Bands Sensitive Convolutional Network for Hyperspectral Image Classification

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