Local Transformer With Spatial Partition Restore for Hyperspectral Image Classification

Convolutional neural network (CNN) has exhibited enormous potentials in hyperspectral image (HSI) classification owing to excellent locally modeling ability. Although performance of CNN-based methods been witnessed, there still have some limitations their internal backbone. On the one hand, long-distance context dependencies is an inborn defect, which leads receptive field limitation and insufficient feature capture HSI. other usually need various sample distribution train cannot infer dynamically, may not inherent changes HSI data well. To overcome above issues, we propose a novel local transformer with spatial partition restore (SPRLT-Net) for classification. Firstly, introduced obtain attention weights dynamically by measuring similarity between related pixel pairs. Secondly, (SPR) module designed split input patch into several overlapping continuous sub-patches as sequential. With obtained at SPR restores sequential original patch. Finally, fully connected layer used SPRLT-Net can global dependencies, dynamical adapt pixels. Experimental results based on spatially disjoint samples randomly selected five benchmark sets demonstrate that outperforms state-of-the-art terms accuracy, generalization performance, computational complexity.