A new automatic regression-based approach for relative radiometric normalization of multitemporal satellite imagery

Relative radiometric normalization (RRN) of multi-temporal satellite images minimizes the radiometric discrepancies between two images caused by inequalities in the acquisition conditions rather than changes in surface reflectance. In this paper, a new automatic RRN method was developed based on regression theory comprising the following techniques: Automatic detection of unchanged pixels, Histogram modeling of subject images, and Calculation of linear transformation coefficients for various categories of pixels according to their gray values in each band. The proposed method applies a new idea for unchanged pixels selection which increases the accuracy and automation level of the detection process. Also, a new idea is proposed for categorizing pixels according to their gray values. In this method, the number and interval of the categories are determined automatically and independently based on the histogram of subject images for each band. Thus, divergent influences of effective parameters such as atmosphere on different gray values are modeled. The method was implemented on two images taken by the TM sensor. Normalization results acquired by the proposed method were compared with the six conventional methods including: histogram matching, haze correction, minimum-maximum, mean-standard deviation, simple regression, no-change and modified regression using unchanged pixels. Experimental results confirmed the effectiveness of the proposed method in the automatic detection of unchanged pixels and minimizing any imaging condition effects (i.e., atmosphere and other effective parameters).