acceptodds
Withdrawn from ICLR 2027

Geospatial Diffusion-based Evolution Synthesis (GeoDES) for Augmentation of Location-Sparse Phenomena

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Geospatial synthesis models are designed to generate realistic environmental conditions, yet they often struggle to capture the detailed dynamics of Location-Sparse Phenomena (LSP): complex events like storms, wildfires and oil spills that occur sparsely and irregularly across the globe. Existing approaches frequently rely on structural simplifications, such as restrictive static regional boundaries or computationally expensive, overly smoothed global grids. Instead, we introduce *Geospatial Diffusion-based Evolution Synthesis (GeoDES)*, a phenomenon-centered spatiotemporal diffusion model. GeoDES' correlated noise schedule, non-autoregressive design and temporal inflative training paradigm serve to increase fidelity and decrease compute requirements for LSP. GeoDES efficiently synthesizes high-quality events to augment hazard datasets and stress-test downstream models. On extratropical cyclone generation tasks, GeoDES outperforms baselines by achieving % lower Peak Vorticity Error and % higher Synoptic Anomaly Correlation Coefficient. Additionally, evaluations on the xBD multi-disaster dataset demonstrate high fidelity, with lower Fréchet Inception Distance than the next-best method. GeoDES brings realistic synthesis to the broad class of LSP and offers a step towards robust geospatial data augmentation.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
geospatial, diffusion, synthesis, augmentation, pretraining, finetuning, low data, video, vision, weather, land, spatiotemporal

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_FbbaOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes GeoDes, a generative model that generates the evolution of Location-Sparse Phenomena (LSP) conditioned on an initial observation. The proposed approach decouples the 2D spatial training and 3D spatial-temporal training, bridged by temporal inflation for model parameter compatibility. Temporaly correlated noises are used in the diffusion forward process of 3D spatial-temporal training. Experiments cover extratropical cyclone synthesis and post-disaster satellite image synthesis.

Critical strengths
  1. GeoDES achieves lower peak vorticity error than the evaluated baselines and near-ideal high-frequency spectral ratio and frequency bias scores on cyclone synthesis.
  2. Clear organization of the evaluation section. Section 3 explicitly identifies four claims and evaluates them through corresponding experiments. This structure makes the supporting evidence easy to locate and read.
Critical weaknesses
  1. The task formulation is not clear in the introduction section. By just reading section 1, I actually cannot get what exact task the model addresses or how the problem relates to real life application. For example is it about weather forecasting? or scenario synthesis for data augmentations? Simply saying 'modeling LSP' is vague to people unfamiliar to this field like me. I think the authors could mention and explain the task a bit at the beginning of section 1.

  2. The methodology rationale is not clear. The introduction section emphasizes the sparisty issue of LSP. However, it is unclear how the proposed components in GeoDes specifically address the sparsity issue.

  3. For the change modeling task, GeoDes's prediction looks still far from the target frame. And the FIDs are also high. GeoDes has highest standard deviation here. In fact, is this result considered useful for the post-disaster prediction task?

  4. Several ablated variants still outperform many baselines on some metrics. This may indicate that the overall gains are due to the neural network architecture, parameter count, or even training hyperparameters, More controlled comparisons would help clarify this.

Questions for authors
  1. What are the \pm numbers in table 1? Are they the standard deviations across the runs? What varies between runs, the training seeds?
  2. Why only averaged across 2 runs in table 2? I expect the standard deviations (if they are) are not reliable in such case.
  3. What does 'sampling paradigm=storm' mean in table 3?
  4. Missing ablation - the use of correlated noise in the diffusion forward process of the 3D model.
Reviewer_dXyEOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes GeoDES, a spatiotemporal diffusion framework for synthesizing location-sparse geospatial phenomena. The method uses phenomenon-centered data representation, 2D spatial pretraining followed by 3D temporal inflation, and temporally correlated diffusion noise to generate event evolution from initial conditions. Experiments on extratropical cyclone synthesis and the xBD disaster dataset show improvements in several generative quality metrics over existing baselines.

Critical strengths
  1. The paper addresses a meaningful problem of synthesizing localized and evolving geospatial phenomena, where capturing fine-grained structures with limited data is challenging.

  2. The experimental results are promising, particularly on cyclone synthesis. The use of domain-specific metrics, such as vorticity error and spectral energy distributions, provides a useful evaluation beyond conventional pixel-level metrics.

Critical weaknesses
  1. Limited novelty and concerns about experimental fairness. The proposed framework mainly combines existing techniques, including 2D-to-3D temporal inflation, SNR-weighted diffusion training, and phenomenon-centered cropping. The temporally correlated noise appears to be one of the few distinctive components, but its effectiveness is not directly evaluated through ablations. Moreover, the comparisons with global weather models involve substantially different task settings, as GeoDES operates on storm-centered patches while global models must predict the full atmospheric state. The SVD baseline also uses different normalization and training strategies, making it difficult to isolate the benefits of the proposed architecture.

  2. Overstated claims and limited experimental scope. Although the paper introduces a general framework for location-sparse phenomena, the main spatiotemporal experiments focus on cyclone synthesis. The xBD experiment only involves pre- and post-disaster image generation, rather than modeling continuous event evolution. Furthermore, the paper emphasizes data augmentation, but no downstream experiments demonstrate whether the generated data actually improve predictive or recognition models. Therefore, the current evidence does not fully support the broad claims of general applicability and augmentation effectiveness.

Questions for authors
  1. Could the authors clarify how the future storm-centered crops are determined during inference? Since the training and evaluation data are aligned using storm center trajectories, how would the generated sequences be used without knowing the future storm locations?

  2. What is the specific contribution of the temporally correlated noise schedule? Are there results comparing the proposed setting with independent Gaussian noise, e.g., ρ=0\rho=0?

  3. How much of the improvement over SVD comes from the proposed architecture rather than differences in normalization and pretraining? Are comparisons under matched training and preprocessing settings available?

Reviewer_DMqkOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The authors introduce GeoDES, a pixel-space video diffusion model for generating compact (geo)spatio-temporal events from an initial condition. They coin such event 'Location-Sparse Phenomena'. The model operates on bounding boxes around the phenomenon of interest and is trained in three steps: (1) 2D spatial pretraining, (2) Make-A-Video-style temporal inflation, (3) 3D fine-tuning with temporally correlated noise and SNR-weighted loss. The authors evaluate their approach on North Atlantic extratropical cyclones obtained from ERA5 data and compare to existing global weather models (ClimaX, Aurora, CEF, CoDiCast). They also generate xBD pre/post-disaster imagery, comparing against Pix2Pix and Palette. Throughout the work, the authors claim superior performance over baselines.

Critical strengths

I want to highlight the following strengths of the submission:

  • The paper addresses an important topic, the modeling of extreme spatio-temporal events. This is especially important as, as the authors mention, models often oversmooth extreme events (outliers) - while those events get increasingly more common, especially with accelerating climate change.
  • The method seems well suited for the problem; While not particularly methodologically novel, the proposed approach nicely combines existing work to suit the problem.
  • While not presented super well (see criticism below), the paper is mostly well written and easy to follow.
Critical weaknesses

The submission has the following shortcomings, ordered roughly from, in my opinions, most to least severe:

  • Throughout the paper, there's a general lack of contextualization with existing work; LSP to me seem not like a new 'thing' or a term to coin. Its highly related to Lagrangian feature tracking and already well established in meteorology (the main experimental domain of the paper) under storm-centred composites and cyclone tracking (the authors cite MCMS themselves). Spatial statistics also has a 'rare events in space and time' angle that is relevant here but not discussed, including space-time point processes and scan statistics.
  • Similarly, in the 'Geospatial ML' literature section, the authors claim that 'Geospatial ML traditionally operates on fixed spatial grids'; this is wrong; Geospatial ML originates in Gaussian Processes and Kriging which both operate in continuous space. And while yes, many raster / discrete models exist and they are dominant especially in the climate / weather domain, there still a rich body of work on continuous models, from modern GPs, to location encoders to neural processes.
  • I have a weird feeling about the results, especially the Aurora results. Aurora should be highly skilful at 2-day lead times. And the authors even acknowledge there's test-period leakage. But a leaky model that scores near zero is very strange. Aurora expects surface fields plus five atmospheric variables on 13 pressure lvls. The authors say that baselines are run on ,sparse, five-variable atmospheric state', So how was zero-shot Aurora initialised with only five variables? Or am I missing something?
  • For the xBD results, the attached figure is not informative at all, the synthetics scenes do not look particularly realistic. I don't find synthetic image quality scores useful here either. Wouldn't a much better evaluation be to run damage detection models on synthetic post event imagery to see if their approach produces more realistic, damage scenes?
  • The results are presented confusingly; What are the numbers in Table 1? This should be explained in much more detail in the caption of the table.
  • Some of the text in the figures is way too small and not legible
  • Ablation effects are mostly within noise.
Questions for authors

Overall, while this paper is in a highly relevant application area (AI for weather and rare events), I found the presentation and the experiments lacking and I don't trust the evaluation suite of the current manuscript.

Please see questions listed in the shortcomings section; I am especially curious to hear the authors thoughts on my concerns regarding the experimental results.

Comments on the paper

1 comment

We thank the AC and the reviewers for their time and attention to GeoDES. We believe that GeoDES is an important step towards generating fine grained structures in geospatial phenomena and showcasing the severe inadequacies of preexisting approaches, such as weather foundation models, for this purpose. While we are grateful for the feedback we have received, we are withdrawing the manuscript and intend to revise according to reviewer feedback in the future.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.