Supervised super-resolution to improve the resolution of hyperspectral images classification maps

Abstract

Hyperspectral imaging is a continuously growing area of remote sensing. Hyperspectral data provide a wide spectral range, coupled with a very high spectral resolution, and are suitable for detection and classification of surfaces and chemical elements in the observed image. The main problem with hyperspectral data for these applications is the (relatively) low spatial resolution, which can vary from a few to tens of meters. In the case of classification purposes, the major problem caused by low spatial resolution is related to mixed pixels, i.e., pixels in the image where more than one land cover class is within the same pixel. In such a case, the pixel cannot be considered as belonging to just one class, and the assignment of the pixel to a single class will inevitably lead to a loss of information, no matter what class is chosen. In this paper, a new supervised technique exploiting the advantages of both probabilistic classifiers and spectral unmixing algorithms is proposed, in order to produce land cover maps of improved spatial resolution. The method is in three steps. In a first step, a coarse classification is performed, based on the probabilistic output of a Support Vector Machine (SVM). Every pixel can be assigned to a class, if the probability value obtained in the classification process is greater than a chosen threshold, or unclassified. In the proposed approach it is assumed that the pixels with a low probabilistic output are mixed pixels and thus their classification is addressed in a second step. In the second step, spectral unmixing is performed on the mixed pixels by considering the preliminary results of the coarse classification step and applying a Fully Constrained Least Squares (FCLS) method to every unlabeled pixel, in order to obtain the abundances fractions of each land cover type. Finally, in a third step, spatial regularization by Simulated Annealing is performed to obtain the resolution improvement. Experiments were carried out on a real hyperspectral data set. The results are good both visually and numerically and show that the proposed method clearly outperforms common hard classification methods when the data contain mixed pixels.