Albedo regression of the Greenland ice sheet

The project investigates the use of machine learning regression algorithms to estimate the broadband albedo factor on the Greenland Ice Sheet. The algorithms use data from the Ocean and Land Color Instrument (OLCI) on the Sentinel 3 Satellite and compare it to in-situ measurements taken from the PROMICE weather stations. Three different models and an extension are investigated: 1. LIN: A multiple linear regression model 2. KNAP: A 2nd order polynomial regression function derived by Knap et. al. (1999) 3. DNN: A multiple-perceptron deep neural network. 3.1 SMOOTH: An extension of the DNN using past and future estimations The LIN and DNN models have an average performance comparable to ESA’s state-of-the-art process SICE (Kokhanovsky et. al. 2019) developed by GEUS. The KNAP model proved to be less accurate and ultimately outdated. The main cause of the performance difference between the KNAP and LIN models is an increase in available spectral bands due to improved satellite sensors. The DNN model has a very small performance advantage over the LIN model, while the SMOOTH model makes a more significant increase in performance by reducing the variance of the DNN model. The study was done on a limited dataset with approximately 2400 samples from only 10 sites. The available features are 21 spectral narrowbands from the top of the atmosphere and the sun’s and satellite’s angles. When the LIN model was applied on a mosaic of Greenland it became clear that it could not handle the effect of angles well. The DNN model proved to do a much better job at this. When comparing the models to ESA’s state-of-the-art model SICE , the DNN model has comparable performance, while the SMOOTH model has a small advantage on average. Further, the DNN model is less restricted compared to the SICE model, as it does not build on various assumptions, and hence is applicable to more areas.