Radiometric Cross-Calibration of HJ-2A/CCD3 Using the Random Forest Algorithm and a Spectral Interpolation Convolution Method with Sentinel-2/MSI
Xiang Zhou,Yidan Chen,Yong Xie,Jie Han,Wen Shao
DOI: https://doi.org/10.3390/rs16224337
IF: 5
2024-01-01
Remote Sensing
Abstract:In the process of radiometric calibration, the corrections for bidirectional reflectance distribution functions (BRDFs) and spectral band adjustment factors (SBAFs) are crucial. Time-series MODIS images are commonly used to construct BRDFs by using the Ross–Li model in current research. However, the Ross–Li BRDF model is based on the linear relationship between the kernel models and is unable to take into account the nonlinear relationship between them. Furthermore, when using SBAF to account for spectral difference, a radiative transfer model is often used, but it requires many parameters to be set, which may introduce more errors and reduce the calibration accuracy. To address these issues, the random forest algorithm and a spectral interpolation convolution method using the Sentinel-2/multispectral instrument (MSI) are proposed in this study, in which the HuanJing-2A (HJ-2A)/charge-coupled device (CCD3) sensor is taken as an example, and the Dunhuang radiometric calibration site (DRCS) is used as a radiometric delivery platform. Firstly, a BRDF model by using the random forest algorithm of the DRCS is constructed using time-series MODIS images, which corrects the viewing geometry difference. Secondly, the BRDF correction coefficients, MSI reflectance, and relative spectral responses (RSRs) of CCD3 are used to correct the spectral differences. Finally, with the validation results, the maximum relative error between the calibration results of the proposed method and the official calibration coefficients (OCCs) published by the China Centre for Resources Satellite Data and Application (CRESDA) is 3.38%. When tested using the Baotou sandy site, the proposed method is better than the OCCs of the average relative errors calculated for all the bands except for the near-infrared (NIR) band, which has a larger error. Additionally, the effects of the light-matching method and the radiative transfer method, different approaches to constructing the BRDF model, using SBAF to account for spectral differences, different BRDF sources, as well as the imprecise viewing geometrical parameters, spectral interpolation method, and geometric positioning error, on the calibration results are analyzed. Results indicate that the cross-calibration coefficients obtained using the random forest algorithm and the proposed spectral interpolation method are more applicable to the CCD3; thus, they also account for the nonlinear relationships between the kernel models and reduce the error due to the radiative transfer model. The total uncertainty of the proposed method in all bands is less than 5.16%.