Improved Three-Dimensional Inception Networks for Hyperspectral Remote Sensing Image Classification

Convolutional neural networks (CNNs) have been applied to hyperspectral image classification. Typically, fixed convolution kernel sizes are used in CNNs. This may be complicated by the high-dimensional features and the spatial-spectral features variability of hyperspectral data. In such cases, the fixed kernel size inhibits the ability to deal with different types of datasets flexibly. Therefore, we adopted a parallel architecture, an improved three-dimensional (3D) inception network, for hyperspectral image classification. Additionally, the dimensionality reduction based on interactive information entropy is used in the adaptive band selection of hyperspectral remote sensing data, then the low-redundant bands with more abundant and discriminative information can be selected. Experimental results based on the publicly available hyperspectral datasets demonstrated that the improved 3D inception network can perform well with limited training samples. Moreover, the combinations of different dimensional convolutions in the 3D-2D and 3D-2D-1D inception networks achieved comparable classification results. The experimental results indicate that the proposed networks are generalized classification models with high classification rates.