Pseudoinvariant feature selection for cross-sensor optical satellite images

Processing of multitemporal satellite images generally suffers from uncertainties caused by differences in illumination and observation angles, as well as variation in atmospheric conditions. Moreover, satellite images acquired from different sensors contain not only the uncertainties but also disparate relative spectral response. Given that radiometric calibration and correction of satellite images are difficult without ground measurements during data acquisition, this study addresses pseudoinvariant feature selection for relative radiometric normalization (RRN) that minimizes the radiometric differences among images caused by atmospheric and spectral band inconsistencies during data acquisition. The key to a successful RRN is the selection of pseudoinvariant features (PIFs) among bitemporal images. To select PIFs, multivariate alteration detection (MAD) algorithm is adopted with kernel canonical correlation analysis (KCCA) instead of canonical correlation analysis (CCA). KCCA, which assumes that the relation between at-sensor radiance is spatially nonlinear, can obtain more appropriate PIFs for cross-sensor images than that of CCA, which assumes that the relation between the at-sensor radiances of bitemporal image is spatially linear. In addition, a regularization term is added to the optimization of KCCA to avoid trivial solutions and overfitting. Qualitative and quantitative analyses on bitemporal images acquired by Landsat-7 enhanced thematic mapper plus and Landsat-8 operational and imager sensors were conducted to evaluate the proposed method. The experimental results demonstrate the superiority of the proposed KCCA-based MAD to the CCA-based MAD in terms of PIF selection, particularly for images containing significant cloud covers.