Hyperspectral Image Classification Based On Multi-scale Space-spectral Joint Feature Network

With the rapid development of remote sensing technology,the use and analysis of hyperspectral image data has become one of the research hotspots in the field of remote sensing.Through the rational use of these remote sensing data,accurate recognition and classification of ground objects can be realized,which has been widely used in Urban Surveying and mapping,meteorological prediction and geological exploration.More and more remote sensing scholars apply deep learning theory to hyperspectral image data in order to better extract the deep features of hyperspectral image.However,in the existing deep learning methods,there is no good solution to the problems of high image dimension and large amount of data.At the same time,the correlation between convolution kernels is not considered in the existing multi-scale feature extraction algorithms.In view of the above problems,this paper has carried out relevant research work,including:(1)A multi-scale spectral spatial residual network based on three-dimensional channel and spatial attention mechanism is proposed for hyperspectral image classification.The network adopts a three-layer parallel residual network structure,and uses three-dimensional convolution kernels of different sizes to continuously learn the characteristics of spectral and spatial residual blocks from the rich spectral and spatial features.Finally,the extracted depth multi-scale features are stacked and input into the 3D attention module to enhance the expressiveness of image features from both channel and space angles,so as to improve the accuracy of classification.Experimental results show that the proposed method achieves better classification results than other deep learning algorithms on three common hyperspectral datasets.(2)A multi-scale spectral spatial weighted kernel network based on adaptive attention mechanism is proposed for hyperspectral image classification.In this network,PCA algorithm and one-dimensional convolution are combined to extract his spectral information,while the spatial information is extracted by three branch structure,using different convolution kernel,and the size of network receptive field is adaptively adjusted by the attention of different branches through attention mechanism.The experimental results show that the network has advantages in filtering and integrating the information extracted from multi branch structure.Compared with other deep learning algorithms,it has the characteristics of short training time and high classification accuracy.