Bands sensitive convolutional network for hyperspectral image classification

Hyperspectral image (HSI) classification deals with the prob-lem of pixel-wise spectrum labelling. Traditional HSI clas-sification algorithms focus on two major stages: feature ex-traction and classifier design. Though studied for decades, HSI classification hasn't been perfectly solved. One of the main reasons relies on the fact that features extracted by embedding methods can hardly match an ad hoc classifi-er. Recently, deep learning methods achieve an end-to-end mechanism and can learn features suitable for classification from the raw data. Inspired by the newly proposed work on deep learning for HSI classification, in this paper, we propose to build a deep convolutional network based on the analysis of spectral band discriminative characteristics. More specif-ically, we first split the spectrum bands into groups based on their correlation relationships. Then we build a band vari-ant CNN submodel, where each group is modelled by one of those submodels. Meanwhile, a conventional CNN model is also learned globally on the spatial-spectral space, to main-tain robustness of submodel changes. Lastly, we concatenate the global CNN model and band-specific CNN submodels to one unique model. In this way, global robustness and band variance are mixed together. Experiments on publicly available datasets demonstrate the great performance of the proposed method.