A CNN with Multiscale Convolution for Hyperspectral Image Classification Using Target-Pixel-Orientation Scheme

Recently, convolution neural network (CNN)-based hyperspectral image (HSI) classification has gained attention due to its remarkable performance when the number of training samples is sufficiently large. Existing approaches involve 1-D, 2-D or 3D CNN based architecture. 3D CNN involves heavy computation and others are inefficient in using spatial and spectral information jointly. In this work, we propose to use pointwise 3-D convolution to extract spectral feature, followed by 2-D convolution to extract spectral-spatial features jointly in an end to end manner. We have shown a variation in inception-like high level architecture for feature extraction. This is guided by the fact that averaging a large number of feature may loose some unique information. On the other hand, existence of spatial variability within a class and similarity among different classes incurs degradation in HSI classification. To combat with that, we propose a novel target-patch-orientation (TPO) scheme to form a spatial-spectral neighborhood of a pixel. The Experimental results reveal that TPO scheme has positive impact on the proposed model.