Day 4

Detailed paper information

Back to list

Paper title A deep learning approach for methane plume detection from hyperspectral satellites
  1. Peter Joyce University of Leeds Speaker
  2. Hartmut Boesch University of Leicester
  3. Yahui Huang National Centre for Earth Observation - University of Leeds
  4. Emanuel Gloor National Centre for Earth Observation - University of Leeds
  5. Martyn P. Chipperfield University of Leeds, Leeds, UK
  6. Cristina Ruiz Villena University of Leicester
Form of presentation Poster
  • C1. AI and Data Analytics
    • C1.07 ML4Earth: Machine Learning for Earth Sciences
Abstract text Rapid identification and quantification of methane emissions from point sources such as leaking oil & gas facilities can enhance our ability to reduce emissions and mitigate greenhouse warming. Hyper and multispectral satellites like WorldView-3 (WV-3) and PRISMA offer very high spatial resolution of atmospheric methane concentrations from their short-wave infrared (SWIR) bands. However, there have been few efforts to automate methane plume detection from these satellite observations using machine learning approaches. Such approaches can not only allow more rapid detection of methane leaks but also have the potential of making plume detection more robust.

In this work, we trained a deep U-Net neural network to identify methane plumes from WV-3 and PRISMA radiance data. A deep residual neural network (ResNet) model was then trained to quantify the methane concentration and emission rate of the plume. The training data for the neural networks were obtained using the Large Eddy Simulation extension of the Weather Research and Forecasting model (WRF-LES). The WRF-LES simulations included an array of wind speeds, emission rates, and atmospheric conditions. The methane plumes obtained from these simulations were then embedded into a variety of WV-3 scenes, to compose the training dataset for the neural networks. The training data labels for the U-Net model were composed of binary mask images where plume concentrations above a certain threshold were differentiated from those below. The training data for the ResNet model consisted of a continuous scale of methane concentrations but were otherwise identical to that of the U-Net model. When evaluating the U-Net model on the test dataset, we found it to be significantly more accurate than the ‘shallow’ machine learning data clustering algorithm, DBSCAN. Furthermore, both trained neural networks provide predictions of satellite images almost instantaneously, whereas the DBSCAN method required a significant amount of human attention. Thus, our neural network models provide a considerable step forward in methane plume detection in terms of both accuracy and speed.

In this presentation, we will give an overview of the process of training our deep neural network models and justify the choices made regarding the architectures of the models. This will be followed by a demonstration of the effectiveness of the models in real-world images. Finally, we will discuss the potential future implementations of our approach. This work has been done by researchers from the National Centre for Earth Observation (NCEO) based at the University of Leicester, University of Leeds, and University of Edinburgh as part of a project funded by the UK Natural Environment Research Council (NERC).