Hyperspectral imagery classification based on semi-supervised 3-D deep neural network and adaptive band selection. (1st September 2019)

• Record Type: Journal Article
• Authors: Sellami, Akrem; Farah, Mohamed; Riadh Farah, Imed; Solaiman, Basel
• Is Part Of: Expert systems with applications. Volume 129(2019)
• Page Start: 246
• Page End: 259
• Publication Date: 2019-09-01
• Subjects: Hyperspectral imagery classification -- Convolutional neural network (CNN) -- Adaptive dimensionality reduction -- Deep learning; Expert systems (Computer science) -- Periodicals; Systèmes experts (Informatique) -- Périodiques; Electronic journals; 006.33
• DOI: 10.1016/j.eswa.2019.04.006
• Languages: English
• ISSNs: 0957-4174

Abstract

Highlights: Adaptive Dimensionality Reduction for the selection of relevant spectral bands. Selecting the most relevant spectral bands using limited number of training samples. Semi Supervised 3D Convolutional Neural Network for image classification. Extracting deep spectral and spatial features based on convolutional encoder-decoder. Enhancing image classification compared to well-established deep learning methods. Abstract: This paper proposes a novel approach based on adaptive dimensionality reduction (ADR) and a semi-supervised 3-D convolutional neural network (3-D CNN) for the spectro-spatial classification of hyperspectral images (HSIs). It tackles the problem of curse of dimensionality and the limited number of training samples by selecting the most relevant spectral bands. The selected bands should be informative, discriminative and distinctive. They are fed into a semi-supervised 3-D CNN feature extractor, then a linear regression classifier to produce the classification map. In fact, the proposed semi-supervised 3-D CNN model seeks to extract the deep spectral and spatial features based on convolutional encoder-decoder to enhance the HSI classification. It uses several 3-D convolution and max-pooling layers to extract these features from the selected relevant bands. The main advantage of the proposed approach is to reduce the high dimensionality of HSI, preserve the relevant spectro-spatial information and enhance the classification using few labeled training samples. Experimental studies are carried out on three real HSI data sets: Indian Pines, Pavia University, and Salinas. The obtained results show that the proposed approach performs better than other deep learning-based methods including CNN-based methods, and significantly improves the classification accuracy of HSIs.