A deep learning based hybrid framework for semisupervised classification of hyperspectral remote sensing images

Since performance of traditional classification methods is extremely dependent on the number of labeled samples and, to gather ground-truth information of hyperspectral images from the earth’s surface is an expensive and time-consuming process. Semisupervised classification is extensively utilized for hyperspectral images to deal with the issue of restricted training samples by combining the power of labeled and unlabeled data. In this paper, a unique semisupervised classification technique depends on a deep learning based hybrid framework (DL-HF) is described in order to utilize the more information as feasible in order to complete the hyperspectral classification issues. To begin, the proposed semisupervised based method DL-HF uses mainly two arrangements for pre labeling of unlabeled data: the neighboring samples create local arrangement based on neighborhood weighted information, and the most similar training data samples perform global arrangement based on deep learning. Then, to expand the training set, a few unlabeled samples along with superior confidence have been chosen. Finally, using the revised training data, self-arrangement which is based on the self-features developed through deep learning used to extract spectral as well as spatial features and generate a classified map. Evaluation results confirmed that the classification effects of proposed DL-HF algorithm are significantly better in contrast to other competing classification schemes on two benchmark hyperspectral datasets: AVIRIS Indian pines and AVIRIS salina valley dataset, in terms of Overall Accuracy (OA), Average Accuracy (AA) and kappa coefficient (k). The overall classification accuracy achieved is more than 94% which is superior to other related classification methods.