Anchored Adaptive Probes with Scale-Aware Rendering for Atmospheric Super-Resolution

Arbitrary-scale spatial super-resolution of atmospheric fields aims to alleviate limitations in acquiring high-resolution data by enabling a single model to meet resolution requirements, including those at unseen scales, without scale-specific training. However, reconstruction performance may be constrained if models do not adequately account for region-specific detail recovery requirements and variations in the target scale. Jointly modeling the spatial allocation and scale-adaptive use of reference information can align the distribution of feature support with regional refinement needs while adapting its spatial range of influence to the target scale. We propose A${}^2$PSR, a reconstruction model that combines anchored adaptive probes with scale-aware rendering. The model treats input grid nodes as stable anchors that preserve basic information and generates probes carrying auxiliary information within grid cells on demand, forming the reference information needed for reconstruction. Scale-aware rendering then adaptively regulates the influence of this reference information to enable continuous-scale reconstruction. Experiments on ERA5 show that A${}^2$PSR performs well within the training scale range of $1$–$2\times$ and outperforms all evaluated arbitrary-scale methods at the tested extrapolation scales, including $22.5\times$, demonstrating good generalization across scales.