Xin He, Yushi Chen, and Pedram Ghamisi
An overview of dual graph convolutional network (DGCN) for hyperspectral image (HSI) classification: data pre-processing, CNN feature extraction, and dual graph learning.
Hyperspectral images (HSI) combine two-dimensional spatial information with high-resolution spectral information. They are used for a variety of applications, such as urban planning, vegetation monitoring, and target detection. To process the images, HSI classification assigns an object category to each pixel of the image. HSI classification remains a challenging task due to high complexity of the data, which represent a hypercube, and limited labelled samples.
Here, we introduce dual graph convolutional network (DGCN) for supervised HSI classification with limited training samples. The model represents a hybrid network of a convolutional neural network (CNN) and two graph networks: the point graph and the distribution graph. CNN extracts spectral-spatial information from individual samples without any pre-processing step to preserve discriminative features. The learned features are passed to the two integrated graph networks to capture the relationships between the samples. The point graph calculates the similarities between the output features of the CNN; the distribution graph explores correlations between the samples with the same label. DGCN thus significantly reduces the number of the required training samples. To mitigate the problem of over-fitting caused by limited sample size, drop edge is incorporated into DGCN. Multi-scale feature cut-out is proposed as a regularization method to further reduce the over-fitting problem and improve the generalization of the model. Compared to the state-of-the-art methods, our model shows superior classification performance on four benchmark hyperspectral data sets (Salinas, Indian Pines, Pavia, and Houston).
IEEE Transactions on Geoscience and Remote Sensing, 2021-03-08.