Research On Hyperspectral Remote Sensing Image Classification Based On Dual Channel Convolution Neural Network

With the development of hyperspectral remote sensing image technology,the use and analysis of data provided by hyperspectral remote sensing image has become a key research issue in this field.Accurate object recognition and classification of hyperspectral remote sensing images can obtain a large amount of useful information from the hyperspectral remote sensing images,which has been widely used in environmental monitoring,geological exploration,military defense and other fields.Traditional classification methods can only extract shallow features,can not make full use of spectral and spatial information,and have limited classification ability.The in-depth learning method can extract the deeper and more comprehensive features of hyperspectral remote sensing images by building a multilayer neural network model.However,the existing in-depth learning models do not make full use of the spatial spectral structure of the image in the feature extraction process,resulting in inadequate feature extraction,poor classification accuracy and other issues.To solve these problems,this paper carries out related research work,including:(1)The classification methods of hyperspectral remote sensing images with good classification results in current research are introduced.Traditional classification methods based on manual feature extraction include support vector machine combined with principal component analysis,composite kernel-based classification and Markov random field-based classification.Classification methods based on in-depth learning include one-dimensional,two-dimensional,three-dimensional convolution neural networks,one-dimensional antagonism generation networks,cyclic neural networks,twin neural networks,spatial spectral residual networks and fast dense spectral spatial convolution networks.(2)A data expansion method for stacking spatial transformation information is proposed to increase the number of available training samples and enrich the spatial information of each pixel point for extracting spatial features.For each pixel point in the hyperspectral remote sensing image and the pixel block within its neighborhood,this method explores the potential spatial information of the pixel point by rotating clockwise,counterclockwise,and row-column transformation to expand the training sample set.A two-channel convolution neural network classification model is pre sented for feature extraction and final classification of hyperspectral remote sensing images.The network uses two parallel convolution neural networks to process different size hyperspectral remote sensing image pixel blocks with different branches,which can better extract more essential spatial-spectral joint features for more accurate and efficient classification.Experiments show that this method achieves better classification results than other deep learning algorithms on all three open source datasets when 10% of the total sample is randomly selected as the training sample..(3)A two-channel convolution neural network classification model that fuses attention mechanism is proposed to achieve an effective classification with fewer samples.The network effectively applies the channel attention module and spatial attention module to the classification of hyperspectral remote sensing images,enabling the network to learn more important and decisive spatial-spectral joint features.After experimental verification,this method can still achieve satisfactory classification results when 3% of the total sample is randomly selected as the training sample.