Classification of forest composition using polarimetric decomposition in multiple landscapes

The Wishart classification, utilizing polarimetric parameters alpha (α) and entropy (H), was performed on airborne L-band Synthetic Aperture Radar to identify dominant scattering mechanisms within each pixel. The scattering classes were then attributed to structural forms of either excurrent (large, central stem with minimal branching) or decurrent (diffuse branching) trees. To test the classification for a variety of forest structures, three contrasting study areas were chosen; a wooded savanna at the Injune Landscape Collaborative Project in Queensland, Australia, a managed transitional boreal-hardwood forest at the Howland Research Forest in Howland, Maine, and a transitional hardwood-pine forest at the Harvard Forest in Petersham, Massachusetts. Two questions are answered in this study: can the polarimetric parameters α and H characterize structurally similar areas within forests? And can they consistently differentiate these areas across multiple study areas? It is shown that the classification explained nearly 80% of the forest composition at the Injune Collaborative Research Site, 47% at the Howland Research Forest and 40% at the Harvard Forest. Classification accuracy decreases with high levels of H, which is a limiting factor at Howland and Harvard forests where canopy heterogeneity, density and moisture content are higher. When high-H Wishart classes (H. > 0.9) are omitted from analysis, accuracy improves to 83% and 86% for the Injune Collaborative Research Site and Harvard Forest respectively. The success of the Wishart classification at low-H indicates that there is a potential for using polarimetric information to characterize forest composition in particular landscapes; namely those that do not exhibit a high, homogeneous H response.