Scale-dependent evaluation of precipitation retrieval models

Determine whether the deterministic and generative precipitation retrieval models preserve the correct scaling behavior of precipitation under spatial aggregation and whether their relative performance rankings change at coarser spatial scales such as 10 km or 25 km.

Background

The study evaluates all models at the native 1 km × 1 km retrieval resolution and compares deterministic architectures, conditional GANs, and diffusion models using pixel-wise, distributional, and spectral metrics. However, hydrological and climate applications often use spatially aggregated precipitation products at substantially coarser scales.

The authors identify the absence of a systematic spatial-aggregation analysis as an unresolved issue. Evaluating model behavior at scales such as 10 km or 25 km could reveal whether the models reproduce the correct scale dependence of precipitation variability and whether the performance ordering observed at 1 km remains valid after aggregation.

References

Scale-dependent evaluation would clarify whether models preserve the correct scaling behavior of precipitation and whether relative performance rankings change with aggregation.

— Generative and deterministic deep learning models comparison for fine-scale precipitation retrievals from infrared brightness temperature  (2610.09859 - Meignin et al., 7 Oct 2026) in Section Discussion