A powerful Greenhouse Gas, Methane is one of the most significant drivers of global warming, and has an impact on climate change 84 times greater than CO2 (per unit mass, over twenty years). It is also a precursor to the formation of tropospheric ozone, which can have a number of indirect effects on agriculture productivity, crop yields and human health. As a result, it is one of the key targets of Sentinel-5P’s TROPOspheric Monitoring Instrument (TROPOMI).
Current retrieval algorithms are based on optimal estimation methods, which have been proven to perform well, but at large computational costs. One of the key bottlenecks are the forward Radiative Transfer Model (RTM) simulations, which are a common problem in EO. Emulation of RTMs using ML is gaining interest, due to the computational and timing gains that can be achieved. Implementing an emulator of the RTM does not require large processing overheads during operational use, allowing faster and less computationally demanding retrievals.
This research project focuses on using AI to learn the physics of RTMs for Greenhouse Gases, in particular for Methane retrievals from S5P-TROPOMI. Applying physics-based ML approaches can significantly reduce processing overheads during operational use, allowing for:
A Neural Network (NN) was trained on a synthetic spectra database of TROPOMI Shortwave InfraRed (SWIR) spectra, based on simulations from the LINTRAN RTM, implemented within the current Methane retrieval algorithm. Various NN architectures were tested to find the best performing emulator. The resulting NN-based emulator can effectively replicate spectra simulated by LINTRAN, explaining 99% of the variance and to within ~3% uncertainty, with a speed up up 10 000 times faster than the RTM. Additional effort is placed on uncertainty characterisation of the emulator, which is a key requirement for S5P-TROPOMI Greenhouse Gas products.
