Enhancing global aerosol retrieval from satellite data via deep learning with mutual information estimation

• A global aerosol optical depth retrieval framework accounting for covariate shift. • Feature extractor in framework can be flexibly replaced with existing networks. • Improved accuracy in pixels lacking corresponding in-situ observations. • Our method demonstrates outstanding performance at the spatial scale. Satellite-based data can provide continuous aerosol observations but suffer from significant uncertainties across various regions. Transfer learning improves model generalization, yet its application in atmospheric research remains limited. Here, we introduce an innovative framework for retrieving global aerosol optical depth (AOD) which named the A erosol domain- Ada ptive N etwork (AAdaN). The framework utilizes a neural network to estimate mutual information, and aligns spatial covariate shift via a transfer loss term. Then, we assess the retrieval potential in unknown scenarios using independent land cover type, and the proposed model demonstrates satisfactory results. The cross-validation shows strong agreement with in-situ measurements, both in sample-based and site-based evaluations. Specifically, the site-based ten-fold cross-validation of our AOD retrievals indicate that all accuracy metrics are satisfactory, with a Pearson correlation of 0.766 and a Root-Mean-Square Error of 0.118, and that about 76.05 % of the retrievals meet the expected error criteria [±(0.05 + 20 %)]. Additionally, the proposed AAdaN achieves stable, high-accuracy aerosol retrievals across various surface and atmospheric conditions, and can generate spatially continuous AOD distributions. This study significantly improves spatial generalization and offers valuable insights for future model development.