Neural Networks Boost CO2 Retrieval for Climate Monitoring

Monitoring greenhouse gas emissions like CO2 and CH4 is crucial for understanding climate change, requiring frequent and widespread spaceborne measurements. However, current full-physics retrieval algorithms, which rely on computationally intensive line-by-line radiative transfer (RT) simulations, are too slow for high revisit and spatial coverage. This research introduces a feedforward multilayer perceptron (MLP) surrogate model designed to accurately and efficiently predict top-of-atmosphere radiances in the CO2 weak band. By using a combined mean absolute error loss on radiances and RT Jacobians, the MLP preserves both spectral accuracy and sensitivity to geophysical parameters. When coupled with the NanoCarb imaging interferometer's instrumental response, this MLP-based forward model offers a precise and efficient way to estimate CO2 concentrations. This promises to enable more effective monitoring of atmospheric emissions from low-cost satellite constellations.