Diagnosing Precipitation with an Graph like Attention Model

Fengyang Xu, Xu Hui, Boyan Xin, Yutong Lu

Keywords: Weather forecast; Precipitation diagnose;

Abstract:

Accurate diagnosis of fine-scale precipitation patterns re-
mains a critical yet underexplored task in Earth system mod-
eling. Unlike sequence-to-sequence forecasting, spatial diag-
nostics demand high spatial fidelity, robustness to intermit-
tency, and interpretability—especially under extreme weather
conditions. In this study, we propose a deep learning frame-
work that operates entirely in the spatial domain, without re-
lying on temporal encoding. The model begins with a pseudo
spectral-space transformation to reorganize spatial modes,
followed by cluster-wise attention that enables localized,
structure-aware feature fusion. A sparse channel selection
module further improves compactness and generalization. We
train and evaluate our method on ensemble forecast fields
and corresponding daily precipitation observations from 2010
to 2018, focusing on daily accumulated precipitation. Com-
pared to strong baselines including LetNet, UNet with SSIM
loss, FourCastNet, and SwinUnet, our method achieves con-
sistently superior performance across standard metrics such
as RMSE, R², Pearson correlation, Hanssen–Kuipers (HK)
score, and multi-threshold Equitable Threat Score (ETS). No-
tably, the model demonstrates improved skill in detecting
extreme precipitation events. Spatial and per-sample analy-
ses further confirm its accuracy and robustness. These re-
sults highlight the promise of structure-aware spatial models
for operational precipitation diagnostics and post-processing
tasks.

Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)

Submission Number: 10405