Research On Hyperspectral Image Classification Algorithm Based On Multiscale Spatial-spectral Fusion

Hyperspectral remote sensing images have rich spectral information and spatial information,which are widely used in mineral exploration,environmental monitoring and precision agriculture.Hyperspectral image classification is one of the important research topics of hyperspectral remote sensing technology.It can use labeled samples to divide unlabeled samples to achieve accurate classification of ground objects.However,most of the space-borne or airborne hyperspectral images have high spectral resolution and low spatial resolution,and the ground objects inside the pixels are mixed.The bands obtained by the imaging spectrometer are interfered by their own and external noises.The spectral bands are redundant,and the phenomenon of “pepper salt” and the phenomenon of “Same thing with different spectrum,different thing with same spectrum” are prominent.Therefore,how to effectively use spectral information and spatial information to improve classification accuracy is one of the hot issues in current research.At the same time,limited training samples often lead to insufficient feature extraction and poor classification effect of existing hyperspectral image classification algorithms,which is one of the urgent problems to be solved.In view of the above two problems,this paper uses superpixel,multi-scale convolution kernel and attention mechanism to improve the utilization of spectral and spatial features,and proposes two mixed convolution classification models.Firstly,the background,significance,research status and problems of hyperspectral image classification are expounded,and the application of convolution network in hyperspectral image classification is briefly introduced.Secondly,aiming at the shortcomings of existing convolutional network classification algorithms,the use of spatial-spectral features and small sample classification are improved by using superpixel segmentation and attention mechanism.Finally,in order to prove the effectiveness and generalization of the proposed model,simulation verification is carried out on three public data sets.The main research contents of this paper are as follows:1.Aiming at the problem of insufficient utilization of spatial-spectral features in existing hyperspectral image classification algorithms,and the problem of insufficient spatial feature extraction of fixed convolution kernels in existing convolution networks.Considering the spatial structure characteristics of ground objects distribution and the spectral consistency of similar ground objects,this paper proposes a hyperspectral image classification algorithm based on hybrid convolution capsule network with multi-feature fusion.Firstly,the principal component analysis algorithm is used to reduce the dimension and obtain the first several principal components.Then,the simple noniterative clustering segmentation algorithm is used to segment and average the 3D image,and the angle cosine clustering is used for clustering.Then,the obtained different features are superimposed,normalized and classified through the proposed mixed convolution capsule network;Finally,three different public datasets are used to verify the proposed hybrid classification algorithm.The experimental results show that the proposed algorithm can effectively utilize spatial and spectral features,and has good classification performance and generalization.2.In view of the problems of over-fitting of deep classification model and low classification accuracy caused by the lack of training samples of hyperspectral images.In this paper,the advantages and disadvantages of maximum pool and average pool downsampling are comprehensively analyzed,and the spatial pool attention mechanism and spectral pool attention mechanism are proposed considering the effectiveness of attention mechanism in hyperspectral image classification.Firstly,principal component analysis is used to reduce the image dimension to 20 dimensions;Then,the original image and the reduced-dimensional image are respectively input into the spectral branch and spatial branch of the proposed bi-branch hybrid convolution network based on the multi-scale pool attention mechanism for classification.Finally,three public datasets are used for verification.Experimental results show that the proposed classification algorithm can obtain better classification results in small samples.