Spectral-Spatial Fused Attention Network For Hyperspectral Image Classification

Hyperspectral image classification has been a research hotspot in the field of remote sensing. Traditional methods are limited by their poor robustness. Recently, convolutional neural network, as the common architecture in deep learning, has achieved superior performance in feature extraction and becomes the mainstream method of hyperspectral image classification. However, due to the redundancy of hyperspectral image, the distinguishing spectral-spatial features for classification are often hard to acquire. In this paper, a novel spectral-spatial fused attention module is proposed for hyperspectral image classification. The module contains three parts. The first part is designed to extract the correlation among the bands. The second part aims to acquire the common spatial positions. Different from the former two parts, the stable spatial features and the contributions of neighborhoods to the center spectrum are explored in the last part. In addition, the identical modules are stacked sequentially in the proposed network to extract the significant spectral-spatial features. The experimental studies on two publicly available datasets reveal the effectiveness of the proposed method