Reconstruction of the full-polarimetric covariance matrix from compact-polarimetric synthetic aperture radar data with convolutional neural networks

Abstract

The focus of this thesis is to find an alternative way to reconstruct a pseudo quadrature polarimetric (quad-pol) covariance matrix from compact polarimetric (compact-pol) data. In the latest years, the compact polarimetry SAR mode was developed and used more and more widely. It provides a good compromise between area covered and information content per pixel [13]. The literature has focused for a long time on quad-pol data in the past. They contain more information compared to compact-pol data. Moreover, several ways to extract useful information from quad-pol SAR images have been developed [8]. Compact-pol data can be considered as a lossy compression from quad-pol data, which has inspired research to find ways to reconstruct the latter format from the former. This allows to apply all the methods and algorithms developed for data analysis of quad-pol data to a reconstructed pseudo quad-pol data. The elaboration of more and more effective deep learning techniques in the last few years has guided us to consider convolutional neural networks (ConvNets) a suitable tool for our problem. ConvNets take advantage of the properties of grid-like topology data [7]. They are able to locate spatial and time local connections. After making assumptions of reflection symmetry for the polarimetric covariance matrix, the reconstruction problem can be formulated as the regression from an image 224x224 with 4 channels, representing the compact-pol covariance matrix, to an image 224x224 with 5 channels, representing the quad-pol covariance matrix. This is the reason why we thought that ConvNets could be a good choice from the available suite of machine learning algorithms. Our results were then compared with previous reconstruction methods, Souyris and Nord's [6, 37], applying the same data set. The methods developed in this thesis showed, on average, slightly worse results than those in the literature. However, we observed that, in same cases, they produced interesting outcomes, for example, a good generalization ability.