| Hyperspectral remote sensing which can obtain spectral information and spatial information for the imaged scenes,and have widely used in environmental monitoring and natural disaster monitoring.However,due to limitation spatial resolution and spatial complexity,there are a large number of mixed pixels.The existence of mixed pixels hinders the application scope of hyperspectral remote images greatly.Therefore,solving the problem of mixed pixels is a relevant problem for the accurate analysis and poses a technical difficulty for quantitative remote sensing.Spectral unmixing is the most effective method to solve mixed pixel issues,aimed at obtaining the pure spectral signatures or endmembers and its corresponding fractional abundance.Recently,the unmixing model based on nonnegative matrix factorization(NMF)technology has been a research hotspot,but there are still some issues.Specifically,three new algorithms developed in this work can summarized as follows.(1)A new spectral weighted sparse non-negative matrix factorization framework is proposed.As the objective function of NMF is non-convexity,it usually falls into local minimum value.To overcome the limitation,some constraints will be added according to the specific problems,sparse constraint is most commonly used.However,this kind of method has the problem that the sparse prior representation of abundance coefficient is insufficient,which leads to the unstability of the algorithm.To address this problem,a new spectral weighted sparse nonnegative factorization spectral unmixing is proposed.The spectral weighted is used to describe the sparsity of abundance coefficients,which further imposing on the solution.Experimental results illustrate the superiority of the proposed method compared with other NMF approaches.(2)A new spectral-spatial weighted sparse nonnegative matrix factorization model is proposed.Currently,most unmixing algorithms based on NMF consider only the spectral information of hyperspectral data,and regarded the pixels as independent distributions when processing the data.In fact,from the spatial structure,there is a correlation between pixels,that is,the hyperspectral data also contains rich feature information in the spatial dimension.Motivated by the first law of geography,a recent trend is to incorporate the spatial information to improve the stability of decomposition,we develop a new blind hyperspectral unmixing method named spectral-spatial weighted sparse nonnegative matrix factorization.Based on the spectral weights mentioned in the previous chapter,the new method introduces spatial weighted and simultaneously acts on the abundance matrix.The existence of spatial weighted enables spatial information to be included in the unmixing model.Experiments proofed that the unmixing model which incorporates spatial information resulting abundance map is smoother and retains more detailed information.Compared with other advanced sparse unmixing algorithms,the proposed method has superiority.(3)A new Deep-Network sparse NMF model for hyperspectral unmixing is proposed.Due to the fact that most of the traditional NMF-based unmixing methods only consider the information in a single layer,even restrictions were added to the model,however,hierarchical features with hidden information was neglected.A single-layer decomposition structure fails to realize the multi-angle expression of the data features,and the unmixing results is restricted.Inspired by the deep learning,a new Deep-Network sparse NMF framework is proposed.The proposed method extends the single-layer of NMF model to deep network,using the deep model obtain the hierarchical features.Considering the rich spatial information,spectral information and sparseness of hyperspectral data,spectral-spatial weighted factors were enforced into the deep NMF model for promoting the sparsity of the result.The proposed method includes pretraining stage and fine-tuning stag two parts,where the former pretrains all factors layer by layer and the latter is used to reduce the total reconstruction error.Experiment results show that the proposed method is more accurate and robust than other methods.In this paper,we proposed three decomposition algorithms for mixed pixel of hyperspectral images,which can improve the application potential of hyperspectral remote sensing images and have important theoretical and application significance. |
