Evaluation of polarimetric configurations for glacier classification
Anthony P. Doulgeris(1), Stian N. Anfinsen(1), Yngvar Larsen(2), Kirsty Langley(3,1) and Torbjørn Eltoft(1)
(1) University of Tromsø, The Auroral Observatory, 9037 Tromsø, Norway
(2) Norut, Box 6434, 9294 Tromsø, Norway
(3) Norwegian Polar Institute, Polar Environment Centre, 9296 Tromsø, Norway
Monitoring arctic glaciers is currently of high global importance for which satellite-borne Synthetic Aperture Radar (SAR) systems are well suited because of their cloud penetrating ability and broad coverage. The scale of the monitoring problem means that reliable automated image classification schemes need to be developed. Fully polarimetric SAR imagery contains the most scattering information leading to the best classification results, however fully polarimetric satellite systems are few and still considered experimental and therefore this type of data is not readily available. For practical reasons too, the wider coverage of single or dual polarisation modes are often chosen for monitoring purposes. We investigate and compare the glacial classification ability of various polarimetric configurations, that correspond to commonly acquired dual and single channel SAR products.
We have an ALOS/PALSAR quad-pol scene over Holtedahlfonna glacier, Svalbard, from December 2007. We also have linear ground truth data derived from field based ground penetrating radar profiles along the glacier. The ground truth consists of several ice classification zones of interest to glaciologists, such as the "firn" accumulation zone which is important for mass balance calculations. Fully polarimetric classifications are commonly performed with the Gaussian based Wishart clustering algorithm. The Wishart clustering algorithm is often initialised with a physical decomposition scheme, like the H/A/alpha or the Freeman-Durden decomposition. However, we have found that a simple intensity based initialisation works well and is applicable even for reduced polarimetric configurations where the physical decomposition interpretation may break down. In this study, we use the more advanced non-Gaussian K-Wishart clustering algorithm, that additionally accounts for potential textural differences in the classes. This is a stochastic EM type algorithm, which performs model parameter estimation and Bayesian classification according to an iterative processing scheme. It differs from the standard Wishart because the underlying model for the scattered signal is based on the multivariate K distribution rather than the Gaussian. The fully polarimetric standard Wishart and K-Wishart performances are compared with regard to the ground truth profiles. Subsequently, combinations of channels are classified and compared to the quad-pol result to investigate the level of classification ability with particular polarisation channel sub-sets. Knowledge of quad, dual and single-pol relative performance may be useful for planning monitoring schedules, associating classification confidence levels to the glacier monitoring data, and as a mechanism for integrating historical SAR data-sets with the newer quad-pol systems.