Increasing the Detection Accuracy of Bi-Temporal Changes via Speckle Whitening of Single-Look Complex Synthetic Aperture Radar Images

This study employs an unsupervised procedure to spatially decorrelate fully-developed speckle in single-look complex (SLC) synthetic aperture radar (SAR) images. The goal is evaluating the extent to which the spatial correlation of the noise induced by the SAR processor affects the detection accuracy of temporal variations of land-cover between two one-look images of the same landscape acquired on different dates. To simulate the scenario, we have spatially correlated a synthetic map of white complex circular symmetric Gaussian noise by using a two-dimensional separable Hamming window in the Fourier domain. The correlated complex speckle field has been modulated by a noise-free optical view, to simulate an SLC SAR image. Subsequently, we have reduced the correlation of the SLC image through a whitening process and calculated the modulus of the complex image. We have applied various methods of statistical change detection for real-valued SAR data, and compared the accuracy of change maps in the following cases: i) ideally uncorrelated noise, ii) correlated noise, iii) correlated noise that has been decorrelated. The study considers three change detection algorithms, ranging from the basic Log-Ratio operator preceded by despeckling to advanced parametric and nonparametric methods based on Kullback-Leibler distance and mean-shift clustering of bivariate scatterplots of local means. Simulation results demonstrate consistent performance improvements, in terms of both geometric accuracy and reduced number of false alarms.