Research On Image-Level Classification Technology Of Hyperspectral Images Based On Deep Learning

Hyperspectral image(HSI)classification methods based on deep learning have made remarkable achievements,which can effectively improve the classification accuracy.Among them,the patch-level classification methods have made great progress by using the local patches after segmentation as the processing unit.However,patch-level classification methods also have some inherent defects,such as difficulty in utilizing global information,the requirement to segment images,and the low efficiency of the model.The image-level classification methods using the global image as the processing unit can effectively avoid these inherent defects.To this end,this paper takes image-level input as the main research body,and focuses on the image-level classification methods of hyperspectral image based on deep learning,focusing on solving the main problems of global classification map detail distortion and class boundary expansion,the contradiction between limited graphics card memory and large-size image input,poor performance in limited samples size,and insufficient generalization ability of the model,fully exploiting its potential and advantages in classification problems.The main work and innovations of this paper are as follows:(1)A classification method for high-resolution features network of hyperspectral images(HSI-HRNet)is designed,which is based on the fully convolutional neural network to perform parallel computing and cross fusion of multi-resolution features of images,so as to alleviate the loss of key information caused by the traditional serial flow mode of features.Two targeted strategies of multi-resolution feature joint supervision and voting classification are used to improve the classification performance.Experimental results show that compared with the existing patch-level and image-level classification methods,HSI-HRNet can obtain superior classification results and improve the effect of the classification map.At the same time,it significantly reduces the training and classification time,which is more efficient in practical applications.(2)A classification method for feature resolution lossless network of hyperspectral image(FOct Conv PA)is proposed,which maintains the lossless flow form of feature resolution in the whole process by deleting the downsampling step,so as to extract elaborate spectral-spatial features,and enhances the network performance by embedding modules such as Octave fully convolution.In addition,in order to solve the conflict between the limited memory of the graphics card and the large size of the image,an image resolution reconstruction classification workflow is proposed,and the workflow can be directly applied to different image-level classification methods,which provides a solution for processing large size images.Experimental results show that the proposed method can further improve the classification performance,and alleviate the widespread phenomenon of classification map detail distortion and boundary expansion in the previous methods.(3)A self-supervised learning framework for spectral variation feature(SVF-SSL)is proposed.By pre-training on a large number of unlabeled hyperspectral image data,the necessary feature learning process can be preloaded,so that the model can learn the ability to extract spectral variation.Experimental results show that SVF-SSL can obtain superior classification performance in the case of limited samples size,such as less noises,high accuracy and detailed description of the global classification map,which alleviates the poor performance of the previous supervised learning methods when the number of training samples are insufficient.At the same time,the network can obtain robust generalization feature extraction ability through self-supervised learning,so as to have the cross-domain classification ability of fast response to the target image,avoiding the tedious and repeated training problem when the previous in-domain classification method faces a new image to be classified.