Natural objects are typically non-Lambertian, and their surface reflections have directionality. Therefore, changes in incident and observation angles can affect observation results. The bidirectional reflection distribution function (BRDF) is used to describe the directional reflection characteristics of objects. Currently, multiple BRDF-related products, such as MODIS (Moderate-resolution imaging spectroradiometer), POLDER (Polarization and directionality of the Earth’s reflectances), and MISR (Multi-angle imaging spectroradiometer) BRDF products, are available for selection. However, BRDF products, such as those aforementioned, have low spatial resolution and cannot be used for operations such as fine-grained vegetation parameter inversion and local climate scale research. Moreover, BRDF products with medium and high spatial resolutions are scarce. In recent years, with the rapid progress in medium- and high-resolution ground observation technologies, the number of medium- and high-spatial-resolution wide-field satellite remote sensing images has increased. For wide-field images, owing to their large width and field-of-view angle, the received surface reflectance is different and BRDF correction is required. Although wide-field satellite sensors provide high-resolution observational data, such data often comprise only single-angle observations. Consequently, BRDF/albedo inversion cannot easily be performed on multi-angle datasets. Therefore, for these types of satellite sensors, surface BRDF inversion can be performed by combining multi angle observation data from low-resolution satellite sensors or BRDF products. However, this may result in the problem of “mixed pixels” caused by the spatial scale differences between different resolution sensors. To address this issue, we investigate BRDF in medium- and high-resolution wide-field images. Based on the imaging characteristics of these wide-field images and using low-resolution sensors, we achieve BRDF kernel parameter inversion between multiple sensors and complete normalization correction of the wide-field images.

The image BRDF correction performed in this study employs atmospheric radiation correction; therefore, the atmospheric radiation correction of the image is first completed. The radiometric calibration coefficients are obtained to perform a radiometric correction of the HJ-2A/B satellite CCD (Charge coupled device) image, and the 6S radiative transfer model is used to establish a lookup table. Then, the atmospheric correction parameters are used to complete the atmospheric correction. Because the XML file of each scene image of the HJ-2A/B satellite CCD camera records only the illumination-observation geometry information of the image central pixel, the imaging angle information of each pixel is further analyzed and obtained using the satellite transit time, pixel latitude and longitude, and wide-field satellite imaging principles to realize pixel-by-pixel normalized BRDF correction of the image. Notably, different types of objects exhibit significant differences in their structural and optical characteristics, whereas similar objects have similar structural and optical characteristics. Assuming that the BRDFs of the same types of objects have similar shapes, the normalized difference water index (NDWI) and normalized difference vegetation index (NDVI) are used to classify different objects, and MODIS reflectance products are combined to obtain multi-angle observation data for different classifications. To address the problem of “mixed pixels” caused by the difference in resolution between CCD images and MODIS products, uniform pixels are selected for spatial scale matching using CCD images as the underlying surface, and only uniform pixel data are retained. Finally, the least squares method is used to invert the observed data to obtain the kernel parameters of different classifications. These parameters are substituted into the kernel-driven model and applied to the HJ-2A/B satellite CCD images to realize the BRDF normalization correction of the images.

This study focuses on two types of land features, the Dunhuang Gobi and farmland in Zhaodong City, Heilongjiang Province, as the research area. A fitting comparison analysis is conducted between the RossThick-LiSparseR (RTLSR) and RossThick Maignan-LiSparseR (RTMLSR) models using the measured BRDF dataset. The results (Figs. 11-14) show that both models fit the two features and that their BRDF shapes are almost identical. The main difference is that RTLSR exhibits a relatively smooth characterization of hotspots. For areas such as the Gobi and farmland, where the hotspot effect is not significant, the RTMLSR model overestimates the hotspot to a certain extent, whereas the RTLSR model is applicable. Subsequently, the RTLSR model is selected to normalize and correct the BRDF of the HJ-2A/B satellite CCD image in the observation direction of the subsatellite point. (Figs. 15-17 and Tables 4-5.) shows that the original image has improved clarity and details after atmospheric correction and that the overall visual effect is close to natural observations after further completion of the BRDF correction. Simultaneously, the reflectance of the BRDF correction has a smaller root mean square error (RMSE) than that of the atmospheric correction and is closer to the reflectance measured in the field , effectively reducing the impact of the BRDF effect.

This study uses MODIS reflectance products MOD09A1 and MYD09A1 as prior knowledge, combined with detailed information on the underlying surface provided by HJ-2A/B satellite CCD camera data, to achieve joint inversion of the kernel parameters of the BRDF kernel-driven model between multiple sensors. To address the problem of “mixed pixels” caused by spatial scale differences between MODIS products and HJ-2A/B satellite CCD cameras, we use spatial position matching to select uniform pixels. The models for different ground object types are selected based on the assumption that the same type of surface BRDFs have similar shapes. The RTLSR and RTMLSR models are compared using two types of ground-measured BRDF datasets. The results show that the RTLSR model applies to areas where the hotspot effect is not noticeable. Applying the RTLSR model to the CCD images of the environment-2 satellite to achieve BRDF normalization correction effectively reduces the influence of the BRDF effect, providing an important method and reference basis for the future application of kernel-driven models in wide-field images.