The methodology for upscaling surface albedo with consideration of surface heterogeneity

Pixel scale land surface truth value is essential for satellite product validation. Due to the heavy field observation burden, it is difficult to obtain large-scale field data. In recent years, machine learning has been widely used in upscaling processes due to its significant advantages in dealing with complex and nonlinear problems. However, surface heterogeneity, which may lead to significant errors in the prediction results, was seldom taken into consideration in the current machine learning upscaling methods. To address this issue, this study proposed a land surface upscaling model which incorporates land surface heterogeneity. A hybrid coefficient of variation (CV) index was adopted to depict the surface heterogeneity, and the random forest (RF) model was applied to upscale single-site in-situ albedo to the coarse pixel scale. The upscaled results were evaluated with pixel-scale albedo reference data. A high accuracy with R 2 of 0.93 and RMSE of 0.026 was obtained. To better understand the performance of the proposed model, 20km × 20km homogeneous and heterogeneous regions were also selected. Results show that models that take surface heterogeneity into account can improve the accuracy of predictions. In regions with low heterogeneity, whether or not surface heterogeneity was taken into account had less impact on the prediction results. However, in regions with high heterogeneity, the model considering surface heterogeneity better captured the spatial and temporal variations in albedo. This study proposed an effective albedo upscaling model with consideration of land surface heterogeneity, which can also be applied to the scale transformation of other land surface parameters.