Research On Hyperspectral Image Classification Based On Deep Learning And Spatial Spectral Combination

Hyperspectral image classification is an important branch of remote sensing image processing.Hyperspectral images are high-dimensional image data with hundreds of spectral bands that include both spatial information and spectral information.Hyperspectral images play an important role in urban planning,water conservancy agriculture,resource detection,geological disaster detection,and military reconnaissance.Therefore,the classification of hyperspectral images is important.Traditional classification methods treat hyperspectral images as ordinary images.The shallow classification models proposed by them often have difficulty in extracting the essential features of the images,which affects the classification accuracy.With the emergence of deep learning,it has been widely used in image processing related fields due to its ability to extract the deep features of input data and the ability to nonlinearly map.However,due to the problems of the same-spectrum foreign matter and the same-object hetero-spectrum in the hyperspectral image,the application of deep learning to hyperspectral images is facing serious challenges.This article focuses on the three problems in hyperspectral images,namely the problem of difficult sample mining,the problem of fewer labeled samples,and the problem of fusion of spectral information and spatial information.We propose our solution.The specific research content is as follows:(1)Aiming at the problem of difficult samples in hyperspectral images,we use the Focal loss in target detection to increase the weight of difficult samples in the network loss.We apply Focal loss in the 3DCNN network we designed and SSRN network.In the experiments on the two data set,the utilization of Focal loss has improved the classification effect.(2)Aiming at the problem of fewer labeled samples in hyperspectral images,we designed a three-dimensional convolutional coding-decoding structure that extracts features from the input hyperspectral data,and then recovers and reconstructs images from these features.The network can allow unlabeled samples to participate in training and assist in the classification of labeled samples,which is the category of semi-supervised learning.In addition,in order to make the feature distance between similar samples as close as possible to improve the classification accuracy,we introduced the center loss in metric learning to control the feature distance of similar samples.In the experiments on two widely used datasets,our network achieved the best classification results.(3)Aiming at the problem of spatial spectrum feature fusion of hyperspectral images,we analyzed the problem of spatial spectrum feature fusion of SSRN network,and then proposed a two-branch spatial spectrum feature fusion classification network.The two branches of the network extract the spatial and spectral features of the hyperspectral image,and apply the spatial attention mechanism after extracting the spatial features,and apply the channel attention mechanism after extracting the spectral features.These two mechanisms can strengthen the extracted features and make the features more discriminative,Thereby improving the classification effect of hyperspectral images.