Uses of Spectroscopy for Tundra Studies

4.1. Long-term Vegetation Changes with NDVI

Long-term satellite data has revealed “greening” of Arctic tundra since the 1970s based on increases to NDVI derived from Landsat and AVHRR time-series data (Myers-Smith et al., 2020; Wang & Friedl, 2019). While tundra greening remains the most common trend across the Arctic, “browning”, represented by a decreasing trend in NDVI values, has occurred in various regions and scales across the tundra (Myers-Smith et al., 2020). Greening and browning trends were one of the first indications that the Arctic tundra was being significantly impacted by climate change. Variations in greening/browning over different years have most commonly been attributed to climate warming (e.g., (Berner et al., 2020; Bhatt et al., 2021; Cooper, 2014)), herbivory by small mammals (Olofsson et al., 2012), and vegetation disturbance and subsequent recovery after extreme warming events (Bokhorst et al., 2012). However, there are significant limitations of the sensitivity of NDVI to high latitude ecosystem change (Huemmrich et al., 2021). For example, recent evidence suggests that some of these changes’ impacts are fine-scale in nature (i.e., < 5-30 m), making many common remote sensing platforms impractical for studying these dynamics (Myers-Smith et al., 2020; Niittynen et al., 2020). Moving beyond greening and browning into the shifting landscape of numerous other metrics unlocked by IS, such as changes in land cover type and biophysical traits, will provide key insights into the magnitude and nature of high latitude ecosystem change.
From the outset, advanced IS data collections, such as from SBG, should be organized and calibrated to allow for future analysis of multi-year trends. In addition, improved land cover descriptions from SBG will enhance the interpretation of the existing NDVI trend analyses by establishing the capacity of different land cover types to respond to environmental change and for that change to be reflected by observable changes in NDVI. Ground measurements collected over extended time series will improve our understanding of the nature of spectral reflectance change associated with measured land cover change and inform remote sensing needs.

4.2. Land Cover and Vegetation Classification

Surface reflectance data have long been used to classify and map vegetation types from landscape to global scales. Accurate data identifying the distribution of and changes to land cover types provide a significant opportunity for understanding Arctic environmental change. Improved mapping and classification of circumpolar land cover and its changes will be key to understanding the effects of global environmental change on Arctic ecosystems (Sections 4.2-4.4). Overcoming the challenges associated with mapping land cover at appropriate levels of thematic, spatial, and temporal detail will ultimately provide a significant advancement in our understanding of Arctic ecosystems.
Mapping Arctic vegetation types at high spatial resolution and with sufficient thematic detail has been challenging in part due to a relative sparsity of spectroscopic data. Global-scale land cover maps, such as the MODIS land cover product (Sulla-Menashe et al., 2019), are typically produced at a level of thematic detail that cannot distinguish between functionally distinct landforms (e.g., low- versus high-centered polygons) and vegetation types (e.g., low versus tall shrublands) present in Arctic tundra. Different arctic vegetation types are often combined into simpler, but less effective classes, or are represented by inappropriate classes (e.g., “grassland”) which do not reflect tundra ecosystem composition. The utility of land cover maps for tracking Arctic environmental change hinges on improving land cover classification, as subtle changes in vegetation properties, such as increased shrub abundance, do not necessarily involve a transition from one class to another within a mapped pixel.
Moving beyond land cover types and into the mapping of plant functional types or finer taxonomic groups from spectra may be possible at continental scales if IS data with large spatial coverage (e.g., ABoVE airborne campaigns (Section 2) and SBG) are harnessed and developed. Acquiring and applying more detailed spectroscopic data for Arctic vegetation types will enable mapping with improved thematic detail, particularly if they are analyzed in tandem with ancillary high spatial resolution datasets that capture important environmental covariates such as topography (e.g., ArcticDEM) and edaphic characteristics (e.g., seasonal inundation, snow depth and hardness, active layer thickness, depth to water). Few studies have yet applied detailed IS data to map Arctic vegetation types (Smith et al., 2021; Thomson et al., 2021), but an increase in available imagery may enable future work in this area.
Land cover maps with classifications designed for Arctic vegetation types are typically limited in spatial or temporal range (Chasmer et al., 2014; Greaves et al., 2019), precluding comprehensive study of Arctic vegetation dynamics, or are coarse in spatial or temporal resolution (e.g., gridded 1 km CAVM) (Raynolds et al., 2019), precluding accurate characterization of the high level of spatial heterogeneity and temporal variability in Arctic vegetation. (Bartsch et al., 2016) suggested that a 30 m spatial grain, which is the proposed spatial resolution for SBG, is sufficient for capturing many of the dynamics of Arctic land cover. However, depending on whether species-level or functional type-level maps are being generated, even higher spatial resolution (e.g., 3 m from Planet) may be insufficient to distinguish Arctic vegetation except at broad thematic levels (e.g., trees vs. shrub vs. water). Therefore, the use and further development of advanced subpixel mixture analysis will enable high accuracy vegetation classifications with reasonable instrument spatial resolution and broad spatial coverage (Thomson et al., 2021). Tapping the information content of higher spatial resolution data (e.g., Section 2) will be essential to preparing the algorithms and analysis pipelines to utilize a spaceborne imaging spectrometer such as SBG that has a finer spectral resolution occurring at an intermediate spatial resolution to map Arctic vegetation (Section 5).
Another key limitation to mapping vegetation in Arctic tundra is the lack of high-quality, georeferenced training data. Existing observations are scattered across numerous countries, land management agencies, and historical datasets. Disparate datasets often do not capture similar levels of detail, and thus can be challenging to integrate. Land cover maps, and the algorithms and data that go into producing them, are only as credible as the underlying training data. Typically, land cover maps are trained on datasets of land cover type that are produced by visual interpretation of very high spatial resolution imagery (e.g., using Google Earth), but the availability of suitable (midsummer) imagery is extremely limited in the Arctic tundra (Section 1). Field data provide the most reliable source of georeferenced Arctic ground verification, but they are inherently limited in scope and are spatially biased towards areas with a long history of research (e.g., northern Alaska’s Dalton Highway corridor). Airborne data (including UAS observations) can bridge the scaling from field data to spatially extensive gridded datasets (Assmann et al., 2020). This scaling will ultimately enable training of machine learning algorithms to effectively map Arctic vegetation at continental scales.
Finally, the unique seasonal characteristics of the Arctic impose additional challenges on mapping tundra vegetation at scale. Phenological differences can help to separate co-occurring and spectrally similar plant functional types (Macander et al., 2017), but the phenology itself is highly variable through space and time since it is sensitive to moisture status and interannual variability in meteorologic conditions (Sections 4.4 and 4.5). Land cover mapping algorithms may misinterpret spectral changes caused by interannual variation as real land cover change. The brief snow-free season in the Arctic tundra may inhibit sufficient characterization of phenology-driven spectral changes, which further reduces our ability to identify spurious change detection. A sufficiently large and representative training dataset, as described above, will help prevent vegetation mapping algorithms from misclassifying changes in moisture status and phenology with changes in land cover in the Arctic Tundra.

4.3. Retrieval of Biophysical Properties and Plant Traits

The strong connection between IS and the biophysical properties of plant leaves and canopies makes it possible to retrieve a host of important vegetation properties with spectroscopy (Serbin & Townsend, 2020). Particularly, the mapping of plant functional traits, i.e., the morphological, biochemical, phenological, and physiological attributes of leaves and canopies (Violle et al., 2007), has been a priority and key focal area of study (Asner et al., 2015; Asner & Martin, 2008; Cawse-Nicholson et al., 2021; Singh et al., 2015; Z. Wang et al., 2019, 2020) . These functional traits, which are closely related to vegetation establishment, growth, and functioning, are key to understanding vegetation responses to climate change, as well as process modeling of terrestrial ecosystems (Gamon et al., 2019; Myers-Smith et al., 2019; Xu & Trugman, 2021; Zakharova et al., 2019). For example, traits that describe leaf photosynthetic capacity (e.g., foliar pigments, nitrogen, and Vcmax), biogeochemistry (e.g., ligno-cellulose, carbon, and macronutrients), and water cycling (e.g., stomatal conductance) are important to characterize ecosystem carbon, water, and energy cycling and response to climate change (Chapin, 2003; Myers-Smith et al., 2019; Ollinger & Smith, 2005; Rogers et al., 2017; Tang et al., 2018; Woodward & Diament, 1991) . Similarly, traits related to vegetation structure, such as leaf area and canopy height, are important for determining ecosystem energy partitioning (e.g., through surface albedo and temperature), as well as surface-atmosphere interactions (Aalto et al., 2018) that feedback to the global climate system (Zhang et al., 2018).
In the Arctic, plant functional traits vary remarkably within and across plant species and over space and time, controlled by the complex, fine-scale gradients in climate, topography, water, and nutrients (Andresen & Lougheed, 2021; Bjorkman et al., 2018; Black et al., 2021; Chen et al., 2020; Thomas et al., 2020). In particular, traits that confer differing competitive advantages, such as those related to plant size and resource economics (e.g., leaf area, seed mass, height, LMA, N, LDMC) (Thomas et al., 2020), are highly sensitive to changes in micro-environments, making them difficult to characterize with traditional field surveys (Metcalfe et al., 2018; Schimel et al., 2015). In addition, the photosynthetic capacity (V cmaxand J max) and response to environmental conditions of Arctic plants are significantly different from the current assumptions in the process models used to forecast Arctic change (Rogers et al., 2017).
Non-vascular plants which dominate large areas of the Arctic, have very different biochemical attributes and possess morphologies that are not yet easily measured (Sections 3.2-3.3) (Holt & Nelson, 2021). Water content varies in non-vascular plants based almost entirely on environmental conditions since they do not actively conduct water, which greatly influences their spectral signatures (Figure 4). Variable water content in the non-vascular ground layer visible to remote sensing instruments presents a primary challenge and significant opportunity to understand ecosystem function where they dominate. Methods using a combination of VNIR, SWIR and MIR show promise for addressing water content in non-vascular plants (Granlund et al., 2018; Neta et al., 2010). Testing these estimations of water content at large spatial scales remains a challenge. Most traits in non-vascular plants exhibit different spectral responses from those of vascular plants (Cornelissen et al., 2007), precluding direct use of existing trait retrieval approaches developed for vascular plants. Recent work by (Thomson et al., 2021) shows that chemometric estimation in non-vascular plants using remote sensing is possible but there are only a few species studied over a small area. Collectively these challenges have created significant uncertainties in our understanding and modeling of Arctic ecosystems (Fisher et al., 2018; Metcalfe et al., 2018; Myers-Smith et al., 2019). Developing algorithms to estimate properties of non-vascular plants using spectra and remote sensing will enable more accurate quantification of plant functional traits.
IS can provide a tool to spatially map a variety of plant functional traits across scales (e.g., from watershed to biome) which has been demonstrated in many other biomes (e.g., Asner & Martin, 2008; Martin et al., 2008; Singh et al., 2015; Z. Wang et al., 2019, 2020). The launch of SBG and other IS missions (e.g., EnMAP) will provide important data to further enable spatiotemporal mapping of traits across the broader Arctic tundra biome (Cawse-Nicholson et al., 2021) . Simultaneously, spectral data from aircraft (e.g., Miller et al., 2019) and low-altitude, near-surface platforms, including automated trams (John A. Gamon et al., 2006; Goswami et al., 2011; Healey et al., 2014), tower-mounted instruments (e.g., Drolet et al., 2014; Hilker et al., 2011), and unoccupied aerial systems (Assmann et al., 2020; Cunliffe et al., 2021; Shiklomanov et al., 2019; Yang et al., 2020), have increased in northern high latitudes. These diverse spectral platforms are likely to revolutionize our means for collecting trait information, which could usher a new paradigm in our understanding and modeling of Arctic vegetation dynamics. For example, using traits derived at watershed and larger scales, the spatial variation in traits across plant species, plant functional types (PFTs), communities, and even ecosystems can be easily characterized (Figure 6). The drivers of spatial variation in plant traits can also be investigated in combination with other core remote sensing datasets, such as topography, climate, soil properties, and vegetation maps, which is a key to understanding plant responses to climate change (Durán et al., 2019) In addition, as a critical uncertainty in process models (Rogers et al., 2017), the spatial information on plant traits and biophysical properties inferred from IS could be directly integrated with models to inform and improve predictions (Fer et al., 2018; Shiklomanov et al., 2021), thereby reducing current predictive uncertainties (Dietze et al., 2014).
The high spatial heterogeneity in vegetation composition, structure, and abiotic environments (Section 3) pose a unique challenge to retrieve plant traits using spectroscopy in the Arctic, as compared to other biomes (Thomson et al., 2021; Yang et al., 2021). Traditional radiative transfer model-based retrieval assumes the underlying vegetation layer to be homogeneous (Jacquemoud et al., 2009), which is not met in tundra landscapes. Empirical modeling that builds statistical relationships between field trait observations and remote sensing spectra using machining learning or latent variable techniques is a powerful alternative (Curran et al., 1997; Singh et al., 2015; Z. Wang et al., 2020; Wold et al., 2001). However, to construct an empirical model, a plot-to-pixel connection is required. This requirement can be easily met in forest or managed ecosystems where a single tree can occupy one or multiple image pixels or a vegetation layer is homogenous across relatively large areas. The Arctic poses challenges to plot-to-pixel connections given the high level of species mixing in imagery pixels of > 5 m resolution, which, combined with the remote and meteorologically harsh environment, restricts the collection of quantitative plot observations to develop trait models.
Unoccupied Aircraft Systems (UAS) remote sensing collects spectral data at a very high spatial resolution and has shown great promise to serve as an intermediate data source to connect ground and high-altitude platforms (Thomson et al., 2021). In addition to the high spatial heterogeneity, the common presence of non-vegetated surfaces (e.g., water, soil, rocks, and litter) and their highly variable spectral characteristics (Section 3.5), present additional challenges to the mapping of traits. Typically, non-vegetated surfaces can be excluded over the course of trait model development and application in low-latitude ecosystems (e.g., Wang et al., 2019), but non-vegetated surfaces are highly mixed with vegetation surfaces in the Arctic, which must be accounted for in trait model development. Lastly, the short growing season and harsh environment means that vegetation spectra and traits can change rapidly during the growing season (Section 4.4). Therefore, trait models built from data collected at a certain time of year may only be applicable to a small time window (e.g., < 1 month), as compared to low-latitude ecosystems where vegetation growth peaks can persist for several months. SBG and other time-series spectral platforms (e.g., PACE, CHIME, DESIS, EnMAP) hold great potential to address this issue by facilitating the development of time-series models that capture seasonal trait dynamics.