Automatic Normalization of Multi-temporal Satellite Imagery using Unchanged Pixels

In this research, a new automatic Relative Radiometric Normalization (RRN) method has been developed based on regression, including the following techniques:- Automatic detection of unchanged pixels,- Histogram modeling of subject images, and- Calculation of linear transformation coefficients for different categories of pixels according to their gray values in each band.The proposed method uses a new idea for unchanged pixels detection, which increases accuracy and automation level of the detection process. Also a new method has been proposed for categorizing of pixels according to their gray values. In this method, the number and interval of the categories are determined automatically based on histogram of subject images for each band independently. Thus, various effects of atmosphere and other effective parameters on different gray values can be modeled. This method has been implemented for two images taken by the TM sensor. Evaluation results show that the accuracy of change detection stage has improved from 95.15% to 96.14% as compared with the accuracy of conventional methods. Moreover, Root Mean Square Errors (RMSE) of the normalization has decreased about 26.44%, 50.13%, 38.54%, 5.27%, 3.75% and 3.31% respectively in the bands 1, 2, 3, 4, 5 and 7, as compared to the case where only one linear transformation is used for all area of an image. Compared to using default values for categorizing pixels and linear transformation for each category, the developed method in this research reduces RMSE of the normalization about 30.82%, 45.9%, 30.53%, 4.19%, 2.82% and 0.32% in the aforementioned bands.