Paolo Pelucchi(He/Him)Universitat de València
Paolo Pelucchi is a PhD student in Remote Sensing at the University of Valencia. With a background in physics and climate science, Paolo is now researching probabilistic machine learning methods to improve the retrieval of aerosol properties from satellite and ground-based sensors. He is pursuing his PhD as part of iMIRACLI, a Marie Curie Innovative Training Network advancing the use of machine learning to uncover the impact of aerosol-cloud interactions on climate. |
 |
TalksTowards probabilistic aerosol retrievals with invertible neural networks24-Apr-24Satellite remote sensing is the primary source of global aerosol observations, providing essential data for understanding aerosol-climate interactions. To solve the inverse problem at the heart of the retrieval process, traditional algorithms must make simplifications, often introducing bias and neglecting uncertainty. In this study, we explore invertible neural networks (INNs) for retrieving aerosol optical depth (AOD) from top-of-atmosphere reflectance, leveraging their ability to handle under-determined inverse problems and quantify uncertainty. INNs model the forward and inverse processes simultaneously, and use additional random latent variables to recover full non-parametric posterior distributions for the retrievals. We train on synthetic datasets generated by combining radiative transfer model simulations and satellite products. The INNs successfully emulate the forward problem, and inversion results demonstrate accurate retrievals as well, with AOD RMSE aligned with the expected accuracy of the MODIS product. The retrieved posteriors give calibrated uncertainty estimates and offer further insights into retrieval quality. We test the INNs under diverse conditions, exploring their potential to enhance aerosol and climate studies. Challenges and next steps in applying INNs to real satellite observations are discussed |