Using forest reflectance modelling to estimate albedo for narrow-view satellites

Abstract: Surface albedo estimations from remote sensing are invaluable for energy balance descrip- tions in climate research. The aim of this work is to evaluate the possibility of using Sentinel-2 satellite data to get the total albedo of a small, vegetated area. In this thesis, a simple machine learning model is created to convert simulated Sentinel-2 reflectance measurements to surface albedo. The Forest Reflectance and Transmission model is used to simulate the reflectance behaviour of a pine forest and a field vegetation stand. The angular behaviour was analysed, and it was found that the directional reflectances at near-nadir view angles are typically lower than the hemispherical reflectance, though this effect also depends on the solar angle. As a result, it is important to include angular modelling in the albedo estimations. Next, a direct- estimation and narrow-to-broadband linear regression model were trained with FRT simulated reflectance data. The ability to predict albedo from the FRT data was high, with r2 > 0.94 for the direct estimation and r2 > 0.999 for the narrow-to-broadband model. Finally, the models for the pine forest and field were compared and found to be distinctly different. In conclusion, it is important to include angular modelling in albedo estimations. The regression models presented in this thesis perform well for the simulated vegetation stands. Moreover, it is valuable to train separate models for different land use classes. With some further improvements, the regression models for Sentinel-2 data have potential for accurate evaluation of surface albedo at a fine spatial resolution.