Research On Multisource Remote Sensing Data Fusion Based On Tensor Decomposition

As an important carrier of the information of the land surface features,remote sensing data is being widely used on a variety of relevant fields,including modern agriculture,resource exploration,and environmental protection,etc.However,restricted by the hardware design of the satellite sensors and the cost of launching,there exists a contradiction among the spectral resolution,the temporal resolution,and the spatial resolution,which makes it impossible for a single satellite sensor to simultaneously provide remote sensing data with high spatial resolution,high temporal resolution,and high spectral resolution.Fortunately,remote sensing data fusion is an effective way to integrate the complementary information of the spatial,temporal and spectral domains of multi-source remote sensing data,to provide better data for relevant remote sensing applications and make up for the shortcomings of single-source remote sensing data.Therefore,this dissertation has carried out the research on the fusion methods of the multisource remote sensing data,including the optical reflectance images and land surface temperature(LST)images,based on the tensor theory.The main works in this dissertation are summarized as follows:(1)In order to effectively capture the high spatio-spectral-nonlocal similarities and the global spatial and spectral piecewise smoothness of the latent high spatial resolution(HSR)hyperspectral image(HSI),a new spatio-spectral fusion method based on global gradient sparse and nonlocal low-rank tensor decomposition model is proposed.Based on the statistics and analysis of several HSI datasets,the relationship between the upsampled low spatial resolution(LSR)HSI and latent HSR-HSI is formulated as a hyper-Laplacian-based term to get rid of the dependence on the unknown spatial blurring kernel.To effectively capture the high spatio-spectral-nonlocal similarities of the latent HSR-HSI,a novel nonlocal low-rank tensor decomposition model is proposed to model the 3-dimensional regular tensors constructed from the nonlocal similar HSR-HSI cubes.The global spatio-spectral total variation regularization is then adopted to ensure the global spatio-spectral piecewise smoothness of the reconstructed HSR-HSI from nonlocal low-rank cubes.The relationship between the HSR multispectral image(MSI)and latent HR-HSI is built using a tensor-based fidelity term to recover the spatial details.The proposed method is verified on the synthetic and real datasets collected by different sensors.The maximum peak signal-to-noise ratio of the proposed method on the synthetic datasets with uniform blur and Gaussian blur can reach 43.819 and 43.033,respectively,which are greater than that of other comparison methods.At the same time,the spectral curve generated by the proposed method on the real dataset is closer to the input LSRHSI.These results show that the proposed method can efficiently preserve the spatial and spectral information of fused images.(2)Aiming at the problem that current learning-based reflectance spatio-temporal fusion methods concentrate on predicting images only from spatial similarity and neglect spectral correlations of remote sensing images,a new tensor sparse representation-based method is proposed for spatio-temporal fusion.In this method,the spectral correlation and the spatial similarity of the nonlocal similar cubes are simultaneously exploited through the tensor-tensor product-based tensor sparse representation model to improve the prediction performance of the fusion model,especially for spectral perseveration.Furthermore,the semicoupled mapping prior knowledge of sparse coefficients across the high-and low-spatial resolution image spaces is exploited to improve the robustness and universal applicability of the fusion model.In addition,to fully capture additional prior spatial information,the proposed method provides a new method to determine the degradation relationship between the target HSR and LSR difference images with the help of the known HSR and LSR difference images.The proposed method was tested on real datasets at both the Coleambally Irrigation Area(CIA)study site and the Lower Gwydir Catchment(LGC)study site.The erreur relative globale adimensionnelle de synthèse(ERGAS)and spectral angle mapper(SAM)of the prediction result on January 12,2002 generated by the proposed method can reach 0.8514 and 1.6846,respectively,and the ERGAS and SAM of the prediction result on December 12,2004 generated by the proposed method can reach 1.2371 and 2.8202,respectively,which are smaller than those generated by other comparison methods.These experimental results show that the proposed method can not only improve the spatial resolution of the target images,but also retain the spectral information more effectively.(3)Aiming at the problem that the existing integrated spatio-temporal-spectral fusion methods neglect the overall internal structure relationship in high-dimensional remote sensing images,a spatio-temporal-spectral fusion method based on semicoupled sparse tensor factorization is proposed to generate synthesized frequent high-spectral and high-spatial resolution images by blending multisource observations.From the perspective of tensor,this method fully explores the differences and connections between the spatial,temporal and spectral characteristics of different remote sensing images,and it can accomplish different kinds of fusion tasks,including spatio-spectral fusion,spatiotemporal fusion,and the spatio-temporal-spectral fusion.Specifically,the proposed method regards the desired high spatio-temporal-spectral resolution images as a multidimensional tensor and formulates the fusion problem as the estimation of the core tensor and the dictionary along each mode.The high spectral correlation across the spectral domain and the high self-similarity(redundancy)features in the spatial and temporal domains are jointly exploited using the low dimensional and sparse core tensors.In addition,assuming that there is a semicoupled mapping relationship between different remote sensing images,the estimation of the core tensor and the dictionaries is formulated as a semicoupled sparse tensor factorization of available heterogeneous spatial,spectral and temporal remote sensing observations.The proposed method is complementary verified by accomplishing different fusion tasks on different satellite image datasets.Experimental results show that the proposed method can provide fine fusion results in different fusion experiments.(4)By considering the complementary advantages of the LST spatial downscaling model and the reflectance spatio-temporal fusion method,a new composite LST data fusion approach based on the spatio-temporal fusion algorithm and the spatial downscaling model is proposed.First,aiming at the problem that current LST spatial downscaling algorithms were proposed mainly by addressing the spatial variability of the relationship model between the LST and the scale factors while neglecting the temporal variability,a new LST spatial downscaling model based on the geographically and temporally weighted regression(GTWR)is proposed.The proposed GTWR-based model can fully capture both the spatial and temporal variabilities of the relationship model by using the temporally and geographically varying regression coefficients for improving the performance of the spatial downscaling model.Second,aiming at the problem that the traditional LST spatial downscaling models ignore the certain difference between the unknown HSR scale factors at prediction time-phase and the HSR scale factors at adjacent time-phase,a composite LST data fusion model is proposed by making full use of the complementary advantages of the proposed reflectance spatio-temporal fusion method based on tensor sparse representation and the proposed GTWR-based LST spatial downscaling model,and then the fusion of the LST image at prediction time-phase is truly realized.The proposed composite LST data fusion model was compared with the traditional spatio-temporal fusion methods on two regions with different spatial heterogeneity.The root mean square error(RMSE)and the mean absolute error(MAE)of the proposed composite LST data fusion model in Zhangye are 1.88℃ and 1.59℃,respectively,which are less than those produced by ESTARFM(RMSE = 2.45℃ and MAE = 1.80℃),and the RMSE and MAE of the proposed composite LST data fusion model in Beijing are 0.94℃ and 0.68℃,respectively,which are less than those produced by ESTARFM(RMSE = 0.96℃ and MAE = 0.72℃).These experimental results show that our model achieved better LST reconstruction results.