Use of SAR image texture in terrain classification

Classification is a common first step in the use of SAR data. Intensity of a pixel is generally used as a feature vector. This is complicated by coherent fading that yields multiplicative noise. Consequently, the first statistical moment of intensity (over some local window) is often used as a feature vector instead. In some cases this leads to unacceptably high rates of misclassification. The 2nd statistical moment also can be used to distinguish categories but is dependent on the composite effects of the sensor (N of looks), the mean backscatter (via multiplicative noise) and the true spatial variance in average backscatter relative to SAR resolution. Thus, using variance measures as feature vectors can lead to increased classification accuracy. However, such measures ignore the observation that the variance for many terrain categories is not stationary and indeed may not be isotropic. Further improvement in classification can be realized by quantifying the translational variance in backscatter using scale-dependent geostatistical semi-variance and lacunarity that spatial structure of image intensity. Simulated SAR data are used to understand the effects of system parameters (such as number of looks and spatial resolution) and target conditions (such as probability of occurrence and stationarity) on geostatistical measures of texture. ERS-1 and JERS-1 SAR data demonstrate the use of these techniques in terrain characterization. These statistics also give measures of heterogeneity of interest to ecologists.