Noise analysis using Tucker decomposition and PCA on spectral images

Abstract

Given the improvement of Remote Sensing (RS) sensors, it has been possible to increase spatial and spectral resolution on many of them. Nevertheless, the amount of data to represent and post-process has become highly prohibitive. Therefore, the need to be able to process such huge data sets, and one of the possible ways to deal with problems is the use of compression methods, however, data loss happen if the need of data size reduction is a must. RS spectral imagery contain high quantities of redundant information along the spectral domain, thus, making possible to use compression methods effectively as for example, tensor decomposition algorithms. In Tucker decomposition (TKD), an interesting and strange phenomenon happens when spatial domain is maintained and spectral domain is reduced, as a preprocessing step of a semantic segmentation task. Under these conditions, it is possible to observe an improvement on Pixel Accuracy (PA) metric when it is compared with the same uncompressed spectral image. Therefore, this work presents a study on how noise affects the Tucker Decomposition compared with Principal Component Analysis (PCA) and its impact in semantic segmentation.