Big Data Analytics and Machine Learning Tools for Space and Earth Observation

Technological developments, such as the internet and wearable technology, generate vast amounts of data about individuals. Where previously data would be collected only to answer specific questions, data can now be collected and stored for potential use in the future, without the need to know what it will be used for. Thanks to data centres and cloud storage the problem of storing these data sets, no matter how large, is now a problem of cost rather than feasibility. As well as data on individuals, data is being collected about the earth; temperature, emissions, population and countless other measures are recorded throughout the day.

This project focuses on data collected by the Sentinel-5 Precursor satellite, and specifically global methane emissions. Currently, there is a retrieval algorithm in place to calculate the methane concentration from the satellite recordings. The primary result of this thesis, presented in chapter 5.2, is the training of a neural network which replicates the task of methane retrieval using the results from this algorithm as ground truth. The main advantage of this method over the retrieval algorithm is speed, the network is considerably faster and still falls within an acceptable level of accuracy.

The final result presented in this work is error correction. In order to improve some of the worse performing areas of the neural network that was trained, a small sample was error corrected. The system trained and designed improves the overall results of the network and could be expanded to further reduce errors.