Automated SAR Sea Ice Interpretation

Synthetic aperture radar (SAR) provides an efficient imaging tool for sea ice monitoring, which has found important applications in both scientific and operational activities such as climatic research and ship navigation. Due to the large volume of SAR data acquired, automated computer based interpretation is desired, the goal of which is to segment the image into homogeneous regions and classify each segmented region into the correct ice type. Unfortunately, significant variations exist with respect to the intensity and texture appearance of SAR sea ice due to the numerous imaging parameters and environmental factors, and hence make the task extremely challenging in both the segmentation and classification processes.

The success of the interpretation system is thus largely dependent upon its adaptivity to intensity and texture. On the other hand, models relatively insensitive to intensity and texture are needed for robust descriptions of the ice types. The two issues are manipulated respectively by the two processes of the low level unsupervised segmentation on image pixels and the high level supervised classification on segmented regions. In the latter, features other than the intensity and texture need to be efficiently incorporated, as suggested by the success of human operators in discriminating ice types using additional high level knowledge such as floe shape and existence of fractures. Computations of such features require the low level segmentation to produce regions neither over-segmented nor under-segmented, the balance of which is difficult to achieve due to the complexity of the sea ice scenes and needs to be guided by the high level supervised classification.

This thesis presents a SAR sea ice interpretation system that combines in a bidirectional interactive manner the low level segmentation and high level classification. The segmentation algorithm is a region growing technique and the classification is based on Markov Random Field (MRF). The two processes are integrated under the Bayesian framework, with both aiming at reducing a defined energy. The overall process keeps refining the segmentation and producing semantic class labels at the same time in an iterative manner. Various features can hence be efficiently combined, and successful results are demonstrated.