Closing spatiotemporal gaps in hydrometeor retrievals: exploiting geostationary infrared observations via probabilistic deep learning

Doctoral thesis, 2026

Comprehensive characterization of clouds and precipitation is fundamental to advancing our understanding of weather and climate systems. However, uncertainties persist in key quantities such as the mass of frozen hydrometeors and surface precipitation. Specialized spaceborne instruments provide high-quality measurements, but their orbital characteristics restrict their spatiotemporal sampling. Geostationary satellites, in contrast, offer continuous monitoring over continental scales, yet they have historically been underexploited. Relying solely on geostationary radiances, this thesis employs neural networks trained against gold-standard satellite products to retrieve frozen hydrometeor masses, cloud probabilities, and surface precipitation. Given the severely ill-posed nature of this inversion problem, the networks output case-specific probabilistic information targeting the irreducible uncertainty of these retrievals.

The appended papers challenge traditional paradigms by demonstrating the value of infrared (IR) observations beyond cloud top properties. Results show that retrievals of frozen hydrometeor masses using only thermal IR channels closely align in performance with those incorporating all visible and IR channels. This alignment overcomes the daytime-only limitation of existing physics-based methods. Extending this approach, machine learning delivers skilful 2D and 3D retrievals of frozen hydrometeor masses and cloud probabilities from only a single IR channel. This capability enables creating the Chalmers Cloud Ice Climatology from multi-decadal records of globally harmonized geostationary observations. Parallel to cloud properties, a multispectral IR-based retrieval developed for Rain over Africa provides rainfall estimates that outperform established products in accuracy, resolution, and latency.

These contributions unlock new applications for geostationary observations, providing unprecedented spatial and temporal continuity of hydrometeor data. The developed retrievals facilitate new insights into clouds, reduce observational uncertainties, and help validate climate models. Moreover, they carry clear societal value by allowing for timely estimates required for risk mitigation. Efforts to leverage these new capabilities are already underway.