Analysis And Application Of Hyperspectral Remote Sensing Data Fusion Method Based On Nonnegative Matrix Factorization

The trade-off exists between the spectral and spatial resolution due to physical limitations, data transfer requirements, SNR and some other practical reasons during the design of remote sensors. The hyperspectral remote sensing data have very high spectral resolution at the expense of loss of spatial resolution, while multispectral or panchromatic data with few spectral channels have high spatial resolution. In most cases, high spatial and spectral resolution data are not available by a single sensor. The hyerpspectral data are limited to application in material classification and recognition due to its low spatial resolution. Therefore, how to obtain data with high spatial and spectral resolution have become one of the hot research issues in the field of remote sensing. In this context, the paper researched the fusion of the hyperspectral and multispectral data, which aimed at obtaining fusion data with high spatial and spectral resolution by integrating both advantages of them. Data fusion methods based on Spectral mixture analysis can enhance the spatial resolution and entertain the spectral feature well at the same time. The paper combined linear spectral mixture model, sensor observation model and NMF(nonnegative matrix factorization) to derive NMF fusion algorithm which basedon optimization with projected gradient method. Firstly, VCA(vertex component analysis) was used to perform the factorization of hyperspectral data to obtain endmember spectra matrix; Secondly, NMF algorithm based on optimization with projected gradient method was used to update the spectral matrix and abundance matrix alternatively to obtain high spectral resolution endmember spectra matrix and high spatial resolution abundance matrix after constant iterations. Finally, fusion data with high spectral resolution and high spatial resolution were obtained by multiplying the two matrices. Sensor observation models of the data were built in the initialization matrix of each NMF unmixing procedure. The research was conducted with synthetic data and real data. The multisource remote sensing data fusion prototype system was designed and developed in end of paper. The main work and achievements of the paper were as followings:(1) Research with NMF problem based on projected gradient method were conducted. Based on linear spectral mixture model, sensor observation model and NMF algorithm, projected gradient methods were derived in the paper to update endmember spectral matrix and abundance matrix. The fusion algorithm had fast convergence speed and avoided the situation that zero elements in decomposition matrix were not updated again.(2) Performance analyses of NMF fusion method were conducted by combining experiments with synthetic data. Experiments with synthetic data downsampled from AVIRIS and HYDICE data were conducted by NMF fusion method and constrained least square method was used as a benchmark. The performances of the two methods were evaluated by quantitative index such as peak signal noise ratio, spectral angle error, root mean square error and universal image quality index. According to the analysis result, NMF fusion method improved the spatial resolution of all band data in hyperspectral data and preserved spectroscopic properties well.(3) Analyses on the performance of NMF fusion method were carried on when applied to experiments with real data. Fusion experiments with HSI hyperspectral and CCD multispectral data of HJ-1A satellite in four regions which had different land cover materials were conducted. Quantitative indexes such as information entropy and image definition were adopted to perform quantitative evaluation and principle components analysis method was used to perform visual evaluation. According to the experiments analysis, value of information entropy and image definition of fusion result were higher than that of original data and principle components of fusion result contained more information and its edge details were more sharper and the spectral information of different materials was well consistent with that of original data as well as, which showed that NMF method can be used to improve the spatial resolution on all wavelength regions and has little spectral distortion.(4) The multisource remote sensing data fusion prototype system were designed and developed. Based on Windows platform, mature.NET, ArcGIS Engine, ENVI/IDL and MATLAB technology were adopted to perform data preprocessing, fusion, evaluation on fusion result and vegetation index calculation functions in the fusion system.