High Range Resolution (HRR) radar is one sensor of interest to the military. This sensor collects data which is a range profile of an aircraft, the result of electromagnetic scattering from the target as a function of the line of sight range. Conventional means of developing an ATR system fail to give adequate results and learning techniques must be used. The primary objective of this research was to develop a workable, robust classification methodology using machine learning and data mining techniques. Specifically the approach should: generate features for classification, determine important features, generate multiple classifiers, determine a method of fusing classifiers for robustness, and be computationally appropriate. Rough Set Theory (RST)guarantees that all possible classifiers using a labeled training set will be generated! There is no equivalent statement for statistical pattern recognition. However, generating all classifiers has been shown to be a NP-hard problem. Therefore, this research had a secondary objective, to find ways to overcome this problem for real world size problems.
To meet these objectives first the data was partitioned using a block and an interleave scheme. This provides classifiers that focus on local and global features. Following partitioning wavelets were used to enrich the feature space for the classification procedure. A subset of the most important classification features was selected using information entropy. This calculation was also used to determine the division point for binary labeling of each range bin. A polynomial time complexity RST method was developed to compute minimal classifiers for each partition. A fusion formula was developed which fused the classifications for all partitions. This was the HRR rough set classifier.
During this research it was found that which wavelet family was used made no statistical difference. It was also determined that an iterated wavelet transform would, in essence, result in a new wavelet tailored to the given problem. This new wavelet produces a 12 percentage point improvement in classifier performance. The rough set classifier produces 93% probability of correct classification and almost 100% probability of declaration. These results are five percentage points better than the most popular method, the quadratic classifier.