Hyperspectral Image Classification Based on SE-Res2Net and Multi-Scale Spatial Spectral Fusion Attention Mechanism

Abstract

<p indent=0mm>In order to extract more discriminative features for hyperspectral image and prevent the network from degrading caused by deepening, a novel multi-scale feature extraction module SE-Res2Net based on the new dimensional residual network (Res2Net) and squeeze and exception network (SENet), and a multi-scale spectral-spatial fusion attention module is developed for hyperspectral image classification. In order to overcome the degradation problem caused by network deepening, the SE-Res2Net module uses channel grouping to extract fine-grained multi-scale features of hyperspectral images, and gets multiple receptive fields of different granularity. Then, the channel optimization module is employed to quantify the importance of the feature maps at the channel level. In order to optimize the features from spatial and spectral dimensions simultaneously, a multi-scale spectral-spatial fusion attention module is designed to mine the relationship between different spatial positions and different spectral dimensions at different scales using asymmetric convolution, which can not only reduce the computation, but also effectively extract the discriminative spectral-spatial fusion features, and further improve the accuracy of hyperspectral image classification. Comparison experiments on three public datasets, Indian Pines, University of Pavia and Grss_dfc_2013 show that the proposed method has higher overall accuracy (OA), average accuracy (AA) and Kappa coefficient compared to other state-of-the-art deep networks.