Aerosols: collaborative neural networks speed up PACE satellite measurements by 43 per cent and make them more accurate

The RemoTAP-Hybrid algorithm, applied to one year’s worth of data from the SPEXone polarimeter, produces the first global seasonal maps of aerosol composition as observed by this instrument.

Aerosols, particles suspended in the atmosphere, remain one of the major sources of uncertainty in climate models. Multi-angle polarimeters such as SPEXone, on board NASA's PACE mission, are among the best instruments to characterize them, but the complete physical inversion of their measurements is slow. A team from the Dutch institute SRON presents RemoTAP-Hybrid, which for the first time combines several collaborative neural networks with this physical inversion, on land and at sea.

Validated over one year of data against the AERONET network of ground-based photometers, the results improve: for the optical thickness of aerosols, the mean square error drops to 0.051 on land and 0.037 on the ocean, and about 78 to 81% of the measurements comply with the requirements of the Global Climate Observing System (GCOS). Compared to the current RemoTAP algorithm, the number of valid restitutions increases by 4% on land and 3% at sea, and the computation time decreases by 43% and 32%.

The team derives the first global seasonal maps of the optical properties, size and chemical composition of aerosols from SPEXone, and notes that adding a mode of fine non-spherical particles could further refine the results.