The discrimination of scenes by principal components analysis of multi-spectral imagery

The potential limitations for reduction of Landsat 5 TM data by principal components analysis (PCA) were studied, using images of two different landscapes at Paraiba, northeast Brazil (coastal/urban and inland/agricultured scenes). Standardized counts were considered for six bands in the solar spectrum. Basic correlation matrices needed for principal axes definition were built by sampling of 1500-3000 pixels. It was found that the use of some simple concepts of factor analysis (Varimax rotation and 'correlation circles') allows the detection of redundant channels in different types of scene. Up to three original variables appeared as non-redundant, and only two principal components were enough to identify main field patterns. When using two principal components and simple clustering methods (e.g., supervised minimal euclidean distance), various types of vegetated landscapes could be discriminated, with quality similar to that obtained by using all six wavebands. These results show that factor analysis in principal components clearly allows to define a minimal set of non-redundant channels (or of principal components), proper for discrimination studies, such that no further data reduction would proceed. Factor analysis procedures allied with simple computational programs offer a powerful means of scene discrimination.