Disentangling fractional vegetation cover: Regression-based unmixing of simulated spaceborne imaging spectroscopy data

The next generation of spaceborne imaging spectrometers will enable hyperspectral analysis of vegetation cover across large spatial extents. Spectral unmixing provides a means to assess subpixel vegetation composition in such imagery. Here we implement a regression-based unmixing approach to generate fractional vegetation cover on a regional scale from a simulated Environmental Mapping and Analysis Program (EnMAP) satellite scene derived from Airborne Visible InfraRed Imaging Spectrometer (AVIRIS) imagery acquired over the San Francisco Bay Area, California, USA, an area with a mixture of temperate and Mediterranean climate forests, woodlands and shrublands. A hierarchical classification scheme was implemented that considered fractional cover of vegetation as a whole (vegetation vs non-vegetation), vegetation life forms (woody vs non-woody vegetation; tree vs shrub vs grass), and tree leaf type (needleleaf vs broadleaf). A Gaussian Process Regression (GPR) model was trained using synthetically-mixed training data generated from an endmember library, and mapping accuracy was assessed using an independent validation dataset across four ecoregions. Our approach was able to effectively model landscape patterns at all levels of the class hierarchy. Site-wide map accuracy was highest when mapping generic vegetation fractions (MAE = 3.8%) and expectedly decreased at more complex hierarchy levels, with highest errors observed when separating tree and shrub fractions. Still, fraction estimates of needleleaf trees (MAE = 10.6%), broadleaf trees (MAE = 13.1%) and shrubs (MAE = 15.3%) were mapped with low overall error. Using Landsat imagery led to an average decrease in map accuracy of 1.9% when compared to hyperspectral image analysis and a maximum decrease of 3.5% when separating broadleaf and needleleaf trees across all sites. Further, a single regional model was shown to yield comparable results to multiple local ecoregion-based models, facilitating the analysis of large regions without creating a separate model for each region. Our results highlight the utility of regression-based approaches for quantitative vegetation mapping, which is of particular interest for future spaceborne imaging spectroscopy missions operating across large areas at moderate spatial resolution.