Research On Hyperspectral Unmixing Based On Autoencoder And Spectral Variability

Hyperspectral imaging technology has been widely used for identifying ground materials such as soil,vegetation,water bodies,and rocks,and for environmental monitoring.However,the large amount of hyperspectral image and the low spatial resolution of ground observations make manual recognition or annotation time-consuming and resource-intensive.Therefore,research on unsupervised hyperspectral image unmixing techniques is helpful for automation and improving efficiency.It enables faster and more accurate acquisition of ground material information and environmental monitoring.Currently,unsupervised hyperspectral unmixing algorithms face challenges such as spectral variability caused by environmental factors,poor convergence of complex models,difficulty in utilizing spatial information,and limited applicability to different scene images.To address these issues,this paper studies hyperspectral unmixing in multiple scenarios.The research focuses on improving spectral mixture models,optimizing unmixing algorithms,and spatial feature utilization.The main research contents are summarized as follows :(1)To address the underfitting problem of traditional spectral linear mixture models in complex environments,a scaled and perturbed spectral mixture model is proposed by incorporating the variation pattern of spectra.This model uses a scaling term to simulate the regular spectral reflectance scaling caused by terrain or lighting,and uses a perturbation term to capture irregular spectral disturbance.To ensure the interpretability of model parameters and improve convergence,a series of perturbation regularization terms are introduced as boundary constraints for model parameters.The alternating direction multiplier method is used to optimize all variables in the model.Experimental results show that the proposed unmixing algorithm can effectively improve the accuracy of abundance estimation on different datasets.(2)To address the problem of local optima in solving spectral models and the lack of adaptivity in regularization terms,a deep learning-based autoencoder is proposed to perform spectral unmixing.Specifically,a novel adaptive abundance smoothing unmixing network based on the autoencoder is proposed.This algorithm uses a multi-layer feedforward network to encode the abundance and a single-layer decoder to reconstruct the data.Furthermore,an adaptive abundance smoothing constraint is proposed to improve the disadvantage of strong prior assumptions.The experimental results indicate that the optimization method based on autoencoder has good flexibility,and the adaptive abundance smoothness constraint can effectively utilize spatial contextual information to improve the robustness of the algorithm.(3)To address the problems of low efficiency in utilizing spatial information in autoencoders and the imbalanced utilization of spectral and spatial information in unmixing,a series of autoencoder unmixing networks are proposed by combining gate mechanisms.The proposed networks use two branch architectures to extract spectral and spatial information respectively,which combines gate mechanisms with convolutional autoencoders and uses gate mechanisms to filter spatial contextual information.This balances the weight of spatial and spectral information and achieves adaptive fusion of the two features to improve the adaptability of the model.To improve the convergence performance of the model,a squared sine distance is proposed as the reconstruction loss function.Experimental results show that the autoencoder with multibranch structure and gate mechanism can achieve accurate unmixing effects under various land cover environments,demonstrating the high applicability of the proposed algorithm.(4)To address the challenges of feature extraction and fusion in autoencoder,a unidirectional local attention autoencoder unmixing algorithm based on augmented linear model is proposed.Based on this algorithm,a framework for autoencoder-based unmixing is developed.The proposed encoder adopts a unidirectional local attention mechanism that focuses on the central pixel and incorporates spatial information from neighboring pixels through two-level spatial filtering to effectively control the weight of spatial information in determining the unmixing results.The decoder employs scaled decoder networks for different endmembers and constrains parameters by modifying its feedforward propagation method.Experimental results demonstrate that compared to other algorithms,the proposed algorithm achieves higher accuracy and robustness in unmixing while maintaining good physical interpretability.