Hyperspectral image classification based on Hybrid 2D-3D CNN

Abstract

In recent years, some deep learning models represented by neural networks have gradually become the focus of hyperspectral image classification research.Based on the existing network structure HybridSN and the high-dimensional characteristics of hyperspectral image data, this paper proposes SE-HybridSN. The channel attention module is added to the structure in order to calibrate the importance of features, achieving the purpose of reducing data redundancy and improving classification accuracy. Principal component analysis (PCA) is used to alleviate the “Hughes” phenomenon. At the same time, embed both 2-Dand3-D convolutional structures in the model, the“spectral-spatial joint features”can be extracted simultaneously and sequentially. Two well-known hyperspectral image datasets, Indian Pines and Pavia University, were selected for this paper to verify the feasibility of the model. When compared with other neural network structures, this method has higher overall accuracy, average accuracy and Kappa.