Measured and predicted synthetic aperture radar target comparison

The DARPA Image Understanding program publicly released measured and predicted synthetic aperture radar (SAR) targets were compared by means of correlation. The training set consisted of three classes (BMP-2, T-72, and BTR-70) at 17 deg depression and 233 azimuths. The test set consisting of seven different serial-numbered targets at 15 deg depression was tested at 196 azimuths. After segmentation and normalization, each test image was correlated with all the training images to generate correlation and classification statistics. Measured correlation scores were higher and more consistent for same serial number training than variant training. The average in-class (0.837) and between-class (0.734) means for measured correlations were higher than both the average in-class (0.707) and between-class (0.675) means for predicted correlations; however, the corresponding variances for in-class (0.056) and between-class (0.048) predicted correlations were higher than in-class (0.026) and between-class (0.036) measured variances. The measured training data declared the target correctly almost 100% of the time; the T-72 and BTR-70 model-predicted data declared the target correctly 80% of the time. The correlation scores varied approximately sinusoidally with aspect. Correlation plots between a single orientation target and the entire training sets showed that a target was highly correlated at both the correct aspect angle and the correct angle rotated 180 deg