Hyperspectral Image Classification using Joint Spatial-Spectral Feature Extraction

Abstract

Hyperspectral images (HSI) are playing a dominant role in many applications as they contain high spectral and spatial information. The HSIs contain hundreds of attached spatial-spectral bands. The extraction of joint spectral-spatial features becomes a difficult task due to the correlation between bands. The deep learning models provide a solution to the extraction of spatial-spectral features by introducing various kernels. Thus, this article is focused on the implementation of deep learning convolutional neural network (DLCNN) based feature extraction with a bio-optimization-based feature selection mechanism. Initially, enhanced guided image filtering (EGIF) is developed to extract the spatial features. In addition, Residual Network 50 (ResNet50) is introduced to extract the color-based three-dimensional spectral features from the HSI dataset. Finally, DLCNN is applied to classify the various bands of HSI. The subjective and objective performance of proposed hybrid network resulted in superior performance as compared to state-of-art approaches in terms of overall accuracy (OA), average accuracy (AA), precision, recall and F1-score.