Mirror Mosaicking Based Reduced Complexity Approach for the Classification of Hyperspectral Images

Convolutional neural networks (CNNs) are top-rated to classify hyperspectral images. Usually, these use the spectral-spatial approach (SSA), in which the patch corresponding to each pixel to be classified extracted from the hyperspectral image. The size of the patches' spatial neighborhood plays a vital role in the complexity of the designed CNN model. The complexity of the model proportionately increases according to the spatial size of patches. Generally, patches are of odd-squared spatial size centering at the corresponding pixel. In this paper, a novel approach based on mirror mosaicking (MMA) has been proposed. It has been compared with the spectral-spatial approach using minimally sized patches. The proposed approach has been proved computationally efficient along with competitive classification performance. A dataset provided by National Ecological Observatory Network (NEON) has been used for the experimentation, which has three major classes, viz. vegetation, soil, and road.