SUB-PIXEL IMAGE CLASSIFICATION OF HYPER-SPECTRAL DATA FOR VEGETATION AND SOIL MAPPINGIN SEMI-ARID ENVIRONMENT

Abstract

The HyMap hyper-spectral data was used to classify photosynthetic
vegetation (PV), non-photosynthetic vegetation (NPV), and exposed soils in a semiarid
savannah environment of McKinlay, northern Queensland, and Australia. This
study aimed to understandhow effective the sub-pixel classificationapproach applied
on hyper-spectral data to distinguish the vegetation and soil features in semi-arid
environment. In contrast to the per-pixel approach this approach treats the pixel
value as reflectance sum of its composite features, and shows its component
abundance. The most commonly used sub-pixel classification technique was used in
this research, namely Linear Spectral Unmixing (LSU). End members were used as
the input class, and the result was compared with the standard maximum likelihood
classification (MLC) using post-classification comparison method The result of this
study shows that LSU produced a patchy distribution of classes throughout the
image. The brown soil tends to be over-estimated with respect to other classes. PV
features were relatively well-mapped compare to other classes. NPV features have
problem with domination of exposed soil reflectance. This is equivalent to the
previous studies result that background soil dominates the spectral reflectance in
this environment. According to the qualitative accuracy assessment, LSU has
higher accuracy in representing PV and NPV compare to the traditional MLC
classification.