| Hyperspectral images(HSI)record not only the spatial information of ground objects but also the spectral information of ground objects,realizing the integration of atlas.At present,all countries in the world attach great importance to the research of hyperspectral image remote sensing technology.With the increasingly mature technology,hyperspectral image remote sensing has been widely used in agriculture,forestry,camouflage detection and other fields.However,due to the low spatial resolution of hyperspectral image,the multi-layer scattering among complex ground objects and the limitation of hyperspectral imager,there are a lot of mixed pixels in hyperspectral image.The existence of mixed pixels limits the development and application of hyperspectral unmixing.Hyperspectral unmixing can provide more detailed spectral and spatial information for subsequent hyperspectral image processing by extracting the end elements in each mixed pixel and estimating the abundance coefficient of the end elements in the mixed pixel.However,the problems of end-member variability in complex real objects and insufficient spatial resolution of images pose great challenges for hyperspectral unmixing.In recent years,as deep learning has made a great breakthrough in the field of image recognition,hyperspectral remote sensing research based on deep learning has been upsurge.In view of the above problems,this paper introduces a deep learning model in the field of image recognition,conducts the research around the directions of Hyperspectral Unmixing(HU),feature extraction,etc.The main contents are as follows:Aiming at the problem that spatial context information is not sufficiently extracted in the current hyperspectral unmixing research,propose a hyperspectral unmixing method based on space-spectrum joint adaptive convolutional network(SSJACN).Based on the self-coding structure,the network uses context adaptive convolution operation and 1× 1× 3 convolution operation respectively for spatial feature extraction and spectral feature extraction.Then,the extracted features are fused and propagated to enhance the utilization rate of spatial-spectral information in the unmixing network.In addition,to make use of the correlation between adjacent pixels,sparse constraints are added to the loss function to further enhance the sparsity of abundance.Finally,the synthetic data set and the real hyperspectral data set were used for experiments.The average spectral Angle distance(SAD)and root mean square error(RMSE)in the synthetic data set were 0.0270 and 0.0321,respectively,and the average spectral Angle distance in the real hyperspectral data set was 0.0792 and 0.0875,respectively.Experimental results show that this method has some advantages compared with other unmixing methods.In order to solve the problem that existing unmixing methods cannot simulate the variability of endmembers,multiscale self-Attention based variational adversarial autoencoder for hyperspectral Unmixing(MSAVAE)is proposed.In order to effectively utilize the local and global features of HSI in the encoder part of the network,a multiscale module(MSM)composed of 1×1 and 3×3 convolution kernel and a self-attention module(SAM)based on Transformer model are designed.The network extrapolates the potential representation of the Endmember from the observed pixels,encodes its variability,and generates the Endmember of each pixel from the potential representation,while estimating the abundance through the inference model.Finally,the abundance of the Endmember and the corresponding Endmember is linearly mixed.The network is used to avoid empty spectrum unmixing of Endmember variability by associating the generated Endmember with the probability distribution of Endmember samples extracted from the known Endmember library.In addition,considering the nonlinear modeling capability of neural networks,the proposed model can be fitted to any Endmember distribution,and the unsupervised method is used to extract and estimate the abundance of Endmember simultaneously.Finally,the synthetic data set and the real hyperspectral data set were used for experiments.The average spectral Angle distance and root mean square error in the synthetic data set were 0.0270 and0.0321,respectively,and the average spectral Angle distance in the real hyperspectral data set was 0.0792 and 0.0875,respectively.Experimental results show that this method has some advantages compared with other unmixing method. |
