Abstract: The linear array push-broom optical satellite sensors are prone to vignetting and striping noise due to optical stitching and uneven sensor response, particularly under high dynamic range and low brightness conditions, where nonlinear effects become more pronounced. To address this issue, this paper proposes a relative radiometric calibration method based on an end-to-end multi-scale residual network. First, a high-quality sample set is constructed for training by selecting samples using a piecewise linear correction. Then, a multi-scale residual network is built, combining multi-scale feature extraction modules and skip connections to extract and integrate the features of vignetting and striping noise, and subsequently remove them from the original image. Experiments using GF1B multispectral images demonstrate that the proposed method effectively removes inter-frame vignetting and intra-frame striping. The striping coefficient decreases by 26.31% and 21.04% compared to traditional linear and piecewise linear methods, while the relative standard deviation decreases by 66.53% and 52.32%, respectively. Compared to statistical methods and deep learning denoising models, the proposed method maintains high accuracy and shows good generalization performance on GF1C and GF1D images.

Key words: relative radiometric correction, multi-scale residual network, feature fusion, vignetting, strip noise