A Deep Learning-Based Intelligent Decision Support System for Hyperspectral Image Classification Using Manifold Batch Structure in Internet of Things (IoT)

In the modern era of Internet of Things (IoT), deep-learning-based systems are exhibited encouraging performance in hyperspectral image (HSI) classification, because of their capability in extracting key deep features from available images. However, this deep learning centered method typically needs a great amount of training samples in IoT and the main issue with HSI is that labeled samples are not adequate and it can force to overfitting/underfitting issue. Motivated by this, the authors proposed a novel deep learning-based intelligent decision support system, which attains promising accuracies with limited training samples using Manifold Batch Structure (MFBS). Three novel approaches have been proposed to design MFBS. Firstly, a manifold batch scanning approach utilized to conclude the spatial association among the neighboring pixels and, the spectral associations between distinct bands. The proposed manifold batch scanning approach integrates the spatial and spectral association within different batches as well as extracts the collective spatial-spectral information. Secondly, since hyperspectral images have ample of unlabeled pixels, hence we refer such samples in the semi-supervised way while construction of convolution kernels. Finally, the MFBS has developed on a network infrastructure that does not include any hyper parameters for alteration. The experiments results on such standard datasets have revealed that MFBS outperforms various related existing HSI classification framework that too in case related to small training datasets.

Automated summary

In this paper, a novel deep learning-based intelligent decision support system is proposed, which achieves promising accuracies with limited training samples using Manifold Batch Structure (MFBS). Three novel approaches have been proposed to design MFBS, including manifold batch scanning approach, referring unlabeled pixels in a semi-supervised way, and a network infrastructure without hyperparameters.