The Research Of Validation Of Remote Sensing Products Based On Pixel Unmixing

With the development of quantitative remote sensing technology, traditional ground-based observation can only obtain small region land surface parameter while large areas by remote sensing. Leaf Area Index(LAI) and Land Surface Temperature(LST) are two important land parameters, which also are widely used in global climate and environmental change. Both MODIS LAI and MODIS LST have an extensive application value, so developing their validation test research has a great significance.In this paper, the study area is the Heihe River Basin of Northwest of China, the experimental data include ground experimental data of ground-satellite synchronous observation experiment, TM remote image with high resolution and the MODIS LAI and MODIS LST product with low resolution. The empirical model of LAI inversion is established through experimental data, and the pixel unmixing model is established by modifying and improving traditional validation in remote sensing, and also compared with traditional testing methods. The research results are as follows:(1) In LAI product authenticity test, the LAI-NDVI empirical models established with the ground measured and NDVI value of TM images, despite some limitation, as physical model is universal but for the study area is applicable, and has a highly targeted, in coordinate with the inversion accuracy.(2) The error which use directly ground measured LAI(LST) or LAI(LST) retrieved from TM 30 m remote sensing image to verify MODIS LAI(LST) is very big, and the main reason caused the big error is the mixing pixel caused by surface spatial heterogeneity and scale effect duo to different spatial scales.(3) Based on the previous linear pixel unmixing study, the optimal solution of the problem solving algorithm is improved so as to satisfy a full-constrained linear pixel unmixing model and the accuracy compared with the ENVI software is higher, the correlation coefficient R2 between LAI and LST inversion results using the new algorithm compared with traditional algorithm and ground-based measurement is increased from 0.495, 0.793 to 0.985, 0.831 respectively, and the accuracy of inversion is greatly improved.(4) Taking advantage of the prior knowledge of high-resolution remote images to solve the problem of pixel unmixing in large-scale. The correlation coefficient R2 is increased from 0.7313 to 0.948 after using the method, the results show that using this kind of pixel unmixing can improve the accuracy and credibility of the process of validation of remote product.(5) The surface heterogeneity, the scale effect and pixel correspondence are analyzed in this paper which cause the uncertainty of validation in remote sensing. A new algorithm contains the advantages of all the traditional validation methods, and the correlation coefficient R2 is increased from 0.7098 to 0.966 after using the new method, which demonstrates that this method have obvious advantages compared with conventional methods, thus improving the accuracy and credibility of the validation in remote sensing.(6) The studies show that LST has a great change over time, due to the transit time of TM is different from MODIS LST product, there will be a big error if using directly the TM inversion result to test the MODIS product. Putting the LST retrieved from TM transformed to the same time scale of MODIS LST with the empirical model established with the LST and solar elevation angle, the correlation coefficient R2 is increased from 0.9103 to 0.968, which indicates that can effectively reducing the error of the validation after time correcting.