Utility of polarimetric imagery as ancillary data for thematic classification as predicted by using measures of spectral separability

Abstract

1991 STS-43 Shuttle polarimetric imagery was used to calculate a degree of linear polarization (DoLP) image, registered to a 1994 NW Madagascar Landsat 5 scene. The Jeffries-Matusita (J-M) distance and Transformed Divergence (TD) were calculated for every class-pair of a 13-class ISODATA classification, using 6 TM (Thematic Mapper) bands, and TM bands plus DoLP. Separabilities were also calculated using just TM bands 7, 4, and 2 and compared against them plus DoLP. There was a small, consistent increase in spectral separability for all class-pairs with the addition of DoLP. A similar experiment was done using 1999 Landsat 7 imagery covering the area of the 1994 NASA BOREAS experiment. The separabilities were calculated for just a 742 combination and 742 plus DoLP created from the airborne POLDER sensor. The maximum increase in separability provided by DoLP was nearly 5-fold greater than for the Madagascar imagery. The separabilities were compared also for bands 742 plus DoLP with 742 plus band 1 and 742 plus band 5. As with band 1, band 5 improved more class-pair separabilities than DoLP, but the average amount of improvement was more than 4-times greater for DoLP for those class-pairs that saw improvement, versus those that were better separated by band 5. The J-M distance predicts better classification performance for DoLP compared to bands 1 and 5, while the TD suggests slightly better performance for band 5 over DoLP or band 1.