AutoML for Methane Plume Detection

Julia’s PhD is on the topic of AutoML for Spatio-Temporal Earth Observation Datasets. She is focusing on the problem of detection of methane plumes, in close collaboration with the SRON methane analysis group.

Methane is a strong greenhouse gas emitted in both natural processes and anthropogenic sources, e.g., cow burps. Other sources include natural gas leaks and garbage dumps.

Mitigating methane emissions

Finding these sources is an important strategy to reduce methane emissions and combat climate change. We can do this by searching for methane plumes in methane measurements made by the TROPOMI instrument. You can label these images and train a neural network to classify them, but to do this accurately, you need additional information like the albedo, wind vector field and quality assurance metrics. In other words, you need data fusion.

Data Fusion

Even though many tasks in Earth Observation require multiple data sources, there is no clear procedure for fusing multimodal data, and no AutoML system exists for this task. Julia is working on creating a Neural Architecture Search system for end-to-end data fusion: no more manual feature extraction.

In future work, Julia wants to apply her work to other EO tasks that will help combat climate change or help society in different ways. During her PhD, she aims to combine data fusion with other ML disciplines like semi-supervised learning and spatio-temporal methods to help leverage the available satellite data to the fullest possible extent.