Aerosol retrieval switches to collaborative neural networks: SPEXone global products gain accuracy and cut computation by up to 43%

Global aerosol retrieval can now start from a collaborative neural network first guess and finish with a full physics inversion, bringing aerosol optical depth, Angstrom Exponent and fine-mode aerosol optical depth closer to ground observations over both land and ocean. The previous RemoTAP used a multimode look-up table for the first guess, which provided a good starting point only for aerosol optical depth and not for microphysical properties or absorbing aerosol. Against AERONET ground observations, the root-mean-square error for aerosol optical depth is 0.051 over land and 0.037 over ocean, with 78.3% and 80.8% of retrievals within the Global Climate Observing System requirement; the Angstrom Exponent error drops from 0.204 to 0.186 over land and from 0.242 to 0.207 over ocean, and fine-mode aerosol optical depth error drops from 0.060 to 0.047 over land and from 0.040 to 0.031 over ocean.

The first guess comes from three retrieval neural networks split by aerosol optical depth range, and a forward model neural network evaluates which result best matches the actual measurement, selecting the state vector that starts the full physics inversion. A single multilayer perceptron neural network had previously improved speed in airborne multi-angle polarimetric retrieval but could not demonstrate improved accuracy in the final products, while the multimode look-up table was limited by fixed refractive index and sphericity grids. Training sets were built by simulating 8 million samples with the RemoTAP forward model, separated into fine non-spherical, fine spherical, coarse non-spherical and coarse spherical mode-dominated subsets; the forward model neural network achieves a relative standard deviation below 1.2% for radiance and an absolute standard deviation below 0.0018 for degree of linear polarization over land.

The aerosol description is extended from a parametric three-mode to a four-mode definition with an added fine non-spherical mode representing insoluble mineral dust and organic matter in fine particles. The earlier three-mode description contained one fine spherical, one coarse non-spherical and one coarse spherical mode, with no explicit representation of fine non-spherical aerosol. An analysis of one year of global AERONET data shows that 21% of fine-mode scenes with aerosol optical depth above 0.05 are non-spherical; adding the mode raises the fraction within requirement for aerosol optical depth over ocean from 76.0% to 80.0% and lowers the Angstrom Exponent error from 0.242 to 0.218.

The work targets researchers using multi-angle polarimeter data and producers of operational aerosol products, with validation covering one year of SPEXone observations from March 2024 to February 2025 against AERONET ground stations. The method is also tested on POLDER-3 measurements, indicating that the collaborative neural network framework is not limited to a single instrument. The products are publicly available for the science community.

The collaborative neural networks have an intrinsic overestimation for small absorbing aerosol optical depth, which the full physics inversion mitigates. The fraction within requirement for fine-mode effective radius over ocean remains 47.1%, and the coarse-mode effective radius has a large root-mean-square error. The derived chemical component volumes depend strongly on the chosen real and imaginary refractive index values of the refractive index base species, which carry considerable uncertainty; for example, estimates of the brown carbon imaginary part in the blue part of the spectrum vary by a factor of five.