Using temporal information in an automated classification of summer, marginal ice zone imagery

Synthetic aperture radar (SAR) imagery is difficult to classify under summer melt conditions, even with the human eye. Backscatter instability causes the intensities of the first-year ice, multi-year ice, and open water classes to intermix, thus making an intensity-based classification invalid. The method presented in this paper supplements backscatter information from SAR data with wind and temporally-analyzed temperature records and regional statistics in order to achieve an automated ice/no-ice classification of summer imagery in the marginal ice zone (MIZ). Referring to a database of prior area statistics (ice percentages, ice types, temperatures), an expert system forms conclusions to guide a current classification of the same area. Using parameters set by the expert system, an algorithmic floe extraction procedure divides the image into two classes (one assumed to contain floes and the other assumed not to contain floes) and then subdivides those two classes into ice and water. Classification results are then compared to the expected values derived from temporal adjustments of prior ice percentages. Unacceptable differences are called to the attention of the user for further inspection and possible manual correction. The study area for testing is the Beaufort Sea, with data taken from the ERS-1 SAR. Results show that temporally-accumulated data can be used to provide a basis for an automated classification of MIZ imagery under summer melt conditions.