Locality Preserved MLP-Mixer for Hyperspectral Image Classification

Recently, more and more multilayer perception (MLP) models have been proposed for hyperspectral image (HSI) classification processing since MLP-based architecture can perform as well as the convolution-based models without convolutions or self-attention. However, convolutional neural network (CNN) focus on achieving both spectral and spatial perception in global receptive field, which failed to consider the intrinsic manifold structure of HSI data. To address this issue, we propose a novel locality preserved MLP-mixer (LPMLP) method for HSI classification by integrating manifold learning and MLP-mixer architecture into one model. In particular, the proposed LPMLP method first constructs an intrinsic adjacency graph to capture the locality manifold structure information and learns a projection to extract the representative features with redundancy removed from the original data. Then, both locality manifold structure information and extracted features are fused and input into MLP-mixer architecture to capture long-range dependencies and extract more discriminative features for achieving better classification performance. Finally, experiments conducted on two HSI data sets validated the effectiveness of the proposed method.