Radiance thresholds and texture parameters for Antarctic surface classification

A sequence of AVHRR images is investigated in order to obtain information about sea ice distribution in the Ross Sea. A supervised classifier, based on decision rules in AVHRR Near IR (2) and thermal (3 and 4) channels, is tuned to the given data set to extract pixels belonging to classes: (a) sea; (b) multiyear ice (ice shelves and bergs); (c) first season ice types (floes and pack); and (d) clouds (thin/thick clouds, water clouds and ice clouds). Pixel-based classification is affected by underestimation of ice clouds due to interference of multiyear ice and pack. To reduce ice/cloud misclassification, an alternative classification approach exploiting textural information is attempted. Firstly, image segmentation is performed by means of a three-stage procedure consisting of: (a) contour detection and segment extraction; (b) region growing; and (c) clustering of segment-based statistics by means of a neural network model. Secondly, three textural features (mean, recursivity and contrast) are extracted as within-segment statistics. Results show that these segment-based textural parameters cannot be clustered in the measurement space in line with their information classes, i.e., they cannot be effectively employed in the classification scheme to reduce the amount of interference between multiyear ice/first season ice types and ice clouds.