A study of methods for classification of coniferous and broad-leaved trees using full polarimetric SAR data

Abstract

It is important to examine forest trees in each species for understanding forest status and managing them. SAR (Synthetic Aperture Radar) has been used to forest survey. SAR can obtain structural information of trees, since its wavelength is long enough to penetrate into the tree canopy and the microwave reflects at stems and other large branches. This shows possibility of SAR to classify forest vegetation into each species based on the tree structural information. In this paper we tried to classify forest trees into coniferous and broad leaved ones using a polarimetric SAR image. First, we derived a coherency matrix for each pixel from the scattering matrix of the SAR data. The image was then decomposed into powers of four scattering models, based on the elements of the coherency matrix. The images of four powers were classified into 5 classes (coniferous, broad leaved, water, field, and urban area). To exclude mismatching area between the SAR image and ground validation data, we built a mask which excludes slopes that are on the opposite side of the mountain ridge as seen from the satellite. In addition to the four-component decomposition, we applied the three-component decomposition method and H, A, α decomposition method for the classification. As a result, the values for overall accuracy were 79%, 65% and 47% and the Kappa coefficient values were 0.51, 0.3 and 0.12 for the four-component decomposition, three component decomposition and H, A,α decomposition methods, respectively.