Improved Multi-Scale Fusion of Attention Network For Hyperspectral Image Classification

With the development of remote sensing technology, hyperspectral images that carry both spectral information and spatial information have attracted much attention. As a significant task,hyperspectral image classification(HSI) needs to fully extract spectral and spatial features to better determine the category. However, when extracting spatial features, the difference in object scale often affects the classification effect. It is difficult to distinguish both larger and smaller objects at the same time. In this paper, we propose an improved spatial multi-scale fusion scheme for the spatial extraction network, combining spatial patches sampled at different scales to achieve the effect of accurately classifying objects of different sizes. And the entire network is based on the attention mechanism, following the attention setting of SSAN, which makes the network pay more attention to pixels in key bands and key spaces.The experimental results on the data set Pavia Center prove that our method has achieved a greater accuracy improvement, reaching expected performance.