Affected by the sensor itself, illumination, atmosphere, terrain and other factors, even if imaging the same region at the same time, the spectral characteristics of ground objects in different remote sensing images are also very different, and the surface parameters, ground object classification and target recognition results of the inversion are also different, which brings great uncertainty to quantitative analysis. The relative radiation correction effect of PIF, method is obvious and the operation is simple, and the accuracy of the effect depends greatly on the selection of the PIF point. The general relative radiometric correction methods are linearization correction without considering the nonlinear difference of multi-temporal images. At present, most radiation normalization methods assume that the transformation relation between images is linear, extract PIF points and establish radiation transformation model. In this paper, Kernel Canonical Correlation Analysis (KCCA) is used for the first time to normalize the radiation between multi-temporal hyperspectral images, which can greatly reduce the nonlinear difference in relative radiation correction. Based on the theory of nuclear canonical correlation analysis, the radiation normalization method of multi-temporal aerial hyperspectral images is proposed. The feature points of PIF are extracted in the nuclear projection space, and the nonlinear model is used for the radiation normalization of hyperspectral images, to improve the radiation normalization accuracy of multi-temporal hyperspectral images. Compared with Canonical Correlation Analysis (CCA), the number and precision of PIF point extraction can be significantly improved. This method can satisfy the radiation normalization between aerial hyperspectral multi-temporal images.