Multi Spectral-Spatial Gabor Feature Fusion Based On End-To-End Deep Learning For Hyperspectral Image Classification
Résumé
The use of low-level spatial information, in addition to the rich spectral information of Hyperspectral Image (HSI) with the potential of deep learning-based methods, has provided high performance in improving the HSI classification. With this regard, Gabor Filtering has been successfully applied for HSI pixels analysis due to its ability in extracting representative spatial features. Therefore, a novel methodology based on Gabor filtering and multi-view Convolutional Neural Networks (multi-view CNNs) was proposed in this paper. Firstly, Gabor texture features at different scales and orientations performed on the first three principal components of HSI were extracted. Then, spectral and spatial features are introduced separately through spectral, and spatial CNN, respectively. Afterward, their learned feature maps are combined and fed into our spectral-spatial CNN (SSCNN) to learn a fused multi-view representation. Experiments applied on two real HSI datasets, including Indian Pines, and Salinas yield competitive classification performances compared to other state-of-the-art deep learning methods.