Spectral-Spatial Hyperspectral Image Classification using Deep Learning

Authors:
Publisher website
Google Scholar
Abstract
Hyperspectral images contain multidimensional data which leads difficulty in processing of a large amount of data. Existing features extraction and selection techniques cannot be applied to extract deep features present in the image. In this paper, a deep learning based spectral-spatial framework is proposed which is a hybrid of feature extraction and feature selection. Firstly, the captured Hyperspectral image is atmospherically corrected due to which a large number of present bands may lost. To accommodate this loss, interpolation is applied on the input image. Afterward, the interpolated image is processed by the random forest algorithm for selection of its important spectral features. Then the spatial-dominated features are extracted using masking approach, which are further merged with the selected spatial features to form hybrid features. These features are passed to stacked autoencoders for the extraction of deep features. These deep features are further processed by logistic regression for classification. The proposed framework generated accurate classification results as compared to the state-of-art classifiers.

This publication has 24 references indexed in Scilit: