HYBRID CANONICAL CORRELATION ANALYSIS AND REGRESSION FOR RADIOMETRIC NORMALIZATION OF CROSS-SENSOR SATELLITE IMAGES
Keywords: relative radiometric normalization, pseudo-invariant features, kernel canonical correlation analysis, nonlinear regression
Abstract. Relative radiometric normalization (RRN) minimizes radiometric differences among images caused by inconsistencies of acquisition condition. In this study, a cross-sensor RRN method is proposed for optical satellite images from Landsat 8 OLI (L8) and Landsat 7 ETM+ (L7) sensors. The data from these two sensors have different pixel depths. Therefore, a rescaling on the radiometry resolution is performed in the preprocessing. Then, multivariate alteration detection (MAD) based on kernel canonical correlation analysis (KCCA) is adopted, which is called KCCA-based MAD, to select pseudo-invariant features (PIFs). The process of RRN is performed by using polynomial regression with Gaussian weighted regression. In experiments, qualitative and quantitative analyses on images from different sensors are conducted. The experimental result demonstrates the superiority of the proposed nonlinear transformation, in terms of regression quality and radiometric consistency, compared with RRN using linear regression.