Abstract

Humid tropical forest types have low spectral separability in Landsat TM data due to highly textured reflectance patterns at the 30m spatial resolution. Two methods of reducing local spectral variation, low-pass spatial filtering and image segmentation, were examined for supervised classification of 10 forest types in TM data of Peruvian Amazonia. The number of forest classes identified at over 90% accuracy increased from one in raw imagery to three in filtered imagery, and six in segmented imagery. The ability to derive less generalised tropical forest classes may allow greater use of classified imagery in ecology and conservation planning.