Pseudoinvariant Feature Selection Using Multitemporal MAD for Optical Satellite Images

Pseudoinvariant features (PIFs) are ground objects with invariant or near-invariant reflectance during data acquisition. The extraction of PIFs from optical satellite images generally plays a crucial role in relative radiometric normalization (RRN) and landcover change detection. Previous studies extract PIFs from bitemporal images while can generally obtain satisfactory results. However, they do not fully consider the problem of inconsistent PIF selection caused by performing pairwise PIF selection on more than two images. To decrease this inconsistency problem, a novel method called multitemporal and multivariate alteration detection (MMAD) is proposed. This method is based on a weighted generalized canonical correlation analysis, which solves canonical coefficients for multivariable and multitemporal data, thereby resulting in consistent PIF selection and RRN. In addition, a new weighting scheme based on pixel similarity, image quality, and temporal coherence is introduced into MMAD to reduce the sensitivity of PIF selection to landcover changes and to stably distinguish PIFs from non-PIFs. Qualitative and quantitative analyses of several multitemporal images acquired by Satellite Pour l’Observation de la Terre 5 are conducted to evaluate the proposed method. Experimental results demonstrate the superiority of the proposed method over related methods in terms of extracted PIF quality and radiometric consistency, particularly for image sequences with considerable landcover changes.