Classification of Remotely Sensed Images Using Independent Component Analysis and Spatial Consistency
We apply independent component analysis (ICA) to learn efficient color representation of remotely sensed images. Among the three basis functions obtained from RGB color space, two are in an opposing-color model by which the responses of R, G and B cones are combined in opposing fashions. This is coincident with the idea of contrasting reflected in many color systems. The interesting point is that there is no summation component that corresponds to illumination in other transforms. Spectral independent components are then used to cluster pixels. After pixel-based classification, we segment an image on the basis of regions by spatial consistency. Experimental results show that this method considerably improves the classification performance of multispectral remotely sensed images.