A Case Study on Pixel-by-Pixel Radiometric Norrnalization Between Sentinel-2A and Landsat-8 OLI

Multi-source medium-to-high resolution remote sensing data differs in their spectral specifications and imaging process. To efficiently use the multi-source data and minimize the surface reflectance retrieval inconsistency from different sensors, it is a key step to apply the radiometric normalization before synthesizing these different datasets with similar spectral bands. Many traditional methods, such as pseudo invariant feature normalization (PIF), time invariant cluster (TIC), automatic scattergram-controlled regression (ASCR) and so on, usually determine the corresponding conversion coefficients of each similar band from different images mainly through the entire image statistics, without considering the different characterizations behaving in variety of the ground covers, so the conversion coefficients are the same for all pixels in a whole band. Unlikely, this paper compensated the differences of variety of ground covers, on the basis of the iteratively re-weighted multiple change detection transform (IR-MAD), proposed a pixel-by-pixel radiometric normalization method. According to matching model and transformation model between the multispectral image pixel spectral and the ground spectral library, the spectral response adjustment matching factor was used to complete the radiometric normalization of the image pairs As a case study, this method was applied to similar bands (visible-near infrared, VNIR) of Sentinel-2A (S2A) and Landsat-8 OLI (L8) data, acquired over quasi-simultaneously in Xianghe County, Hebei Province, on September 12, 2017. Compared with using traditional method IR-MAD alone, the method in this paper could effectively minimize the differences among the medium-to-high resolution surface reflectance images Before processing, the mean and standard deviation of reflectance image of the S2A in the VNIR bands was 0.0763±0.0518,0.0891±0.0481,0.0861±0.0609,0.2955±0.0693. While the L8 was 0.0530±0.0361, 0.0752±0.0371, 0.0757±0.0482, 0.2711±0.0645, all lower than S2A. After the normalization, the corresponding statistical value of S2A normalized image dropped to the value which was 0.0556±0.0375,0.0758±0.0370,0.0741±0.0419,0.2713±0.0610. But the result of only using IR-MAD was 0.0589±0.0390, 0.0774±0.0367, 0.0763±0.0463, 0.2719±0.0602. It indicates that the method proposed from us could achieve a more satisfactory normalization effect Next step in our research, we will further verify and improve this method for more images.