Performance Evaluation of Classic and Accurate SVD Computation in a Multispectral Image Segmentation Problem

Totally nonnegative (TNN) matrices have wide range applications. Recently a more accurate algorithm for computing singular value decomposition (SVD) was developed, promising to improve current application's precision. In this document, the algorithm is explored and tested in a multispectral image processing application: dust storm detection (segmentation). The multispectral data is posed as TNN matrices, then Bidiagonal Decompositions and Singular Values are computed for feature extraction. When we compared the traditional SVD numerical solution and the high relative accuracy SVD algorithm, we found that the latter shows slight improvement over the traditional approach. For visual assessment, we present the event of March 18, 2008, dust storm in central Mexico, and the visual results match with the numerical results.