Machine Learning on-board of Hyperspectral Satellites for Methane leak detection

In this talk we describe the current landscape of methods for methane detection in satellite images with machine learning and highlight existing limitations. We adapt modern machine learning architectures targeting low compute environments and data with high spectral dimension. In the end, we propose models which are more accurate and faster than existing machine learning approaches and classical methods. Finally, we also outline directions for next research directions and highlight where these models can be put for general use for processing multi- and hyper- spectral data on low compute devices. We expect a rise of this type of data in the application areas of agriculture, medical imaging and for example in processing data from recently announced multispectral mobile phone sensors.