Project:	Advancing Deep Learning Approaches Towards Operational Monitoring of Greenhouse Gas Plumes Observed by the EMIT Imaging Spectrometer (JPL AO No. 16185)
Disciplines:	Computer Science, Statistics / Earth Science / Geography
Mentor:	Brian Bue, (JPL), bbue@jpl.nasa.gov, Phone: (818) 354-2234
Background:	Methane and carbon dioxide are important Greenhouse Gases (GHGs), atmospheric concentrations are rising, and most emissions are from anthropogenic source categories with high uncertainties. A small percentage of methane and carbon dioxide point sources dominate emissions, so locating and quantifying emissions from these sources will increase understanding of anthropogenic emissions and provides efficient, cost-effective and impactful mitigation opportunities. For nearly a decade, we have been actively involved in developing and validating automated GHG plume detection systems driven by data captured by spaceborne imaging spectrometer platforms including EMIT and Carbon Mapper, and by airborne spectrometers such as the Next Generation Airborne Visible Infrared Imaging Spectrometer (AVIRIS-NG) and the Global Airborne Observatory (GAO). Our prior work demonstrated robust automated detection of methane plumes using a Convolutional Neural Network (CNN)-based detection system. Our current focus is to develop and validate a robust plume detection system using our CNN-based approach and/or comparable deep learning methods (e.g., pixelwise transformer models) aiming towards operational deployment for the EMIT mission. This demands thorough and rigorous analysis of the capabilities and limitations of the current/propose plume detection system on spatiotemporally diverse observations captured in a wide range of imaging conditions.
Description:	The overall goal for this project is to generalization performance of the methane plume detection system on spatiotemporally diverse observations collected by the EMIT imaging spectrometer, supported by preliminary EMIT results along with prior results from airborne AVIRIS-NG and GAO imaging campaigns. Students will contribute to the following efforts: • Train and evaluate performance of new and existing plume detectors on a large corpus of labeled and expert-curated plumes and background examples extracted from recent EMIT observations, and from prior AVIRIS-NG and GAO airborne methane imaging campaigns. • Conduct sensitivity studies to assess plume detectability with respect to (e.g.,) - Sensor-specific factors such as spatial resolution (i.e., spaceborne @ 30-60m pixel size, airborne @ 3-7m pixel size) and retrieval SnR, foreground vs. background clutter - Separability from known classes of false enhancements (e.g., image artifacts, confuser materials) - Source infrastructure type (e.g., oil/gas, landfills, agriculture, mining facilities) - Alternative GHG retrieval approaches versus established baselines. • Assess the impact of model / data / spatiotemporal bias and distribution drift using ~3 years of labeled, curated & quality controlled global spaceborne observations and airborne observations spanning nearly a decade of regional campaigns covering diverse locations throughout the United States + Canada. • Work closely with an interdisciplinary team of subject matter experts to interpret plume detection results on new data, assess out-of-sample generalization performance, identify and characterize novel methane emission sources and false enhancements, and guide data triage, curation and quality control efforts.
References:	PBS NewsHour Story: https://www.pbs.org/newshour/show/nasa-scientists-track-climate-changing-methane-leaks-from-the-air EMIT VISIONS Open Data Portal: Greenhouse Gases: https://earth.jpl.nasa.gov/emit/data/data-portal/Greenhouse-Gases/ Carbon Mapper Background: https://carbonmapper.org/about AVIRIS-NG GHG Mapping: https://avirisng.jpl.nasa.gov/greenhouse_gas_mapping.html California Methane Survey: https://www.nature.com/articles/s41586-019-1720-3
Student Requirements:	Required: experience training/evaluating CNN-based image segmentation / object detection models with python deep learning/ML/analytics libraries (e.g., pytorch, opencv, scikit-learn, scipy/numpy/pandas). Optional: familiarity with remote sensing image data and python-based geospatial analysis libraries (e.g., GDAL, rasterio, spectralpython, cartopy, geopandas)
Location / Safety:	Project building and/or room locations: . Student will need special safety training: .
Programs:	This AO can be done under the following programs:   Program    Available To        SURF@JPL    both Caltech and non-Caltech students  Click on a program name for program info and application requirements.