Radiometric indices-based spectro-spatial approach for hyperspectral image classification

Abstract

For the Remote Sensing (RS) community, research on image classification remains a challenging issue aiming towards continuously enhancing understanding images content. Quite recently, Deep Learning (DL) based approaches have shown promising performances on various datasets. Nevertheless, for Hyperspectral images (HSI), using DL approaches is hampered due to the limited number of labeled data and the high dimensionality of the feature space, leading to the curse of dimensionality problem. In this paper, we propose a new classification approach of HSI images based on Low-Dimensional Feature Vector (LDFV) representation consisting of specific radiometric indices. This handcrafted image representation significantly reduces the feature space while keeping relevant spectral information. Conventional classifiers such as SVM or KNN can be used thereafter on the bases of the computed LDFV. Next, we consider the spatial information in order to revise the classification results. Our approach is tested on multiple HSI datasets including the ‘Indian Pines’ and the ‘Salinas Scene’ datasets, showing interesting results compared to the state-of-the-art methods while significantly reducing the processing time.