AI Based Data Fusion for Multimodal Three Dimensional Geospatial Methane Data Cube Generation

Identifying methane sources and fluxes require a combination of different sensing modalities, forward dispersion and inversion modeling. Many of the data sources are either spatially or temporally sparse requiring a data fusion/conflation to generate training data for AI models. We discuss our approach to combine Sentinel 5p data with EMIT hyperspectral data, Sentinel 2 NIR channel in combination with High Resolution Rapid Refresh weather data to generate a three dimensional data cube. We also augment the data cube with Lagrangian Dispersion Model simulations where all sensing modalities are converted to CH4 concentration. We use the concentration data to initialize Stochastic Time-Inverted Lagrangian Transport model to identify emission footprint and highlight the location of potential emission sources.

The 3D datacubes is built on open source SpatioTemporal Asset Catalogs (STAC)and openEO API to harmonize the spatial and temporal span of different data sources. We demonstrated the development and advantages of such datacube that can stream the data either for data reconstruction of partial observations as well as generate the data for CH4 source detection using self-supervised foundation models. We discuss the performance and use cases for the 3D methane cube for identifying emission sources in oil and gas basins of western USA.