Estimating Surface Albedo of Arctic Sea Ice Using an Ensemble Back-Propagation Neural Network: Toward a Better Consideration of Reflectance Anisotropy and Melt Ponds

Abstract

The surface albedo of Arctic sea ice is a critical variable in the Earth’s energy budget, and various Arctic sea ice albedo datasets have been derived from remote sensing data. However, the influences of reflectance anisotropy and melt ponds were not well considered in previous studies. To improve the estimation accuracy and calculation efficiency, we developed a method for estimating the surface albedo of Arctic sea ice using an ensemble back-propagation neural network (EBPNN) model. A bidirectional reflectance distribution function (BRDF)/albedo dataset of snow, white ice, and melt ponds was constructed using the asymptotic analytical radiative transfer (AART) model, and the nonlinear relationships between the surface reflectance of Moderate Resolution Imaging Spectroradiometer (MODIS) and surface albedo were established using the EBPNN model (prediction correlation coefficient (R) = 0.996, and root mean square error (RMSE) = 0.023). Our results are consistent with the in situ measurements and the Medium Resolution Imaging Spectrometer (MERIS) product (R = 0.916, RMSE = 0.085, and bias = 0.001), in which the reflectance anisotropy of the snow/ice and the influence of melt ponds were fully considered, and the calculation efficiency was significantly improved by using the EBPNN model. The results of this study provide new insights for estimating the Arctic sea ice albedo from satellite observations, and the proposed EBPNN method can be used to generate a long-term Arctic sea ice albedo dataset with a superior estimation accuracy and calculation efficiency.