more so in the summer. These two regions are influenced by stronger chlorophyll absorption for western snowberry compared to wolfwillow, based on their spectral signature. On the other hand, the green peak (around 550 nm) is similar for both shrubs, and is therefore not useful for classification in the spring. Nevertheless, this spectral region is moderately important during summer, where the reflectance of wolfwillow is significantly higher than that of western snowberry. Lastly, in the far-SWIR, there is moderate separation for a narrow hyperspectral region in spring, which is not represented in the broadband simulations. Although this region is significantly different in all seasons based on the two-sampled t
-test, it is not strong enough to reflect its difference in the separability metrics. This region is most possibly related to the differences in water and moisture absorption between the two species.
Overall, when looking at the differences between the hyperspectral and broadband results for the separation of the two shrubs, we notice that the results are almost the same, except for a narrow region in the far-SWIR, which is not included in the broadband results. This leads us to the conclusion that hyperspectral data would not really improve the classification results for our specific study purposes, and that use of Landsat 8 or Sentinel-2 data would suffice. In addition, the increased number of spectral bands that Sentinel-2 data provide do not offer enhanced detection capabilities, since the NIR region that includes the red-edge bands is not one of the sensitive regions for western snowberry and wolfwillow classification throughout the seasons.
However, we must point out that our current simulated broadband results represent the leaf/branch scale and not the canopy scale. The reflectance properties of the two shrubs could be different at that scale due to canopy architecture, such as leaf angle distribution, density, biomass, and leaf area index, in which shadows and occlusions also play a role. In addition, these simulations do not represent satellite data conditions, which are strongly affected by the atmosphere, and which capture the land surface at a broader scale, in which topography also has a significant role. Furthermore, since Landsat and Sentinel-2 data capture the surface at a broader scale (10–30 m), each image pixel is usually a mixture of different land cover types (e.g., woody plants, grass, bare ground, rock). This is especially the case when WPE is at an early stage. Overall, grasslands can undergo different WPE stages (i.e., early, moderate, or advanced), resulting in different woody cover within an image pixel [78
]. A field-based study showed that the earliest WPE that could be identified was when it reached between 10% and 25% of an image pixel [37
]. However, more research with remotely sensed imagery is needed to verify this result. Nevertheless, even with mixed pixels, there exists a number of spectral unmixing techniques that could enhance WPE species specific mapping with coarse resolution pixels [79
]. With this technique, each pixel gets assigned to a fraction of its land covers, which are defined by endmembers. Two endmember classes that could be used for that purpose are the spectral signatures of western snowberry and wolfwillow that were used in this study. For the above reasons, the optimal season and bands detected in the current study for separation between the two woody shrubs might not coincide with their actual detection on the landscape. Therefore, the current results should be cross-validated with satellite-based remote sensing data, such as Landsat 8 and Sentinel-2. We plan to implement this in future research that will establish specific broadband multispectral indices optimally correlated with the two shrub species of this study, and with research that will investigate potential improvements in their detection with spectral unmixing techniques.