We explore some variants of “Gaussianization” for characterizing the distribution of background pixels in multi-spectral and hyperspectral imagery, and then use this characterization to develop algorithms for target detection. We consider two very different problems – anomalous change detection and gas-phase plume detection – as ways to explore the applicability of Gaussianization for remote sensing image analysis. One variant is an extension of the Gaussianization concept to non-Gaussian reference distributions, and in particular, we show that using the multivariate t as the reference distribution often leads to better target detection performance. Since we are no longer, strictly speaking, Gauss-ianizing, we call the method iterative rotation and remarginalization. In our scheme, the remarginalization is achieved with a parametric transformation function that is built up from a linear basis of (hard or soft) hinge functions, which provide explicitly differentiable and enforcably monotonic remarginalization functions. An efficient knot-pruning strategy enables rapid training of these functions. Also, for remote sensing imagery with many spectral channels, we have found it advantageous to pre-whiten the data with axes aligned to principal components, and then selectively to Gaussianize only the top principal components, treating the lower-variance directions as “already Gaussian.” This provides a computationally faster and empirically more effective Gaussianization for spectral imagery.
|