Master Thesis Defense by Atlas Lundbergh Varsted

Title: Predicting albedo in Greenland using Machine Learning

Abstract:
Earth observation satellites such as Sentinel-3 are used to measure albedo continuously over Greenland. However, we can’t measure albedo through clouds using the Sentinel-3 satellite. We can use machine learning to fill in these gaps in the measured data. Using Gaussian processes with passive microwave background (PMB) data as the predictor, we show that machine learning can be used to reliably predict albedo values where measurements are not available. We get a good fit model for the prediction of albedo using a spacial approach. The model has some uncertainties in the results around the coastlines of Greenland. These are the areas in which we see the most changes in the ice and snow throughout the year. We also show that a temporal approach, while more limited, is potentially useful. We conclude that our models are effective at predicting albedo. The features of the models show that by combining spatial and temporal approaches to maximise data coverage for the predictor, the combined model can be greatly improved with smaller uncertainties. Extending the model in this way would require significantly more computing power and time, which was not available to us in this work.

 Supervisor: Aslak Grinsted
Censor: Sebastian Bjerregaard Simonsen (DTU)