Hyperspectral Classification via Global-Local Hierarchical Weighting Fusion Network

The fusion of spectral–spatial features based on deep learning has become the focus research in hyperspectral image (HSI) classification. However, previous frameworks usually performed feature aggregation only at branch ends. Furthermore, first-order statistical are considered process, which is not conducive to improving discrimination features. This article proposes a global–local hierarchical weighted end-to-end classification architecture. architecture includes two subnetworks for spectral and spatial For subnetwork, band-grouping strategies designed, bidirectional long short-term memory used capture context information from global local perspectives. pooling strategy attention combined construct module enhance discriminability learned by convolutional neural network. stage, weighting mechanism developed obtain nonlinear relationship between both experimental results four real HSI datasets GF-5 satellite dataset demonstrate that method proposed more competitive terms accuracy generalization.