Relative radiometric normalization is one of the main means of producing surface reflectance products from high-definition optical satellite images， the relative radiometric normalization method based on Iteratively Reweighted multivariate alteration detection is currently the most widely used method for high-resolution optical satellite image. However， the classical IR-MAD method is only applicable when there is little change in features between the reference image and the image to be corrected， and the method fails when there is a large change in features. In this paper， an improved IR-MAD algorithm is proposed by introducing a pseudo-invariant region constraint mechanism. The main idea is to use a Gaussian mixture model to extract the invariant regions between images， thus improving the quality of the pseudo-invariant points， so that effective radiometric normalization results can still be output when the feature changes a lot. The experiment was conducted using three representative sets of images， with GF-2 as the image to be corrected and Sentinel-2 as the reference image. The results show that compared to the classical IR-MAD method，（1） The pseudo-invariant points extracted by our method are mainly concentrated in the invariant regions， and the spatial distribution is more reasonable； （2） In dataset 1 representing general changes， the average coefficient of determination of our method is increased by 8.8%， while in dataset 2 and 3 representing strong changes， the average coefficient of determination of our method is 95% and 89%， respectively， while the traditional method is lower than 50% or even negative； （3） The spectral reflectance information of typical features in the processed image by our method is closer to the reference image. Therefore， our method provides a new and effective approach for relative radiometric normalization of high-resolution optical satellite images in complex situations.