Unsupervised classification of Arctic sea ice using neural network

Analysis of multichannel passive microwave data indicates that the emissivities of sea ice tend to be uniform within each ice regime but varies considerably from one regime to another. The regimes include the marginal ice zone, where new and young ice dominates, the seasonal ice region, where first year ice with different snow cover and surface characteristics are observed, and the perennial ice region, where deformed or undeformed second and older ice types are located. In this investigation, these ice regimes have been studied over an entire fall/winter period and their persistence, duration, and variability are evaluated. The procedure involves the use of an unsupervised cluster analysis program, that isolates in 7-dimensional space (representing data from the 7 SSM/I channels), the radiometrically distinct ice surfaces. A neural network is then utilized to improve the classification using the cluster analysis results to train the system with a backpropagation model. The time evolution of each cluster is analyzed and the neural network is further enhanced to filter noise in the data that may be associated with atmospheric and other external effects. The areal extent of sea ice in each regime has been quantified and results show that the seasonal region undergo tremendous changes but the other regions were basically constant with time