FROM INFRARED TO PASSIVE MICROWAVE: DEEP LEARNING EMULATION WITH UNCERTAINTY DIAGNOSTICS

Petkovic^1,2,3, M. Arulraj¹, M. Orescanin⁴, S. Steckler⁴, P. Ortiz⁵, R. Ferraro¹, H. Meng⁶

¹University of Maryland, College Park, ²Earth System Science Interdisciplinary Center, ³Cooperative Institute for Satellite Earth System Studies, ⁴Naval Postgraduate School, Monterey, CA, ⁵USMC, ⁶NOAA/NESDIS/STAR

Satellite passive microwave (PMW) sensors provide critical information for precipitation, cloud, and hydrological retrievals, yet their observations are limited in spatial and temporal coverage due to the low-Earth-orbit platforms on which they are typically deployed. In contrast, geostationary InfraRed (IR) imagers offer virtually continuous, high-resolution coverage but do not match the physical sensitivity of PMW to hydrometeors and atmospheric water vapor. Bridging these two observational domains by emulating PMW radiances from IR measurements can reduce spatio-temporal sampling gaps and improve the readiness of satellite-derived products for operational use by providing an insight into their common information content.

This work explores Bayesian Deep Learning (BDL) for emulating microwave brightness temperatures from geostationary IR radiances, focusing on both predictive performance and uncertainty quantification. Over ocean, a BDL regression model was trained to map GOES-16 ABI observations to synthetic GPM Microwave Imager (GMI) brightness temperatures across multiple frequencies. Results show that while deterministic and Bayesian models achieve similar accuracy (2K to 6K, frequency dependent), BDL uniquely decomposes predictive uncertainty into epistemic and aleatoric components. Epistemic uncertainty decreases with additional training data, while aleatoric uncertainty remains tied to irreducible scene variability such as cloud structures, demonstrating how uncertainty diagnostics can guide data collection strategies and model refinement.

To extend applicability over land surfaces, where surface heterogeneity increases retrieval difficulty, transfer learning is employed by adapting the ocean-trained model to land conditions. This stage was performed using a deterministic framework. Nevertheless, the transferred model retained expected predictive capability, highlighting the potential of leveraging ocean-based training to improve performance in more complex land environments. Together, these results demonstrate that uncertainty-aware deep learning and deterministic transfer learning provide complementary pathways for advancing IR-to-PMW emulation. BDL offers interpretability and guidance over ocean training domains, while transfer learning enables expansion to land, supporting the development of more reliable and operationally relevant remote sensing algorithms.