Mapping Mediterranean natural vegetation species using heterogeneous endmember analysis of hyperspectral imagery
Summary
Vegetation is a crucial component of system Earth as it plays an essential role in the water balance and the carbon dioxide cycle. Knowledge on the distribution and dynamics of vegetation is therefore very important. Imaging spectroscopy is a promising tool to monitor and understand complex vegetation patterns and dynamics. To map vegetation species is one of the most challenging objectives, as reflectance spectra of vegetation are very similar. Conventionally, areas are classified per species which assumes homogeneous compositions. However, natural vegetation compositions are often mixed – i.e. heterogeneous. Mapping species in natural environments with a homogeneous approach is therefore difficult.
This research proposes a new method of species mapping: spectral unmixing on the basis of heterogeneous endmembers as opposed to homogeneous endmembers. Next to that, two upcoming analysis methods are compared to their conventional version: object-based compared to pixel-based image analysis and image analysis based on original compared to continuum removed reflectance spectra.
Six Mediterranean vegetation species were mapped in the Peyne catchment, southern France. These species typically occur in dense forests of mixed composition. Three different approaches were applied to investigate the effects: 1) linear spectral unmixing on the basis of heterogeneous and homogeneous endmembers. 2) Image analysis with an object-based approach and a pixel-based approach. 3) Image analysis with continuum removed reflectance spectra and original reflectance spectra. Lastly, the accuracy was assessed and correlations were checked to determine the differences in reliability.
It is concluded that linear spectral unmixing on the basis of heterogeneous endmembers produces substantially better results than linear spectral unmixing on the basis of homogeneous endmembers. Average root mean square error for all species for the heterogeneous approach is 23 compared to 32 for the homogeneous approaches. No differences in accuracy were found for the object-based compared to the pixel-based approach, and for the image analysis based on original reflectance spectra compared to continuum removed reflectance spectra. A substantial negative correlation between average root mean square error per plot and the level of heterogeneity was found, indicating bias in favour of heterogeneous pixels.