Imaging spectrometry-derived estimates of regional ecosystem composition for the Sierra Nevada, California
Introduction
Our understanding of how terrestrial ecosystems are changing, and will continue to change as a result of ongoing changes in climate, increasing atmospheric CO2 concentrations, and land-use, relies heavily on terrestrial ecosystem and biosphere models. These models incorporate the impact of these different environmental forcings on terrestrial vegetation dynamics, biogeochemistry, and water and energy fluxes (Fisher et al., 2014).
Traditionally, the composition of the plant canopy within terrestrial ecosystem and biosphere models has been represented in a coarse-grained manner, with each of the earth's major biomes being represented as a single, homogeneous Plant Functional Type (PFT) (e.g. C3 grasses, needle-leaved evergreen trees, or cold- or drought-deciduous trees) (Huntzinger et al., 2012; Fisher et al., 2014). However, results of recent terrestrial biosphere modeling studies (e.g. Levine et al., 2016; Sakschewski et al., 2016) have highlighted the importance of fine-scale diversity and heterogeneity in canopy composition for predictions regarding the responses and resilience of terrestrial ecosystems to climatological perturbations. In these models, the plant canopy within each climatological grid cell is compositionally heterogeneous at the scale of individual plants within the canopy, i.e. at spatial scales of meters (see Fisher et al., 2014 for a recent review).
The incorporation of fine-scale ecosystem heterogeneity in terrestrial biosphere models has, in large part, been motivated by the desire to incorporate a greater diversity of PFTs into model simulations because diversity is an important source of ecosystem resilience (see Moorcroft, 2006 and references therein). Building upon the development of regional and global-scale databases of plant attributes, such as TRY (Kattge et al., 2011), GLOPNET (Wright et al., 2004) and the database of Atkin et al. (2015), plant trait databases are increasingly being used to define the characteristics of plants within the terrestrial biosphere models they are representing (e.g. Verheijen et al., 2013; Fyllas et al., 2014; Sakschewski et al., 2016). Examples of important composition-related parameters within terrestrial biosphere models are given in Table A1: they include leaf photosynthetic properties that govern rates of carbon fixation and transpiration, seasonal patterns of leaf phenology, patterns of plant carbon allocation to leaves, stems and roots, rates of tissue turnover, and resulting rates of plant height and diameter growth and rates of stem mortality.
Successful implementation of this new generation of terrestrial biosphere models in regional and global scale modeling studies requires more highly-resolved information about spatial variation in the composition of plant canopies than the information provided by existing global classifications of vegetation types. The ability to more accurately specify plant canopy composition is important because it is used to specify plant physiological and ecological properties that determine the biophysical and biogeochemical functioning of ecosystems being simulated.
Remote sensing measurements offer a promising way to provide spatially-resolved, spatially-extensive information about the fine-scale variation in the composition and structure of plant canopies that can be used by the new generation of terrestrial biosphere models to improve their predictions of vegetation dynamics and the associated exchanges of carbon, water and energy between land and the atmosphere. Metrics of fine-scale ecosystem structure, such as forest canopy height (Dubayah and Drake, 2000; Drake et al., 2002; Lefsky et al., 2005) and aboveground biomass (Saatchi et al., 2007; Saatchi et al., 2011; Baccini et al., 2012), have been derived using lidar and radar active remote sensing technologies, and analyses have shown that these measurements can be successfully used to initialize and improve terrestrial ecosystem and biosphere simulations (Hurtt et al., 2004; Hurtt, 2010; Antonarakis et al., 2011). Regarding metrics of fine-scale ecosystem composition, Antonarakis et al. (2011) showed that, in addition to measures of ecosystem structure, further improvements to terrestrial ecosystem simulations required spatially-resolved, spatially-extensive information about the fine-scale or sub-pixel variation in the composition of the plant canopy. This is because, as noted earlier, plant composition determines important physiological and ecological properties of the ecosystem, which, in turn, governs its biophysical and biogeochemical functioning.
Currently, estimates of regional and global-scale terrestrial canopy composition used in terrestrial ecosystem and biosphere model simulations are derived from multi-spectral remote sensing estimates of broadly-defined land cover categories or global biome types. For example, the MODIS global land cover product (Friedl, 2010) distinguishes 8 global biome types, the Global Land Cover Product (GLOBCOVER; Bontemps et al., 2011) distinguishes 18 global vegetation types and the NLCD (National Land Cover Database) Land Cover product distinguishes 9 vegetation types. Fig. 1a shows the NLCD (Homer et al., 2015) classification of the region investigated in this study, in which the vegetation in each pixel is assumed to be homogenous and comprised of one of five vegetation types; deciduous, evergreen, mixed forest, shrubs, and grasses. The impact of land cover type uncertainty on terrestrial ecosystem and biosphere model predictions was investigated by Quaife et al. (2008), who showed that incorrect assignment of land cover classes to PFTs, the difficulties in differentiating vegetation types, and information loss at coarse resolutions resulted in significant differences in predictions of terrestrial carbon fluxes.
Over the past two decades, studies have shown that the higher spectral resolution of imaging spectrometry sensors can yield more taxonomically- and functionally-resolved estimates of spatial variation in canopy composition (Goodenough et al., 2003; Kokaly et al., 2003; Asner et al., 2008; Féret and Asner, 2013; Ferreira et al., 2016; Clark et al., 2018). For example, Martin et al. (1998) classified eleven forest cover types in Harvard Forest, Massachusetts, Kokaly et al. (2003) distinguished nine forest types within an area of Yellowstone National Park, van Aardt and Wynne (2007) distinguished three pine species in a Virginia forest, and Clark et al. (2018) recently classified twenty-one forest types in the Californian Bay Area.
One of the most commonly used methods for estimating vegetation composition from imaging spectrometry measurements is Spectral Mixture Analysis (SMA; Adams et al., 1986), which estimates the fractional contribution of different reference spectral signatures to the reflectance signature of each pixel (Keshava and Mustard, 2002). Compared to other classification methods, such as Random Forests, which assign a single land cover type per pixel, the fractional cover estimates obtained from SMA aligns well with the need to specify fine-scale spatial variation in composition and structure of plant canopies in modern terrestrial ecosystem and biosphere models. See Fassnacht et al. (2016) for a recent review of classification methods applied to forested ecosystems. Ideally, SMA uses the fewest possible endmembers (i.e. ‘pure’ spectra corresponding to each vegetation or land cover type) to characterize the fractional composition within each pixel (Sabol, 1992; Adams and Gillespie, 2006), and requires some prior knowledge of the ecosystem. Multiple Endmember Spectral Mixture Analysis (MESMA; Roberts et al., 1998) builds on traditional SMA by allowing endmembers to vary in type and number on a per pixel basis with combinations of endmembers with the best fit (lowest RMSE) being chosen for each pixel. MESMA also accounts for variation in brightness between endmembers by including a shade endmember, which can be either an image-based endmember or photometric (i.e. spectrally flat, or near-zero reflectance). MESMA has been used for a number of vegetation mapping studies: Roberts et al. (1998) and Hamada et al. (2011) mapped vegetation in arid landscapes of southern California; Li et al. (2005) mapped coastal marsh vegetation in California; and Youngentob et al. (2011) mapped two Eucalyptus sub-genera in Australia.
As noted by Clark et al. (2018), however, until recently, the number of studies examining the ability of medium-resolution spaceborne hyperspectral sensors to estimate ecosystem composition has been limited. Three recent studies that have addressed this issue are Roth et al. (2015), who used MESMA to map the dominant plant species in five different ecosystems across North America, Clark (2017) who used MESMA and machine learning to map ten vegetation cover types in the Californian Bay Area, and Clark et al. (2018) who used machine learning to classify twenty-one forest types in the Californian Bay Area. However, in all three of these analyses MESMA was used to identify and validate the dominant cover type within each pixel rather than producing estimates of the fractional composition of vegetation within each pixel.
Antonarakis et al. (2014) showed how high spatial resolution (10 m) imaging spectrometry-derived estimates of the sub-pixel level fractional PFT composition of an eastern mixed-temperate forest canopy could be combined with lidar-derived estimates of forest canopy structure to provide a spatially-resolved estimate of above-ground plant composition and structure across a ~4 km2 area of the Harvard Forest Long-Term Ecological Research (LTER) site in the Northeastern United States. Their analysis showed that incorporating this remote sensing-derived estimate of above-ground ecosystem state into an individual-based terrestrial biosphere model simulations yielded marked improvements in the accuracy of the model's predictions of the ecosystem's carbon fluxes compared to those obtained from a conventional “potential vegetation” simulation (i.e. a simulation in which canopy composition is estimated by performing a long-term, multi-century simulation until the vegetation within the model comes into equilibrium with the climatological forcing data). Moreover, the improvements were similar in magnitude to the improvements obtained from simulations initialized with ground-based inventory measurements of canopy composition and structure (Antonarakis et al., 2014).
In this analysis, we use imaging spectrometry measurements from the recent National Aeronautics and Space Administration (NASA) Hyperspectral Infrared Imager (HyspIRI) Preparatory Campaign (Hochberg et al., 2015) to estimate sub-pixel level plant functional type composition across a ~710 km2 area in the Southern Sierra Mountains of California. The purpose of the HyspIRI preparatory campaign is to provide multi-temporal imaging spectrometry data over large areas of California at spatial resolutions of ~18 m and 30 m, i.e. at a spatial resolution approaching the characteristics of anticipated global imaging spectrometry missions, such as HyspIRI (Lee et al., 2015) the Environmental Mapping and Analysis Program (EnMAP, Guanter et al., 2016) and the Hyper-spectral Imager SUIte (HISUI, Matsunaga et al., 2016), and NASA's planned Surface Biology and Geology mission (NASEM, 2018).
The purpose of this study is to evaluate the ability of such instruments to provide reliable, spatially-extensive estimates of fractional PFT composition suitable for constraining individual-based terrestrial ecosystem and biosphere model simulations of large-scale ecosystem dynamics and functioning. As in Antonarakis et al. (2014), key characteristics of the methodology employed here compared to earlier studies are: (i) the direct alignment between definitions of the PFTs whose abundances are being estimated by remote sensing and the PFTs defined within the terrestrial biosphere model; (ii) use of the sub-pixel PFT fractional abundances to define the fine (meter-scale) sub-grid scale variation in the composition of the plant canopy within the individual-based ecosystem model. The analysis presented here moves beyond that of Antonarakis et al. (2014) in three important respects with regard to assessing the relevance of a global imaging spectrometry mission for constraining regional-to-global scale terrestrial biosphere model simulations. First, we assess the ability of this methodology to work with imaging spectrometry measurements collected at spatial resolutions of planned global missions and to be successfully applied at large spatial scales and over heterogeneous landscapes. Second, we investigate the ability to successfully estimate ecosystem composition in regions of complex terrain that have significant topographic heterogeneity. Third, we evaluate the ability to successfully estimate plant canopy composition across a variety of ecosystem types, including: mixed tree-grass savannahs, deciduous woodlands, and coniferous forests.
Section snippets
Study area and data
The study area is a 710 km2 area (~67.5 km × 10.5 km) in the Sierra Nevada Mountains northwest of Fresno, California (Fig. 1a–c) (National Elevation Dataset, 2009). It spans an elevational gradient ranging from 125 to 3175 m above sea level: the western third lies in the foothills of the Sierra Nevada and is dominated by oak savanna ecosystems; the central and eastern portions are higher elevation montane to subalpine ecosystems, moving eastwards from oak-savannas to pine/oak forest, through
Conventional MESMA classification
The PFT composition of the CZO Study Area estimated by conventional MESMA is shown in Fig. 2c. The analysis indicates that the western, lower elevation (<875 m) portion of the study area is comprised of Mediterranean savannah ecosystems dominated by grasses interspersed with oaks, shrubs and western hardwoods. At mid-elevation sites (875–1700 m), located in the center portion of the study area, canopy composition changes to a mixture of Shrub-dominated, Oak-dominated, and Cedar/Fir dominated
Discussion
Knowledge regarding the large-scale spatial distribution of canopy composition across heterogeneous landscapes is critical for improving terrestrial biosphere model predictions of carbon, water and energy fluxes, and how these will be affected by changes in climate, atmospheric CO2, and land-use. Existing national and international large-scale ecosystem canopy composition products derived from multi-spectral imagery such as that shown in Fig. 1a, do not have the fine-scale diversity and
Conclusions
This study evaluated the ability of imaging spectrometry measurements, collected over a functionally and topographically heterogeneous ~710 km2 area in the Southern Sierra Mountains of California, to estimate sub-pixel fractions of seven PFTs; Grass/NPV, Shrub, Oak, Western Hardwood, Western Pine, Cedar/Fir, and High-elevation Pine for use in terrestrial biosphere and ecosystem model simulations. Applying the conventional MESMA approach, in which all PFTs are evaluated for every pixel, to the
Acknowledgements
The authors would like to acknowledge the NASA HyspIRI Preparatory Activity NNH11ZDA001N-HYSPIRI grant “Linking Terrestrial Biosphere Models with Remote Sensing Measurements of Ecosystem Composition, Structure, and Function”. We thank Michael Goulden and Anne Kelly at the University of California, Irvine for the forest inventory plot data at the Southern Sierra CZO flux tower sites. We thank Susan Ustin and Dr. Huesca-Martinez at the University of California, Davis and the CZO/NEON projects for
References (84)
- et al.
Spectranomics: emerging science and conservation opportunities at the interface of biodiversity and remote sensing
Global Ecology and Conservation
(2016) - et al.
Invasive species detection in Hawaiian rainforests using airborne imaging spectroscopy and lidar
Remote Sens. Environ.
(2008) - et al.
Quantifying forest canopy traits: imaging spectroscopy versus field survey
Remote Sens. Environ.
(2015) Comparison of simulated hyperspectral HyspIRI and multispectral Landsat 8 and Sentinel-2 imagery for multi-seasonal, regional land-cover mapping
Remote Sens. Environ.
(2017)- et al.
Mapping of forest alliances with simulated multi-seasonal hyperspectral satellite imagery
Remote Sens. Environ.
(2018) - et al.
Endmember selection for multiple endmember spectral mixture analysis using endmember average RMSE
Remote Sens. Environ.
(2003) - et al.
Sensitivity of large-footprint lidar to canopy structure and biomass in a neotropical rainforest
Remote Sens. Environ.
(2002) - et al.
Review of studies on tree species classification from remotely sensed data
Remote Sens. Environ.
(2016) - et al.
Mapping tree species in tropical seasonal semi-deciduous forests with hyperspectral and multispectral data
Remote Sens. Environ.
(2016) - et al.
Leaf optical properties with explicit description of its biochemical composition: direct and inverse problems
Remote Sens. Environ.
(1996)
Imaging spectroscopy and the airborne visible/infrared imaging spectrometer (AVIRIS)
Remote Sens. Environ.
Estimating life-form cover fractions in California sage scrub communities using multispectral remote sensing
Remote Sens. Environ.
Special issue on the hyperspectral infrared imager (HyspIRI): emerging science in terrestrial and aquatic ecology, radiation balance and hazards
Remote Sens. Environ.
Completion of the 2011 National Land Cover Database for the conterminous United States-representing a decade of land cover change information
Photogramm. Eng. Remote. Sens.
Mapping vegetation in Yellowstone National Park using spectral feature analysis of AVIRIS data
Remote Sens. Environ.
An introduction to the NASA Hyperspectral InfraRed Imager (HyspIRI) mission and preparatory activities
Remote Sens. Environ.
Mapping vegetation in an urban area with stratified classification and multiple endmember spectral mixture analysis
Remote Sens. Environ.
Determining forest species composition using high spectral resolution remote sensing data
Remote Sens. Environ.
How close are we to a predictive science of the biosphere?
Trends Ecol. Evol.
Mapping chaparral in the Santa Monica Mountains using multiple endmember spectral mixture models
Remote Sens. Environ.
Comparing endmember selection techniques for accurate mapping of plant species and land cover using imaging spectrometer data
Remote Sens. Environ.
Differentiating plant species within and across diverse ecosystems with imaging spectroscopy
Remote Sens. Environ.
Atmospheric correction for global mapping spectroscopy: ATREM advances for the HyspIRI preparatory campaign
Remote Sens. Environ.
Canopy foliar nitrogen retrieved from airborne hyperspectral imagery by correcting for canopy structure effects
Int. J. Appl. Earth Obs. Geoinf.
Mapping two Eucalyptus subgenera using multiple endmember spectral mixture analysis and continuum-removed imaging spectrometry data
Remote Sens. Environ.
Chapter 4 spectral mixture analysis
Spectral mixture modeling: a new analysis of rock and soil types at the Viking Lander 1 site
J. Geophys. Res. Solid Earth
Using lidar and radar measurements to constrain predictions of forest ecosystem structure and function
Ecol. Appl.
Imaging spectroscopy- and lidar-derived estimates of canopy composition and structure to improve predictions of forest carbon fluxes and ecosystem dynamics
Geophys. Res. Lett.
Impacts of the 2012–2015 Californian Drought on Carbon, Water and Energy Fluxes in Californian Sierras: Results From an Imaging Spectrometry-Constrained Terrestrial Biosphere Model
Radiometric estimates of nitrogen status of leaves and canopies
GLOBCOVER 2009-Products Description and Validation Report
Comparison of tree basal area and canopy cover in habitat models: subalpine Forest
J. Wildl. Manag.
Existing Vegetation, [ESRI Geodatabase]
Existing Vegetation, [ESRI Geodatabase]
Existing Vegetation, [ESRI Geodatabase]
Existing Vegetation, [ESRI Geodatabase]
Cited by (18)
Continental-scale hyperspectral tree species classification in the United States National Ecological Observatory Network
2022, Remote Sensing of EnvironmentCitation Excerpt :The dimensionality reduction algorithm used in this study identified groups of adjacent bands in relatively discrete spectral regions that overlap with spectral regions used in multispectral satellites, supporting the idea that multispectral satellite sensors can access a large amount of spectral information for species classification (Laurin et al., 2016). Hyperspectral satellite data is still limited to few prototype datasets with relatively low spatial resolution (Loizzo et al., 2018; Diaz et al., 2018; Bogan et al., 2019), compared to multispectral satellites with sub-meter resolution (e.g. WorldView3). Our results show that most of the information required for species classification across NEON sites overlap with WorldView3 satellite multispectral bands supporting that species identification at the tree and plot level with satellite data is feasible (Immitzer et al., 2012, Hartling et al., 2019, Ferreira et al., 2019).
Multi-season unmixing of vegetation class fractions across diverse Californian ecoregions using simulated spaceborne imaging spectroscopy data
2021, Remote Sensing of EnvironmentCitation Excerpt :Regarding the mapping of vegetation class fractions, temporal information of multi-season or time series data has been successfully integrated into unmixing procedures (Hansen and DeFries, 2004; Schug et al., 2020; Somers and Asner, 2013b). Considering the HyspIRI Airborne Campaign data, both single- and multi-season imaging spectroscopy data have been exploited through unmixing (Bogan et al., 2019; Clark, 2017; Cooper et al., 2020a; Wetherley et al., 2018). However, none of these HyspIRI-based studies have exploited multi-season imaging spectroscopy data from the multiple, ecologically-distinct flight boxes for unmixing of vegetation class fractions, nor have any drawn comparisons to unmixing based on temporally-dense multispectral data archives.
Combining simulated hyperspectral EnMAP and Landsat time series for forest aboveground biomass mapping
2021, International Journal of Applied Earth Observation and GeoinformationScalable mapping and monitoring of Mediterranean-climate oak landscapes with temporal mixture models
2020, Remote Sensing of EnvironmentCitation Excerpt :While useful for mapping changes associated with specific events, these methods do not fully exploit the dense image time series that are now available. Other methods focus on aerial orthophoto, imaging spectroscopy, and/or LiDAR surveys (e.g. Asner et al., 2016; Bogan et al., 2019; Fricker et al., 2019; Miraglio et al., 2020; Navarro et al., 2019; Swatantran et al., 2011). These studies have achieved impressive results, but remain spatially and temporally limited until data coverage expands considerably – and even then will not allow for retrospective analysis.
Disentangling fractional vegetation cover: Regression-based unmixing of simulated spaceborne imaging spectroscopy data
2020, Remote Sensing of Environment