Hyperspectral Image Classification Based On Contextual Feature And Recurrent Neural Network

In recent years,due to the unique properties and the large amount of information contained in hyperspectral data,hyperspectral image processing has become one of the hot spots in the field of remote sensing image research.And the hyperspectral image classification task plays a substantial role in geological exploration,crop detection,national defense and other fields,worthy of more in-depth study.However,in the hyperspectral image classification task,the acquisition and learning of the data feature has always been the focus and diffi-culty of the research.How to extract the abundant and effective features directly affects the classification result.In this paper,we start from the contextual features,introduce the relevant ideas in the field of natural language processing into the hyperspectral image classi-fication problem,and use the deep learning framework to extract and integrate the low-level features of hyperspectral images,to get better and more discriminative high-level semantic features,while merging feature learning and classification as a whole framework to enhance the classification result.The main works of this paper are as follow:1.We propose a spatially constrained bag-of-visual-words model for hyperspectral image classification,extending the concept of document and word in natural language processing field to hyperspectral image processing field,while improving the traditional bag-of-visual-words method for hyperspectral image characteristics.This method builds documents by superpixel segmentation,and constructs visual words on the basis of traditional low-level features,then calculates the visual words histogram feature of the document as semantic feature,which is fed into classifier for completing the classification.The method integrates the spatial-spectral information,and utilizes the local spatial feature efficiently,and obtains the semantic feature to improve the classification result.2.We propose a local space sequential recurrent neural network model for hyperspectral image classification,extracting more discriminative high-level semantic feature of hyper-spectral data.Based on traditional low-level features,this method extracts the local spatial sequence feature of hyperspectral data samples and uses the recurrent neural network to integrate and abstract the local spatial information into high-level semantic feature,while enhancing the influence of beneficial pixels and reducing the effect of other harmful pixels,improving the classification accuracy.3.We proposed a non-local sequential recurrent neural network model for hyperspectral image classification,adding global information on the basis of local spatial sequence fea-ture and enhancing the representation of high-level semantic feature.This method finds the non-local similar samples on the global scale and extracts the corresponding LSS feature of them including the original samples,which together construct the non-local spatial sequence feature(NLS)of the original samples.NLS not only preserves the local spatial information,but also integrate the information of non-local similar samples,so the high-level semantic feature contains much more knowledge which is better for hyperspectral image classifica-tion.