Non-parametric smoothing for relative radiometric correction on remotely sensed data

Digital change detection methods have been broadly divided into either pre-classification spectral change detection or post-classification change detection. Since all spectral change detection methods are based on pixel-wise operations, or scene-wise plus pixel-wise operations, accuracy in image registration and scene-to-scene radiometric normalization is more critical for these methods than for other methods. A wide range of algorithms has been developed to adjust linear models. This paper proposes an automated radiometric normalization process that automatically extracts the training dataset, and uses a non-parametric smoother to adjust a non-linear mapping in order to minimize the effects of the influences of radiometric differences on image interpretation and classification. In order to investigate how the proposed normalization improves the performance classification, and assess the effectiveness of this technique, we carried out classification experiments on three image sets, and compare their results.