Estimating fractional vegetation cover from multispectral unmixing modeled with local endmember variability and spatial contextual information

Vegetation fractional cover (fCover) is an important canopy structural variable for understanding the climate-vegetation feedback. Trees and non-tree vegetation may respond differently to climate changes, yet traditional fCover estimation methods focus on quantifying fractional cover for general vegetation. Satellite-based spectral unmixing is more advantageous in this regard as it allows for trees, non-tree vegetation-specific fCover mapping. However, existing multispectral unmixing based fCover estimation studies rarely consider local endmember variability or spatial contextual information. This leads to a lack of spatial detail at locations dominated by similar land covers and a discontinuity among neighboring pixels in the fCover map. To address this, we proposed a Local endmember extraction, Global abundance inversion, and Global fCover mapping with consideration of Spatial contextual information (LGGS) framework to enhance current multispectral (Landsat-8) unmixing based fCover estimation. The proposed LGGS approach was first tested at a forested landscape (Site-A) in Livengood, Alaska then transferred to two study sites (Site-B, Site-C) with various site conditions for demonstrating its generalizability. Our results suggest that LGGS automatically locates high-quality endmembers from Site-A with spectra close to the reference data with median spectral angle mapper values = 0.067, 0.12, and 0.124 for trees, non-tree vegetation, and bare soils respectively. Moreover, incorporating local endmember variability and spatial contextual information improves the overall spatial details and continuity of the mapped fCover, resulting in smaller unmixing errors (mean root mean squared differences = 0.0117-0.0154) and higher structural similarity (mean structural similarity index measure similar to=0.99) to the observed spectra than previous efforts. Overall, our mapped tree fCover resembles the reference map much better (Pearson correlation coefficient: 0.86, root mean squared error: 0.21) than existing satellite products including MODIS vegetation continuous fields, Landsat global forest cover change, and Copernicus global land service products in terms of both numerical range and spatial pattern. The proposed fCover estimation method relies on freely available multispectral imagery therefore can be applied to other vegetated regions for large-scale canopy cover monitoring and eco-related studies.