| Iceberg and ship identification in satellite synthetic aperture radar (SAR) data is an essential part of offering an operational iceberg surveillance program. Identification here refers to detection of ocean SAR targets and then classification of these targets as iceberg, ship, or unknown. Maximizing the detection and minimizing incorrect classification of iceberg and ship targets are required. Because coarser resolution satellite SAR data is at times not as intuitive as satellite optical data for manual human interpreted target classification, this process can be labor intensive, subjective, and error prone. Therefore, it is desired that an automated method for iceberg or ship identification be implemented. The methodology investigated here follows a well known standard in supervised pattern recognition, the maximum likelihood-quadratic discriminant function. The goal here in this thesis is to build class models from known iceberg and ship targets. Each class model is based on features that describe targets such as brightness, texture, and shape. Based on these descriptors as training input into the discriminant functions, future unknown targets can be compared with the class model for best fit. The best fit (or minimum distance) is used to assign class status for these unknown targets. One major consideration when using this type of pattern recognition approach is feature selection. Feature selection is based on the notion that some subset (subspace) of the descriptive metrics will lead to improved classification accuracy when comparing discriminant functions. Sequential forward selection and variants of exhaustive search algorithms are implemented and compared. RADARSAT-1, ENVSIAT AP (HH/HV), and EMISAR SAR iceberg and ship targets are used for algorithm training, feature selection, and performance estimation. |
