Discovering Spatial Correlations between Earth Observations in Global Atmospheric State Estimation by using Adaptive Graph Structure Learning
Hyeon-Ju Jeon, Jeon-Ho Kang, In-Hyuk Kwon, O-Joun Lee · 2025
Read Discovering Spatial Correlations between Earth Observations in Global Atmospheric State Estimation by using Adaptive Graph Structure Learning by Hyeon-Ju Jeon, Jeon-Ho Kang, In-Hyuk Kwon, O-Joun Lee (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2508-07659.txt.
Introduction
Hyeon-Ju Jeon and co-authors teach a forecast model to learn which far-apart observations sway each other. Their 2025 paper deals with weather prediction, where standard systems read past air states and fresh Earth observations at fixed grid points. Observations drift in place, so their spatial links shift in complex ways that strict formulas miss. A graph network with structure learning can find those links, yet loose learning adds too many edges and blurs the signal.
Their fix caps each node's links by distance and need, in a system named CloudNine-v2. On real East Asian air and observation data, it cut forecast error against rival graph models. The gain held even where the air churns most, the team reports.
Catalog record
| CREATOR | Hyeon-Ju Jeon, Jeon-Ho Kang, In-Hyuk Kwon, O-Joun Lee |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 35 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2508-07659
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Discovering Spatial Correlations between Earth Observations in Global Atmospheric State Estimation by using Adaptive Graph Structure Learning Hyeon-Ju Jeon, Jeon-Ho Kang, In-Hyuk Kwon, O-Joun Lee Abstract This study aims to discover spatial correlations between Earth observations and atmospheric states to improve the forecasting accuracy of global atmospheric state estimation, which are usually conducted using conventional numerical weather prediction (NWP) systems and is the beginning of weather forecasting. NWP systems predict future atmospheric states at fixed locations, which are called NWP grid points, by analyzing previous atmospheric states and newly acquired Earth observations without fixed locations. Thus, surrounding meteorological context and the changing locations of the observations make spatial correlations between atmospheric states and observations over time. To handle complicated spatial correlations, which change dynamically, we employ spatiotemporal graph neural networks (STGNNs) with structure learning. However, structure learning has an inherent limitation that this can cause structural information loss and over-smoothing problem by generating excessive edges. To solve this problem, we regulate edge sampling by adaptively determining node degrees and considering the spatial distances between NWP grid points and observations. We validated the effectiveness of the proposed method by using real-world atmospheric state and observation data from East Asia. Even in areas with high atmospheric variability, the proposed method outperformed existing STGNN models with and without structure learning. 1 Introduction Accurate weather prediction is essential for preventing disasters, managing resources, and ensuring public safety. It has a direct impact on numerous sectors of society ( Zhou et al. 2021 ; Wu et al. 2024 ) . Traditional numerical weather prediction (NWP) systems use data assimilation methods ( Kwon et al. 2018 ; Bonavita et al. 2015 ; Clayton et al. 2012 ) to combine observational data and physical model states. However, these methods are often inefficient. The assimilation process typically requires extensive preprocessing and labor-intensive quality control to address inconsistencies, missing values, and varying formats across observation types. Additionally, when observational information is projected onto the model’s initial state, much of the original signal from the raw measurements can be distorted or lost, particularly when the model’s resolution or physical assumptions do not align with real-world conditions. Graph neural networks (GNNs) have emerged as a powerful framework for representing unstructured and relational data in various scientific domains. In meteorological applications, GNNs have been used to model spatial dependencies among heterogeneously distributed ground-based observation stations ( Jeon et al. 2022 ; Jeon et al. 2024b ) . They have also been used for tasks such as NWP error correction ( Wu et al. 2024 ) and determining the impact of observations on the analysis state ( Jeon et al. 2024a ) . Furthermore, spatiotemporal GNNs (STGNNs) incorporate temporal dynamics, enabling sequence-to-sequence prediction of evolving physical systems ( Lam et al. 2023 ; Li et al. 2018 ) . However, most existing STGNN variants construct a pre-defined fixed graph structure shared across all time steps. This implicitly assumes a static observation topology. This assumption is particularly limiting for meteorological networks, where the coverage and sensor footprints of observation platforms, such as polar-orbiting or geostationary satellites, dynamically change over time. Consequently, it becomes challenging to use the original observations without any information loss when the observation configuration varies over time. Although some recent studies ( Wang et al. 2025 ; Wu et al. 2019b ) have explored dynamic graph modeling based on spatial proximity between observations, most approaches still rely on heuristically defined static graphs. Such static representations cannot capture the broad and dynamically changing influence radii characteristic of regions with high atmospheric variability, such as coastal or mountainous areas. As illustrated in Figure 1 , prediction errors in these regions are often significantly higher. Our empirical evaluation confirms that models relying on fixed adjacency matrices incur up to 30% higher MAE in high‐variability regions compared to low‐variability regions, underscoring their inability to capture dynamic local interactions. (a) Spatial distribution of nodes by variability. (b) Mean absolute error (MAE) by node group. Figure 1 : Comparison of node-level error statistics and spatial distribution. Mean absolute error (MAE) of predictions for four meteorological variables (V, U, T, and Q) across Asia, Europe, and North America, stratified by nodes with low and high temporal variability. The variability index (VI) for each node was computed as the standard deviation over 24 time steps. Nodes were categorized as “low variability” (bottom 25%) or “high variability” (top 25%) according to the VI distribution. Recent research ( Jiang et al. 2023 ) has been studied graph structure learning to address these limitations and more effectively capture dynamic, context-dependent interactions in spatiotemporal data. However, observation platforms are heterogeneous, including ground stations, radiosondes, ships and various satellite sensors. These platforms have different spatial distributions, sampling frequencies and observed variables. The spatial coverage and temporal availability of observations can fluctuate dramatically due to orbital patterns, sensor outages or meteorological phenomena. Furthermore, regional atmospheric variability, particularly in complex terrain such as coastal or mountainous areas, highlights the need for a graph structure that can adapt dynamically. To address these issues, this study aims to extend existing graph structure learning method, which often assumes node homogeneity and temporal stationarity, by proposing a novel framework capable of capturing the intrinsic characteristics of meteorological networks, such as multi-source heterogeneity, dynamically evolving spatial relationships and localized atmospheric variability. We propose an adaptive graph-structured learning framework that dynamically infers regional adjacency matrices from observation features (i.e., sensor measurements) and metadata such as geographic coordinates and sensor attributes. Unlike traditional methods, which rely on fixed or heuristically defined graph topologies, our approach dynamically infers a regional adjacency matrix for each prediction target (grid point) at every time step. This process utilizes observations, as well as metadata, including geographic coordinates and sensor types. Specifically, we construct a k -hop subgraph for each grid point that incorporates NWP state vectors, and heterogeneous observations from multiple platforms. Node features are projected into a unified embedding space through node-type-specific encoders. Then, we use a differentiable Gumbel-Softmax mechanism to learn the k most relevant edges for each node, selecting neighboring nodes adaptively based on feature similarity and spatial proximity. These dynamic adjacency matrices are then integrated into a spatiotemporal graph neural network (STGNN) architecture. A GNN-based encoder captures spatial dependencies and a GRU-based decoder aggregates temporal information over a sliding window. This design enables our model can flexibly accommodate local atmospheric variability and rapidly changing network topology, thereby achieving robust and accurate predictions in a variety of meteorological scenarios. As a result, the proposed model allows us to accommodate both feature similarities and graph structure, achieving significant improvements in high-variation regions without sacrificing generalization elsewhere. Our experimental results validate that the proposed model achieves accurate and robust integration of diverse observational data, leading to enhanced predictive performance in spatiotemporal weather forecasting. The main contributions of this work are as follows: • We propose a novel, adaptive framework for learning graph structures that can dynamically infer regional adjacency matrices for each prediction target. This framework integrates observation features and metadata to enable context-aware connectivity in meteorological networks. • We introduce a differentiable Gumbel-Softmax-based edge selection mechanism that identifies the most relevant neighboring nodes for each grid point, allowing the model to flexibly capture both feature similarity