Research On Hyperspectral Images Super-resolution Reconstruction And Classification Algorithm

Hyperspectral imaging technology could obtain hundreds of continuous spectral bands from surface objects,which contains a large amount of spectral information and spatial information.These rich spectral information and spatial information could be used to accurately distinguish different land cover types.In recent years,hyperspectral images have been widely applied in many fields,including environmental monitoring,mineral exploration,agricultural production,change area detection,plant disease diagnosis,land cover classification.However,hyperspectral image usually has extremely low spatial resolution due to the limitation of hardware sensors.Therefore,there is an urgent need to obtain high quality hyperspectral images to provide a data base for many subsequent related applications such as image classification.In addition,due to the complex compo-sition of hyperspectral image features,a large amount of labeled data is required as the training set when performing hyperspectral image classification,as well as the problem of possible unknown class targets in the actual scene is not considered.To address the above issues,this paper focuses on the super-resolution reconstruction and classification methods for hyperspectral images.The main research work and innovation of this paper are as follows.(1)To address the issue of low spatial resolution of hyperspectral image,this paper proposes a hyperspectral image super-resolution reconstruction method DesP3 D based on pseudo-3D convolution and dense connection,which combines dense connection struc-ture and pseudo-3D convolution to force the network to learn deeper detail information,and can establish mapping relationships from low to high resolution at a faster training convergence rate.Experiments on three datasets show that the DesP3 D method proposed in this paper can effectively recover high quality hyperspectral images.(2)To address the problems of relying on a large number of labeled samples for training and the classification effect to be improved when classifying hyperspectral im-ages,this paper proposes a hyperspectral image closed-set classification method PBEM based on the EM attention mechanism,which applies the expectation-maximizing atten-tion mechanism to hyperspectral image classification and can effectively reduce the dif-ferences between samples of the same category while maintaining the differences between targets of different categories.Experimental results on four hyperspectral datasets show that the proposed PBEM method has significant advantages over other methods in terms of classification performance and time consumption in the case of very small training samples.(3)For hyperspectral image classification without considering the unknown class of objects in the environment,this paper proposes a DANCE domain adaptive hyper-spectral image open-set classification method RELS,which systematically deals with the hyperspectral open-set classification problem using a domain adaptive model based on self-supervised learning.Adequate open-set classification experiments are conducted on three datasets.The experimental results show that the proposed method can get good classification results for both known category samples and unknown category samples.