Explainable Graph Neural Networks for Observation Impact Analysis in Atmospheric State Estimation

TL;DR

The paper tackles the problem of quantifying observation impact on atmospheric state estimation within NWP data assimilation by introducing a meteorological graph that fuses observation and NWP data. It constructs ego-centric $k$-hop subgraphs around NWP points, uses a self-supervised GCN with node feature reconstruction to learn context-aware embeddings, and then applies an MLP to predict current states, optimized with $L_{ssl}$ and $L_{reg}$. To interpret the model, it employs gradient-based explainability methods—Contrastive Saliency, Grad-CAM, and Layer-wise Relevance Propagation—to quantify and visualize observation influence. The approach yields improvements over fully connected baselines and non-self-supervised GNNs, while providing interpretable insights into observation-type importance and temporal dynamics, with potential practical benefits for observation prioritization in data assimilation.

Abstract

This paper investigates the impact of observations on atmospheric state estimation in weather forecasting systems using graph neural networks (GNNs) and explainability methods. We integrate observation and Numerical Weather Prediction (NWP) points into a meteorological graph, extracting $k$-hop subgraphs centered on NWP points. Self-supervised GNNs are employed to estimate the atmospheric state by aggregating data within these $k$-hop radii. The study applies gradient-based explainability methods to quantify the significance of different observations in the estimation process. Evaluated with data from 11 satellite and land-based observations, the results highlight the effectiveness of visualizing the importance of observation types, enhancing the understanding and optimization of observational data in weather forecasting.

Figures (2)

• Figure 1: The averaged impact of each observation type.
• Figure 2: Time series of observations impact.