Spectral-spatial feature extraction method for hyperspectral images classification using multiscale superpixel and covariance map

Abstract

In this paper, a hand-crafted spectral-spatial feature extraction (SEA-FE) method for classification of hyperspectral images (HSIs) is proposed to improve the classification performance, especially in the limited labelled training samples. Usually, spatial information (SPI) is extracted from the neighborhood of each pixel. To overcome the shortcoming of the traditional method, i.e., fixed square window (SW), superpixel analysis is used to construct the neighborhood regions. Also, to reduce the problems of selection the optimal superpixel size, multiscale framework is applied where each superpixel is known as a feature map (FM). Then, SEA-FE combines the FMs together to exploit the spatial structure by calculating the covariance map (CM) as feature coding strategy (FCS). The CMs are mapped from manifold space (MS) to Euclidean space (ES) to serve as direct input for classical learning methods. The experimental results on three HSI datasets demonstrate the effectiveness of the SEA-FE compared to several FE methods.