Abstract:
A land-cover classification is needed to deduce surface boundary conditions for a
soil?vegetation?atmosphere transfer (SVAT) scheme that is operated by a
geoecological research unit working in the Andes of southern Ecuador. Landsat
Enhanced Thematic Mapper Plus (ETM + ) data are used to classify distinct
vegetation types in the tropical mountain forest. Besides a hard classification, a
soft classification technique is applied. Dempster?Shafer evidence theory is used
to analyse the quality of the spectral training sites and a modified linear spectral
unmixing technique is selected to produce abundancies of the spectral
endmembers. The hard classification provides very good results, with a Kappa
value of 0.86. The Dempster?Shafer ambiguity underlines the good quality of the
training sites and the probability guided spectral unmixing is chosen for the
determination of plant functional types for the land model. A similar model run
with a spatial distribution of land cover from both the hard and the soft
classification processes clearly points to more realistic model results by using the
land surface based on the probability guided spectral unmixing technique.