Hyperspectral Image Hybrid Convolution Classification under Multi-Feature Fusion

To address the problem of insufficient utilization of spatial-spectrum features in existing convolutional neural network classification algorithms for hyperspectral remote sensing images, we propose a hyperspectral image classification strategy based on a hybrid convolution capsule network under multi-feature fusion. First, a combination of principal component analysis and non-negative matrix decomposition is used to reduce the dimensionality of a hyperspectral dataset. Second, the principal components obtained through dimensionality reduction are used to generate a multidimensional feature set through super-pixel segmentation and cosine clustering. Finally, the superimposed feature set is used to extract spatial-spectrum features through a two-dimensional and three-dimensional multi-scale hybrid convolutional network, and a capsule network is used to classify them. We performed experiments on different hyperspectral datasets, and the results revealed that under the same 20-dimensional spectral setting, the proposed strategy significantly improves the overall accuracy, average accuracy, and Kappa coefficient compared to traditional classification strategies.