Deep CNN feature fusion with manifold learning and regression for pixel classification in HSI images

Abstract

ABSTRACTSupervised classification and target recognition of Hyperspectral images (HSI) is a challenging task due to high dimensionality and spectral mixing. Straightforward cognitive computation and target classification lead to high computation cost and low recognition accuracy. Limited availability of training samples makes the recognition process very slow and inaccurate. The main purpose of this work is to improve the classification accuracy for high-dimensional images by the fusion of posterior probability obtained from the two-stage probabilistic framework. The first stage addresses the issue of high dimensionality and the second stage addresses the spectral mixing problem. Both stages provide the prediction probability of pixels in a particular class. In stage-1, we have addressed the imbalance between dimensionality and training samples problem for which we have integrated the deep CNN based spatial and spectral features in combined data-cube form, using ‘off-the-shelf’ CNN models. Subsequently, a graph-based non-linear manifold embedding has performed to extract and fuse the region-wise external information. A probability of prediction has obtained by using LDA classifier. These probabilistic values have denoted as a global probability, as an outcome of stage-1. In the stage-2, the spectral mixing issue was addressed by computing the regional probabilities of class mixing for each pixel. The regional probabilities have calculated by using a regional subspace regression approach. Subsequently, the probabilistic output, obtained from stage-1 and stage-2, has been combined with a linear decision fusion method using regularizers. The experiments have conducted on three real Hyperspectral images, i.e. Indian pines (IP), Pavia University (PU), Salinas Valley (SV) datasets. The probabilistic fusion of stage-1 and stage-2 yields to the maximum overall accuracy of 97.38%, 95.10%, and 99.88% for IP, PU and SV datasets. The over-all accuracies have compared with past methods, and it has found that the proposed framework is providing higher prediction accuracies than previous state-of-art methods.