pith. sign in

arxiv: 2510.12328 · v6 · pith:URPUUL6Tnew · submitted 2025-10-14 · 💻 cs.LG

Leveraging Teleconnections with Physics-Informed Graph Attention Networks for Long-Range Extreme Rainfall Forecasting in Thailand

Kiattikun Chobtham , Kanoksri Sarinnapakorn , Kritanai Torsri , Prattana Deeprasertkul , Jirawan Kamma This is my paper

Pith reviewed 2026-05-25 07:50 UTC · model grok-4.3

classification 💻 cs.LG
keywords graph attention networksphysics-informed neural networksextreme rainfallteleconnectionsThailand rainfall forecastinggeneralized Pareto distributionlong-range forecasting
0
0 comments X

The pith

Physics-informed graph attention networks using teleconnections outperform baselines in long-range extreme rainfall forecasting for Thailand.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper introduces a graph neural network approach that models rainfall gauge stations as nodes in a graph to forecast extreme rainfall events over long ranges in Thailand. It incorporates physical knowledge through edge features based on orographic precipitation and uses climate indices to capture teleconnections between distant weather patterns. By combining graph attention with LSTM layers and a specialized extreme value distribution, the method addresses limitations in traditional machine learning for rare events. This matters because accurate long-term forecasts can support better water resource management in regions vulnerable to extremes. Experiments show improvements over standard baselines and the operational SEAS5 system in predicting extremes and generating high-resolution maps.

Core claim

The authors claim that their Attention-LSTM model, which applies attention using edge features from orographic-precipitation physics and processes embeddings with LSTM layers, when combined with Spatial Season-aware GPD for Peak-Over-Threshold mapping, achieves better performance in predicting gauge-station rainfall extremes across Thailand than well-established baselines, remains competitive with state-of-the-art methods, and improves upon the SEAS5 operational forecasting system for extreme events while enabling high-resolution mapping for decision-making.

What carries the argument

Physics-informed Graph Attention Network with LSTM (Attention-LSTM) that uses orographic-precipitation physics to initialize edge features for attention, followed by LSTM processing and Spatial Season-aware Generalized Pareto Distribution for extreme value mapping.

If this is right

  • The method outperforms well-established baselines across most regions including extreme-prone areas.
  • It remains strongly competitive with the state of the art.
  • It improves extreme-event prediction compared to the SEAS5 system.
  • It produces high-resolution maps supporting long-term water management decisions.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Adapting this graph-based teleconnection approach to other countries could enhance global extreme weather forecasting capabilities.
  • The explainability via teleconnections might help identify key climate drivers for policy.
  • Combining with more advanced physics could address potential non-stationarity issues in changing climates.
  • Deployment in operational settings may require validation on future unseen events to confirm robustness.

Load-bearing premise

That the chosen climate indices, the orographic-precipitation physics used to initialize edge features, and the graph topology of gauge stations together capture the dominant drivers of long-range extreme rainfall variability without substantial missing processes or non-stationarity.

What would settle it

A direct comparison on a held-out set of recent extreme rainfall events where the model does not show improvement over SEAS5 or baselines would falsify the performance claim.

Figures

Figures reproduced from arXiv: 2510.12328 by Jirawan Kamma, Kanoksri Sarinnapakorn, Kiattikun Chobtham, Kritanai Torsri, Prattana Deeprasertkul.

Figure 1
Figure 1. Figure 1: Simulation of the simple thermodynamic equation of Smith linear model 2.2 Recurrent Neural Networks (RNNs) and Graph Neural Networks (GNNs) Recent advancements in forecasting long-term climate variability have increasingly relied on ML and deep learning models. Recurrent Neural Networks (RNNs), LSTM, and Gated Recurrent Unit (GRU) were employed for sequence-to-sequence prediction (Hochreiter and Schmidhube… view at source ↗
Figure 2
Figure 2. Figure 2: The Attention-LSTM architecture 3.1 Dynamic Attention Coefficients Our novel approach introduces attention coefficients (weight matrices) of GATs into state￾of-the-art LSTM. Static edges (E) of graph structures will be constructed using the Pearson correlation as the statistical method, referred to as teleconnection. This served as feature selection for data-preprocessing in a ML process. The weight matrix… view at source ↗
read the original abstract

Accurate rainfall forecasting, particularly for extreme events, remains a significant challenge in climatology and the Earth system. This paper presents novel physics-informed Graph Neural Networks (GNNs) combined with extreme-value analysis techniques to improve gauge-station rainfall predictions across Thailand. The model leverages a graph-structured representation of gauge stations to capture complex spatiotemporal patterns, and it offers explainability through teleconnections. We preprocess relevant climate indices that potentially influence regional rainfall. The proposed Graph Attention Network with Long Short-Term Memory (Attention-LSTM) applies the attention mechanism using initial edge features derived from simple orographic-precipitation physics formulation. The embeddings are subsequently processed by LSTM layers. To address extremes, we perform Peak-Over-Threshold (POT) mapping using the novel Spatial Season-aware Generalized Pareto Distribution (GPD) method, which overcomes limitations of traditional machine-learning models. Experiments demonstrate that our method outperforms well-established baselines across most regions, including areas prone to extremes, and remains strongly competitive with the state of the art. Compared with the operational forecasting system SEAS5, our real-world application improves extreme-event prediction and offers a practical enhancement to produce high-resolution maps that support decision-making in long-term water management.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The manuscript presents a physics-informed Graph Attention Network with LSTM (Attention-LSTM) for long-range extreme rainfall forecasting at gauge stations across Thailand. Climate indices are preprocessed as node features; edge features are initialized from a simple orographic-precipitation physics formulation on a graph of stations; embeddings are processed by LSTM layers; extremes are handled via a novel Spatial Season-aware Generalized Pareto Distribution (GPD) applied after Peak-Over-Threshold mapping. The central claim is that the method outperforms well-established baselines across most regions (including extremes-prone areas) and remains competitive with the operational SEAS5 system while producing high-resolution maps useful for water management.

Significance. If the performance claims were supported by rigorous, quantitative validation, the integration of teleconnection indices, physics-derived edge features, and spatial extreme-value modeling on a gauge graph could represent a useful advance for long-range extreme rainfall prediction in monsoon regions. The graph-attention mechanism for explainability via teleconnections and the season-aware GPD are potentially valuable contributions, but the absence of any reported metrics prevents assessment of whether these elements deliver genuine gains.

major comments (2)
  1. [Abstract] Abstract (final paragraph): the claims that the method 'outperforms well-established baselines across most regions' and 'remains strongly competitive with the state of the art' (including SEAS5) are asserted without any quantitative metrics, station counts, cross-validation scheme, error bars, ablation results, or statistical tests. This absence is load-bearing for the central performance claim and prevents evaluation of whether gains are robust or artifacts of region/threshold selection.
  2. [Model/Experiments] Model and Experiments sections: the assertion that the chosen climate indices plus orographic edge initialization on the gauge-station graph capture the dominant long-range drivers is not tested against omitted processes (e.g., MJO, sub-seasonal ENSO modes, land-surface feedbacks, or non-stationarity). Without such verification, the claim that Attention-LSTM embeddings encode relevant teleconnections rests on an unverified completeness assumption that directly supports the outperformance narrative.
minor comments (2)
  1. [Abstract] Abstract: the novel 'Spatial Season-aware Generalized Pareto Distribution' is introduced without citation to prior spatial or seasonal extreme-value literature, leaving unclear what is genuinely new versus an incremental adaptation.
  2. [Abstract] Notation: the acronym 'POT' is used without prior expansion on first use.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the detailed review and constructive feedback. We address the two major comments point-by-point below, agreeing where revisions are warranted and providing clarifications on the model design choices.

read point-by-point responses

  1. Referee: [Abstract] Abstract (final paragraph): the claims that the method 'outperforms well-established baselines across most regions' and 'remains strongly competitive with the state of the art' (including SEAS5) are asserted without any quantitative metrics, station counts, cross-validation scheme, error bars, ablation results, or statistical tests. This absence is load-bearing for the central performance claim and prevents evaluation of whether gains are robust or artifacts of region/threshold selection.

    Authors: We agree that the abstract should be self-contained with key quantitative support. The Experiments section of the manuscript reports detailed metrics (including regional performance comparisons, station counts, cross-validation results, and comparisons to SEAS5), but these were not summarized in the abstract. In the revised version we will add concise quantitative statements to the abstract (e.g., average skill improvements, number of stations evaluated, and reference to statistical testing) while preserving length constraints. revision: yes

  2. Referee: [Model/Experiments] Model and Experiments sections: the assertion that the chosen climate indices plus orographic edge initialization on the gauge-station graph capture the dominant long-range drivers is not tested against omitted processes (e.g., MJO, sub-seasonal ENSO modes, land-surface feedbacks, or non-stationarity). Without such verification, the claim that Attention-LSTM embeddings encode relevant teleconnections rests on an unverified completeness assumption that directly supports the outperformance narrative.

    Authors: We acknowledge that exhaustive ablation against every possible omitted driver (MJO, sub-seasonal ENSO modes, land-surface feedbacks, non-stationarity) is not performed. The selected indices follow established literature for the Thai monsoon, and the orographic edge initialization follows a simple physics-based formulation. We will add an explicit Limitations subsection discussing these omitted processes and their potential impact, together with a statement that future extensions could incorporate additional drivers. The current performance gains are presented as empirical evidence rather than a claim of completeness. revision: partial

Circularity Check

0 steps flagged

No circularity: claims rest on experimental comparisons with independent baselines

full rationale

The paper describes a physics-informed Attention-LSTM GNN that initializes edge features from orographic-precipitation physics and climate indices, then evaluates extreme-event skill via POT mapping with a Spatial Season-aware GPD. No equations, derivations, or self-citation chains are present in the supplied text that reduce the reported outperformance (vs. baselines or SEAS5) to quantities defined by the model's own fitted parameters. The central claims are empirical and falsifiable against external operational forecasts; the chosen indices and graph topology are treated as modeling choices, not as self-defining the target skill. This is the normal non-circular case for an applied ML forecasting paper.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 1 invented entities

The central claim rests on domain assumptions about teleconnections and orographic physics plus one newly introduced statistical entity; no free parameters are explicitly listed in the abstract.

axioms (2)
  • domain assumption Teleconnections (preprocessed climate indices) exert predictable influence on Thai rainfall at seasonal lead times
    Invoked when the model ingests climate indices to capture long-range drivers
  • domain assumption Orographic-precipitation physics provides useful initial edge features for the graph of gauge stations
    Used to initialize the attention mechanism in the graph attention network
invented entities (1)
  • Spatial Season-aware Generalized Pareto Distribution (GPD) no independent evidence
    purpose: Peak-over-threshold modeling of extreme rainfall that accounts for spatial and seasonal variation
    Introduced as a novel method to overcome limitations of traditional machine-learning models for extremes

pith-pipeline@v0.9.0 · 5771 in / 1606 out tokens · 40438 ms · 2026-05-25T07:50:30.423103+00:00 · methodology

Review history (2 revisions) →

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

23 extracted references · 23 canonical work pages

  1. [1]

    provided insights into the statistical analysis of climate variables or climate indices, by examining historical trends and precipitation patterns at both global and regional scales in many regions. In the northern peninsular Malaysia area, ENSO exerts a strong influence on the timing and intensity of the monsoon (Moten et al., 2014), affecting global rai...

    work page 2014

  2. [2]

    was proposed to learn a spatiotemporal continuous-time process governed by atmospheric physics. Research recommendations emphasize that data-driven ML and traditional physical modeling can complement each other in predicting extreme events (Saha et al., 2024; Camps-Valls, 2023). However, the ML models remain limited to short-term predictions, and many mod...

    work page 2024

  3. [3]

    can help elucidate the relative contributions of various climate drivers and their interactions in modulating climate (Runge et al., 2019; Du et al., 2024), but a causal model must be represented by a directed acyclic graph and time-series prediction remains particularly challenging (Ravuri et al., 2021). In this work, we cast the long-term rainfall predi...

    work page 2019

  4. [4]

    (𝑘,l) = 𝐶!𝑖𝜎ℎ

    2 Related Work 2.1 Orographic Precipitation As mentioned in the introduction, various studies suggest integrating physics with data-driven ML. The simple thermodynamic equation of Smith linear model (Smith, 2003), governed by the physics of orographic precipitation, provides the spectral formulation that captures the basic structure of rainfall fields. Th...

    work page 2003

  5. [5]

    Recurrent Neural Networks (RNNs), LSTM, and Gated Recurrent Unit (GRU) were employed for sequence-to-sequence prediction (Hochreiter and Schmidhuber, 1997; Cho et al., 2014)

    Simulation of the simple thermodynamic equation of Smith linear model 2.2 Recurrent Neural Networks (RNNs) and Graph Neural Networks (GNNs) Recent advancements in forecasting long-term climate variability have increasingly relied on ML and deep learning models. Recurrent Neural Networks (RNNs), LSTM, and Gated Recurrent Unit (GRU) were employed for sequen...

    work page 1997

  6. [6]

    Recently, GNNs are a well-established neural network architecture designed for processing graph-structured data

    with LSTM (Wavelet-CNN-LSTM and Wavelet-LSTM) showed superior performance in monthly gauge-rainfall predictions compared to traditional models. Recently, GNNs are a well-established neural network architecture designed for processing graph-structured data. GNNs can take input in the form of a directed or undirected graph describing the connectivity struct...

    work page 2025

  7. [7]

    GCNs were formulated for undirected graphs, operating on the normalised graph Laplacian

    and LSTMs and proposed GC-LSTM to predict future links, achieving outstanding performance and outperforming state-of-the-art methods. GCNs were formulated for undirected graphs, operating on the normalised graph Laplacian. Their convolution is implemented using a localised first-order Chebyshev polynomial of the Laplacian. Pareja et al. (2020) proposed Ev...

    work page 2020

  8. [8]

    #$ !%&∗𝑌(𝑡−𝑖)+&𝛽!

    The Attention-LSTM architecture 3.1 Dynamic Attention Coefficients Our novel approach introduces attention coefficients (weight matrices) of GATs into state-of-the-art LSTM. Static edges (E) of graph structures will be constructed using the Pearson correlation as the statistical method, referred to as teleconnection. This served as feature selection for d...

    work page 1964

  9. [9]

    5 Dataset and Experimental Setup 5.1 Dataset and Graph Structure Modeling We use seven climate indices detailed in Table 2, processed from NOAA OISST (Huang et al

    Workflow of the operational long-range rainfall forecasts based on physics-informed Attention-LSTM and Spatial Season-aware GPD extreme value mapping. 5 Dataset and Experimental Setup 5.1 Dataset and Graph Structure Modeling We use seven climate indices detailed in Table 2, processed from NOAA OISST (Huang et al. 2021), together with historical rainfall d...

    work page 2021

  10. [10]

    The MJO can be characterized as an eastward-moving pulse of cloud and rainfall near the equator that typically recurs every 30 to 60 days

    in the South China Sea and Gulf of Thailand Real-time Multivariate MJO series (RMM1) The major fluctuation in tropical weather occurs on weekly to monthly timescales. The MJO can be characterized as an eastward-moving pulse of cloud and rainfall near the equator that typically recurs every 30 to 60 days. It is calculated using multivariate EOF (MV-EOF) of...

    work page 1981

  11. [11]

    (2024) for 1982–2024

    Climate features used in the study Following the spectral method of orographic precipitation (Subsection 2.1), we simulate orographic rainfall using each TMD gauge station elevation and ERA5 monthly average U and V wind components at 200 hPa, following the study of Saha et al. (2024) for 1982–2024. These serve as the Attention-LSTM physics-informed initia...

    work page 2024

  12. [12]

    The selection of climate indices is based on feature-selection criteria using Pearson correlation

    Each climate feature influences rainfall at individual stations through spatial teleconnections and temporal relationships. The selection of climate indices is based on feature-selection criteria using Pearson correlation. We construct a teleconnection when the average absolute correlation is higher than 0.4. We incorporate domain knowledge into the GNN e...

    work page 2016

  13. [13]

    of equation (13) evaluated with the Chi2 test. We consider a teleconnection present when the test p-value is below a significance threshold of 0.1, indicating that the lagged influence of the teleconnection improves each station-level rainfall predictability for each cluster in Thailand. Stations The clusters and High-Quality (HQ) stations Training, valid...

    work page 1982

  14. [14]

    We then present statistics of Spatial Season-aware GPD mapping in Subsection 6.2 and compare predictive skill with related work in Subsection 6.3

    Empirical Results Based on dataset and the experimental setup in Section 5, we report the graph-based model results in Subsection 6.1. We then present statistics of Spatial Season-aware GPD mapping in Subsection 6.2 and compare predictive skill with related work in Subsection 6.3. Finally, Subsection 6.4 presents operational forecasts evaluated against th...

    work page 1982

  15. [15]

    (𝑘,l) and (b) twelve-cluster areas for HII rainfall stations The spectral transform of precipitation 𝑃

    Reanalysis graph-based structure (a) the spectral transform of the precipitation 𝑃"(𝑘,l) and (b) twelve-cluster areas for HII rainfall stations The spectral transform of precipitation 𝑃"(𝑘,l) for each TMD rainfall station, simulated using ERA5 U and V wind compenents for 1982–2024, is shown in Figure 4a. A clear relationship between precipitation and high...

    work page 1982

  16. [16]

    model using Accuracy for each cluster of TMD stations. 6.4 Real-world Applications compared with ECMWF Seasonal Forecasts In ECMWF operational seasonal forecasting system from SEAS5, probabilities of precipitation anomalies or the mean of precipitation over 51 ensembles summarise the expected climate over the coming months from a statistical view of poten...

    work page 2025

  17. [17]

    Both predictions are initialized in March 2025 with a six-month lead time, covering March through August

    provides the highest average Accuracy across clusters (see details in Table 5). Both predictions are initialized in March 2025 with a six-month lead time, covering March through August

    work page 2025

  18. [18]

    For a preliminary assessment, we evaluated the 74 gauge stations as sample grid points

    Because the inputs and outputs of ECMWF SEAS5 and our proposed model are different, a direct comparison is not strictly fair. For a preliminary assessment, we evaluated the 74 gauge stations as sample grid points. For monthly RMSE at these stations for March–August 2025, Attention-LSTM (Q =

    work page 2025

  19. [19]

    (RMSE = 101.93 mm/month, average rainfall = 234.23 mm/month). 7 Conclusion and Future work This work presents an interpretable graph-based model for long-range rainfall forecasting at gauge stations across Thailand, along with the forecasting system for real-world application. The proposed Attention-LSTM is constructed using statistical feature selection ...

    work page doi:10.3389/fclim.2022.1031226 2023

  20. [20]

    Camps-Valls, G., Gerhardus, A., Ninad, U., Varando, G., Martius, G., Balaguer-Ballester, E., Vinuesa, R., Diaz, E., Zanna, L., & Runge, J. (2023). Discovering causal relations and equations from data. Physics Reports, 1044, 1–68. https://doi.org/10.1016/j.physrep.2023.10.005 Chang, C.-H., Johnson, N. C., & Yoo, C. (2021). Evaluation of subseasonal impacts...

    work page doi:10.1016/j.physrep.2023.10.005 2023

  21. [21]

    Saha, A., & Ravela, S. (2024). Statistical-physical adversarial learning from data and models for downscaling rainfall extremes. Journal of Advances in Modeling Earth Systems, 16, e2023MS003860. https://doi.org/10.1029/2023MS003860 Scarselli, F., Tsoi, A. C., & Hagenbuchner, M. (2018). The Vapnik–Chervonenkis dimension of graph and recursive neural networ...

    work page doi:10.1029/2023ms003860 2024

  22. [22]

    https://doi.org/10.3390/app15168854 Smith, R. B. (2003). A linear upslope-time-delay model for orographic precipitation. Journal of Hydrology, 282(1–4), 2–9. https://doi.org/10.1016/S0022-1694(03)00248-8 Sun, Q., Zhang, X., Zwiers, F., Westra, S., & Alexander, L. V. (2021). A Global, Continental, and Regional Analysis of Changes in Extreme Precipitation. ...

    work page doi:10.3390/app15168854 2003

  23. [23]

    https://doi.org/10.3390/w15162979 Waqas, M., Wannasingha Humphries, U., Chueasa, B., & Wangwongchai, A. (2025). Artificial intelligence and numerical weather prediction models: A technical survey. Natural Hazards Research, 5(2), 306–320. https://doi.org/10.1016/j.nhres.2024.11.004 Wu, T., Chen, F., & Wan, Y. (2018). Graph attention LSTM network: A new mod...

    work page doi:10.3390/w15162979 2025