Deep Spatial-Spectral Feature Extraction Network for Hyperspectral Image Classification

In this article, a deep spatial-spectral feature extraction network (DSSFEN) is proposed to enhance the spatial-spectral feature discrimination capability and improve the performance of hyperspectral image (HSI) classification. Multi-scale enhanced graph band sort module (MSEGBS) is designed to extract spectral features. A new dual-view spatial fusion CNN module (DSFC) is designed to extract comprehensive spatial features. To exploit the inner relationship of the spectral and spatial information, we design a novel spatial-spectral cross fusion attention module (SSCFA). Compared with some existing methods, extensive experiments on three popular HSI datasets show that our proposed DSSFEN can achieve better classification performance with 2% samples as training set.