OPTIMIZATION AND VALIDATION OF MAIA AEROSOL RETRIEVAL USING POLDER OBSERVATIONS

Abstract

NASA’s Multi-Angle Imager for Aerosols (MAIA) is a satellite-borne multi-angle polarimeter (MAP) instrument that aims to investigate the human health effects caused by exposures to various types of particulate matter (PM) and is anticipated to launch in 2025. The Step-1 aerosol retrieval algorithm deployed by MAIA is refined and tested by applying the algorithm to the data acquired by the satellite-borne instrument—POLarization and Directionality of the Earth's Reflectances (POLDER), and the results are compared with the ground-based Aerosol Robotic Network (AERONET) and retrieval of Generalized Retrieval of Aerosol and Surface Properties (GRASP). To address the quilting effect observed across the edges of adjacent retrieval patches—each composed of multiple pixels that are retrieved simultaneously—a three-step strategy was implemented within the retrieval algorithm. As a significant part of this study, further advancement came with the development and integration of a neural network (NN) based radiative transfer and optics module, which has significantly accelerated the radiative transfer modeling process, increasing efficiency by three to four orders of magnitude without compromising modeling accuracy. The innovative use of automatic differentiation for the analytical Jacobian matrix calculation from the trained NNs has further accelerated computations and improved the precision of derivative estimations. Accuracy assessments over MAIA's Los Angeles primary target area have shown that aerosol optical depth (AOD) retrievals are comparable to those from AERONET and GRASP-HP. Compared to AERONET reference AODs, the RMSEs of MAIA-NN based retrieval are 0.152, 0.050, 0.036, and 0.033 for wavelengths 565, 670, and 1020 nm, respectively. In contrast, the RMSEs of GRASP-HP based retrieval are 0.084, 0.084, 0.081, and 0.082, respectively. The three-step strategy not only mitigates the visual quality of aerosol maps but also substantially improves retrieval performance by enforcing physically plausible constraints. As a result, there was a RMSE performance enhancement for MAIA-NN retrievals, with a notable improvement of 49%, 59%, and 62% at 565, 670, and 1020 nm wavelengths, despite a slight 4.8% decrease at 440 nm. Remarkably, the NN-based approach has proven to be more than 400 times faster than the traditional method, taking less than 1 second per pixel on average, indicating its transformative capability for near-real-time aerosol retrieval from MAP satellite data.