Hyperspectral image classification approach via attention-driven spectral-spatial feature correlation enhancement

Abstract Given the characteristics of human object observation for classification, feature correlation in Hyperspectral Images (HSI) is key to accurately classify image. To better capture spatial and spectral features and explore their intrinsic correlations, this paper proposes a HSI classification approach via attention-driven Spectral-Spatial Feature Correlation Enhancement (ASSFCE). First, a Shallow spectral-spatial Feature Extraction (SFE)module is designed, which integrates 3DCNN and 2DCNN pipelines to simultaneously acquire fine spectral details and discriminative spatial information. Second, to fully explore global and local HSI feature correlations, the extracted shallow features are partitioned into multiple groups with the same channel number. A Feature Fusion module for Spectral Correlation Enhancement (SCEFF) based on Improved Depth-wise Separate Convolutions (IDSCs) is then proposed to facilitate inter-group feature interaction. Additionally, a Transformer Encoder (TE) module is employed to further extract more distinctive contextual information. Experiments on Indian Pines, Pavia University, and Houston 2013 datasets verify the approach effectiveness. Experimental results demonstrate that the proposed method can effectively capture the spatial-spectral correlation property of the HSI and achieve superior classification performance with overall accuracy of 99.21%, 99.78% and 99.41%.