Revealing Climate Change from Moon Images with Dimensionality Reduction and Machine Learning

Earth’s albedo is an important factor governing our planet’s energy balance, and accurate measurements can help increase our understanding of the climate system as such, and specifically constrain global climate models. Terrestrial albedo is proportional to the intensity of earthshine, which is the solar radiation reflected off Earth onto the moon. Analyzing pictures of the moon can reveal the Earth’s albedo at a lower cost and without the same systematic errors as the classical approach based on satellite images. Our proof-of-concept analysis is conducted in a model-world where realistic lunar images are generated by a geometric model of sunlight scattering between the Earth and the Moon. The analysis presented here is based on 1.2 million synthetically generated images of the Moon at the same phase and, as a consequence the dataset has a low variance suitable for our initial investigations. A random forest machine learning model was trained on the dimension-reduced dataset to predict the albedo and reached the target range of ≤0.3% relative error rate. Several machine learning models and feature extraction algorithms were explored to identify the best possible configuration for this aim and compared in this work.