An iterative approach to multisensor sea ice classification

Abstract

Characterizing the variability in sea ice in the polar regions is fundamental to an understanding of global climate and the geophysical processes governing climate changes. Sea ice can be grouped into a number of general classes with different characteristics. Multisensor data from NSCAT, ERS-2, and SSM/I are reconstructed into enhanced resolution imagery for use in ice-type classification. The resulting twelve-dimensional data set is linearly transformed through principal component analysis to reduce data dimensionality and noise levels. An iterative statistical data segmentation algorithm is developed using maximum likelihood (ML) and maximum a posteriori (MAP) techniques. For a given ice type, the conditional probability distributions of observed vectors are assumed to be Gaussian. The cluster centroids, covariance matrices, and a priori distributions are estimated from the classification of a previous temporal image set. An initial classification is produced using centroid training data and a weighted nearest-neighbor classifier. Though validation is limited, the algorithm results in an ice classification that is judged to be superior to a conventional k-means approach.