acceptodds
Under review as a conference paper at ICLR 2027

Temporal Resolution Conditions the Global–Local Trade-off in Wildfire Spread Prediction

Abstract

Which inductive bias should a spatiotemporal operator have? We compare a gated global–local operator—spectral graph mixing fused per cell with local convolution—against convolutional recurrent baselines on 27 cross-state wildfire cases, and find that the answer depends on how far the fire front moves per model step. On the slowest fires of the main pool, under per-fire ROC-AUC, the operator ties the recurrent baseline at h=1 and leads it from h=3 on, the margin growing with horizon (+0.006, +0.047, +0.064 on a 6 h lattice); on the fastest it loses, and removing either branch, adding spectral or local capacity, or widening the training ring does not close that gap. A preregistered intervention that doubles per-step travel on the same fires by re-indexing them onto steps of twice the length moves the winner across families (+0.094 ROC-AUC, p=0.0022), so on these fires the dependence is interventional, not merely observed. The mechanism is a division of labor: the global path should predict for slow long-range spread and condition the local path on fast short-range steps, and neither role substitutes for the within-window local recurrence that carries fast regimes. The same ordering holds on the 419-fire WSTS benchmark, where every fire lies at the fast end. The ranking is a property of the (architecture, discretization) pair rather than of the architecture alone. All comparisons are graded verified, faithful, or historical under a protocol-recovery audit.

Primary area:
learning on time series and dynamical systems
Keywords:
Neural Operator, Spatiotemporal Forecasting, Dynamical Systems, Evaluation Methodology, Wildfire Spread Prediction

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_4DP8OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

To the best of my knowledge, this paper introduces a method based on two-branch neural networks for wildfire spread prediction. The paper also measures different models’ performance with different spread rate of wildfire and temporal resolutions.

Critical strengths

None

Critical weaknesses

This paper should be clearly rejected.

  • The paper is unclear and unpolished. It does not demonstrates the existence of the addressed problem of wildfire spread. The experiments are not understandable and there are many missing high-level explanations and details. I couldn't even understand the abstract. For me, this looks like a conclusion not an abstract. The contribution and introduction look like experimental results.
Questions for authors
  • What is the main aim of this paper? What is the contribution? Can you clarify.
  • L234-236: “We omit a transformer: three frames give temporal attention little to attend over, and an untuned spatial transformer adds capacity choices we cannot control at this sample size.” This is not true! What is meant here? Why can’t 3 frames work with transformers?
Reviewer_f22SOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies the influence of time discretization on the relative performance of different spatio-temporal inductive biases in wildfire spread prediction. The gated global-local architecture GLOW, which combines spectral graph blending and local convolution, was studied and compared with the convolutional loop baseline and other baselines under cross-state transition and autoregressive rollout evaluations. The main finding is that model rankings depend on the amount of fire spread per prediction step. GLOW performs favorably on slower fires in the main evaluation pool, while ConvLSTM performs better on faster fires. A controlled temporal-resolution intervention, which changes the same fires from 6-hour to 12-hour steps, further demonstrates that the relative advantage can shift between model families. The paper also includes branch ablations, training-objective analyses, and external dataset evaluations to investigate the observed behavior and its limitations.

Critical strengths
  1. The paper does not merely focus on improving the accuracy of predictions, but rather studies under what circumstances the biases of different models are effective. The discovery that model ranking relies on time discretization provides a useful perspective for model evaluation and selection in dynamic prediction tasks.
  2. The paper designs time-resolved interventions and extensive empirical analyses. A 6-hour and 12-hour discrete comparison of the same fire provided more compelling evidence than observing and comparing different fire events separately. Cross-state evaluation, branch ablation and protocol auditing have further strengthened empirical research.
Critical weaknesses
  1. The experimental results in the paper show that GLOW performs better on slow fires (Figure 5), but the results in the extended dataset (Table 3) provide relatively weak support for this finding, and the observed advantage is also smaller. I understand that these two groups of experiments adopted different fire samples and result summary methods, and the author has already explained this. However, considering that the performance differences of the model under different propagation states are one of the core contributions of this paper, I hope to see more evidence outside the main analysis set to support the conclusion of slow fires
Questions for authors

1.In Figure 5, the advantage of GLOW in slow fires is relatively obvious, but in the extended dataset (Table 3), this advantage is significantly weakened. I understand that these two groups of experiments used different fire samples and evaluation setups, which the author has also explained in Appendix B. Given this difference, the author believes that it is currently very certain that the advantages in slow fires can be extended beyond the main analysis set.

2.The paper holds that global spectral branching is more suitable for slow fires, while local recursive mechanisms are more important for rapid fires. However, the ablation experiment simultaneously altered other parts of the network architecture. Could the author further explain which experimental results most directly support this mechanism explanation?

1 comment

We thank R2 for the careful review. R2 found the two weakest points of our evidence. We made new measurements for both. The first result changes our claim.

Two problems in the submitted Table 3.

  1. Its pool (53 fires) contains all 27 main-pool fires. Only 26 fires are new.
  2. It used the historical GLOW checkpoints. Its caption says "same checkpoints as Table 1", but Table 1 uses the verified checkpoints.

New measurement. We evaluated the verified GLOW checkpoints with the Table 1 protocol on this pool and on a second pool of 49 fires. Both contain the 27 main-pool fires. Outside the main pool, they add 26 and 22 fires, with no overlap. We used the ConvLSTM checkpoints of Table 1. On the main-pool fires, the new GLOW scores agree with Table 1 (differences below 10⁻⁴ at h = 1, 3 and 5).

Result outside the main pool (Table 6 of the revised PDF). Of the 48 fires, 16 are scorable.

Stratum (cells/step)FiresΔGLOW wins
Slow (<2)5−0.030 (95% CI [−0.094, +0.049])1/5
Mid (2–4)9+0.0244/9
Fast (>4)2−0.0241/2
All16+0.001—

The relation between rate and Δ is r(log rate, Δ) = +0.06 (p = 0.82). One mid fire (AZ_2020_44) has a ConvLSTM score of 0.30 and Δ = +0.40. Without it, Δ for all fires is −0.026.

What this means.

  1. Outside the main pool, we do not replicate the slow-fire lead. This sample of five slow fires cannot exclude a small lead, but it does not support one.
  2. This test cannot examine the fast end. Of the six fast fires in the expanded pool, four are main-pool fires.
  3. On the WSTS benchmark (419 fires, daily steps), ConvLSTM beats GLOW on 324 fires (77.3%). This result has the historical grade, and its inputs also have the time feature of correction 4. Cells per step do not compare across the two datasets. The rebuilt dataset gives a second test (our reply to R1, Part 2).

Corrected table (Table 8 of the revised PDF; verified GLOW, positives-weighted, 27 scorable fires):

StratumFires (of which main pool)GLOWConvLSTMΔ
Slow (<2)9 (7)0.6870.659+0.028
Mid (2–4)12 (4)0.7610.727+0.034
Fast (>4)6 (4)0.6890.793−0.103

The slow and mid leads come from main-pool fires. On the new fires, Δ is −0.067 (2 slow fires, 0 wins), −0.004 (8 mid fires, 3 wins) and −0.023 (2 fast fires, 1 win). Table 8 shows the main-pool fires and the new fires in separate rows. Its weights differ from Table 2, which gives unweighted means.

Our revised claim. GLOW's score relative to ConvLSTM decreases as the spread rate increases. The main pool and the rebuilt dataset show this. A slow-fire lead of GLOW appears only in the main pool. We removed the general slow-fire claim from the abstract and from §1. Table 2 of the revised PDF compares the three sets of fires.

Reviewer_yJF9OpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This work studies a wildfire spread prediction problem, which is of great importance for many countries in the world, especially given the increase in frequency and intensity of such events. The scope of the paper lies in the study of different operators and their properties but it is very unclear. It introduces a global-local operator named GLOW and compares against some traditional baselines for the task like ConvLSTM and Fourier Neural Operators.

Critical strengths
  1. The research question seems interesting as a general problem.
  2. The specific task under study is highly important
Critical weaknesses

The paper is very poorly written, to a point that significantly impedes my ability to assess its scientific value. There are significant issues with the clarity and communication of this work.

Overall. The presentation of the paper is way below the bar for ICLR. As it stands I cannot truly assess the scientific quality from this paper. The paper needs full restructuring, better and more clear motivation, an introduction to the application and the problem formulation, introduction to the data used for this study, clear articulation of the methodology, better examination of the prior work, clear statement of limitations and articulation of the findings.

Below I list a subset of potential clarity/presentation issues:

Starting from the abstract, the authors start with a generic research question: "Which inductive bias should a spatiotemporal operator have?" and then proceed to describe an experiment for a very specific task i.e. Wildfire spread prediction, which can by no means answer the question they themselves pose. Moreover their sentences are, to me, impossible to follow. For example quoting from the abstract:

"Which inductive bias should a spatiotemporal operator have? We compare a gated global–local operator—spectral graph mixing fused per cell with local convolution—against convolutional recurrent baselines on 27 cross-state wildfire cases, and find that the answer depends on how far the fire front moves per model step. On the slowest fires of the main pool, under per-fire ROC-AUC, the op- erator ties the recurrent baseline at h=1 and leads it from h=3 on, the margin growing with horizon (+0.006, +0.047, +0.064 on a 6 h lattice); on the fastest it loses, and removing either branch, adding spectral or local capacity, or widening the training ring does not close that gap."

No build-up, no explanatation. What is a cell? What is h? What is the main pool? What is the data? What is gated global-local operator---spectral graph mixing fused per cell with local convolution? All these are incomprehensible. The main text follows this paradigm.

For example in L44-49:

What is a free-running rollout? What is a front? What is a cell per model step?

L51-53: What is measured per-step displacement? what is base-FNO (never defined). " global spectral path in isolation , degrades (0.736 →0.731 →0.662)" What do these numbers represent? What is the global spectral path in isolation?

L80-L82. What are the input data? What is the problem you tackle exactly? What are teacher-forced and free-running modes?

Figure 1 is also fully unclear, with a bolded title that conveys no information. The caption needs improvement.

The contribution list is unclear and raise more questions than provides answers. A contribution list should be a clear indicator of the findings of a paper. Here the titles of each bullets do not convey anything to the reader. For example:

  • Bullet 1: A regime-conditioned ranking on a unified spread-rate axis. -- Again what is a "main pool"?

  • Bullet 2: What does re-indexing mean? This is the first time this is discussed.

  • Bullet 3: "removing a branch changes which regime wins rather than uni- formly hurting performance (at h=5 the spectral-only arm matches the dual-branch model on two of three states); pure-focal training, widened rings, gate-capacity increases (normalization- confounded) and unconstrained fusion do not unlock fast-fire performance; "

    - What is a branch? This is the first time this is mentioned. 
    - What is a spectral arm?
    - What is the dual-branch model? 
    - What is pure focal training? 
    - Widening rings? gate-capacity? 
    - What is fast-fire? There has been no introduction to the actual application, and no real introduction to the motivation of this work.
    

L114-127 : "Across this literature the architectures are predominantly local, and compar- isons are made at a fixed spatiotemporal resolution under teacher-forced or single-horizon evalua- tion. We complement it by testing California-to-27-fire transfer with free-running feedback, and by treat- ing the temporal sampling interval itself as an experimental variable. The question is whether rank- ings remain stable when the protocol changes, not which model wins one leaderboard."

This is section 2 and we still don't know the actual problem, the input data, and the expected output.

Section 3.

L148. In fact this is the fist definition of a frontier set.

L151. F(x) and φ(b,e) are undefined. We can only guess what they represent.

L162 What does endogenous mean in this case?

Section 4.

L175 We discuss the input features to the proposed method but we still don't know what kind of input the method is expected to have.

L182: K1/K5 is undefined. We have no clue what they represent. In fact in L177-188 there are 3 different (probably) Ks. K_seq, Top-K and K1/K5.

L190 Uk is undefined as well.

Section 5.

This is the first time we discuss about the input data. Apparently we are dealing with a VIIRS dataset.

L316: What is WA and AZ maximum? OR? All undefined.

Questions for authors

The presentation of this work does not allow my to have meaningful questions.

3 comments

We thank R3 for the detailed list. We checked each point. The paper uses these terms before it defines them, and we accept this. R3 asks for definitions, a problem statement and a clear abstract. We give all three. Appendix A of the revised PDF has the full glossary. Below are the terms that R3 asked about.

TermDefinition
cellOne node of the 1 km grid.
front, frontier set C_tThe front is the boundary of the burned area. C_t holds the unburned cells with a burned cell in their 8-neighborhood. The loss and the metrics use only C_t.
taskFor each cell of C_t, the model gives the probability that it ignites within the next 12 h (two 6 h steps).
hThe forecast horizon in model steps. On the 6 h lattice, h = 5 is 30 h.
latticeThe time step of the data: 6 h in the original dataset, 12 h in the rebuilt dataset.
main pool, scorable fireThe 27 held-out fires of Table 1. A scorable fire has at least 10 positive cells at the scored horizon (15 main-pool fires at h = 5).
free-running rolloutThe model receives its own prediction, with a threshold of 0.5, as the next burn state. Few cells pass this threshold, so the burn input stays close to the last observed state.
teacher-forcedThe model receives the observed burn state at each step.
spread rate (cells per model step)The mean number of newly burned cells per step after ignition. Steps without growth count as zero. The submission called it measured per-step displacement.
per-step travelHow far the front moves in one model step. We measure it with the spread rate.
slow / mid / fast fireA spread rate below 2, from 2 to 4, and above 4 cells per 6 h step.
branchOne of the two feature paths in GLOW. The spectral branch mixes graph Laplacian modes. The local branch has two blocks of two 3×3 convolutions.
gate, gated global–local operatorThe gate is a scalar sigmoid in each cell that mixes the two branches. GLOW is this gated model.
dual-branch modelGLOW, with both branches and the gate.
spectral-only arm (no-local)GLOW without the local branch and the gate.
base-FNO, global spectral path in isolationFour spectral blocks on one frame. It has no GRU, no local branch and no gate.
pure focal trainingTraining with the focal loss only. We remove the curriculum, PDE, CRPS and perturbation terms.
widened ringTraining candidates up to Chebyshev distance 2. The evaluation uses distance 1.
gate capacityA two-layer MLP replaces the linear gate.
re-indexingThe 6 h → 12 h subsampling of the same fires. The revised text does not use this word.
K_win (K1, K5)The number of input frames: three by default, one or five in K1 and K5.
n_modes (submitted "top-K")The number of graph Laplacian modes, which is 32.
U (submitted U_K)The n_modes eigenvectors of the grid-graph Laplacian with the lowest eigenvalues.
x_t, φ, F_θx_t = φ(b_t, e_t) is the cell feature from the burn state b_t and the terrain and wind e_t. F_θ maps x_{t−2:t} to a burn probability for each cell.
endogenousProduced by the predictions of the model during rollout. We removed this word.
WA, OR, AZThe US states Washington, Oregon and Arizona, the three held-out regions.

The problem in four sentences (§1 and §3 of the revised PDF).

  1. Task. For each frontier cell of a 1 km grid, predict the probability that it ignites within the next 12 h. In rollout, repeat this prediction for h steps of 6 h.
  2. Input. The last three 6 h frames of each cell. Each frame has burn state, burned neighbors, terrain (elevation, slope, aspect), HRRR wind, cell position and a time index.
  3. Data. VIIRS active-fire detections on a 6 h lattice. We train on California fires. We test on 27 fires from Oregon, Washington and Arizona.
  4. Metric and comparison. ROC-AUC on the frontier set. We compare GLOW, ConvLSTM, logistic regression and a spectral graph baseline with one evaluation protocol.

A new abstract (186 words; also in the revised PDF):

We study how the spread rate of a fire and the time step of the data change the ranking of wildfire-spread models. The models use satellite fire detections on a 6 h lattice. For each 1 km cell at the fire front, they predict the probability that the cell ignites within the next 12 h. They repeat this prediction over several steps. We compare GLOW with ConvLSTM and two baselines on 27 held-out fires from three US states. GLOW fuses a spectral graph branch and a local convolutional branch. GLOW's score relative to ConvLSTM decreases as the spread rate increases. A rebuilt dataset with a 12 h lattice and 84 scorable fires shows the same direction, although both models have low skill at long horizons. Per fire, GLOW leads on the slow fires among the 27, but not on the slow fires among 16 further held-out fires. A coarser lattice for the same fires also shifts the comparison, but per-step travel alone does not explain the shift. On the rebuilt dataset, a convolutional recurrence inside the local branch reduces the fast-fire deficit, at a short-horizon cost.

Edits in the revised PDF.

R3 pointEdit
The abstract opens with a general question.The new abstract states the wildfire question in its first sentence and the task in the next two.
"cell", "h" and "main pool" are not defined.§1 defines them. Appendix A is a glossary.
The motivation and the problem come late (L80–82, §2).§1 gives the motivation, then the task, input, data and metric.
Free-running rollout, front, cells per model step (L44–49).§3.1 and §3.2 define them.
L51–53: base-FNO, "global spectral path in isolation", 0.736 → 0.731 → 0.662.These are the mean per-fire ROC-AUC of base-FNO in the slow, mid and fast thirds of the main pool. The submitted Figure 1 is now Figure 5, with a caption that says this. §4.2 defines base-FNO.
L175, L182: the input and the Ks.§3.1 lists all 13 features. We use three names: K_win (frames), n_modes (modes) and K_ens (ensemble members). Appendix A maps the older symbols.
U_k, F(x) and φ(b,e) are not defined (L151, L190).§3.1 defines F_θ and φ. §4.1 defines U.
L148: the frontier set appears late.§1 and §3.1 define it.
L162: "endogenous" is unclear.We removed the word.
VIIRS first appears in §5.§1 and §3.2 describe the data.
L316: "WA and AZ maximum".§1 defines the states. §5.1 now says that ConvLSTM has the highest score in every WA and AZ column.
Figure 1 is unclear.Figure 1 is new: the per-fire difference by spread rate on both datasets. Figure 2 now uses the branch names of the text and labels the gate.
The contribution list uses undefined terms.§1 gives a problem statement, the approach and three findings (our reply to R1).
The methodology needs a clear statement.§4 now has three parts: the model, the baselines and variants, and the training objective.
Prior work needs a better examination.§2 now also covers SimVP, Earthformer and SwinLSTM, next to the WSTS baselines (UTAE among them).
The limitations are not clear.§6 starts with a numbered list of six limitations.

New order of the paper. §1 Introduction: task, data, question and findings. §2 Related work. §3 Problem formulation and data. §4 Method. §5 Results: main comparison, rate dependence, change of lattice, mechanism controls. §6 Limitations and conclusion.

These edits change the structure of the paper. The revised PDF contains all of them.

Reviewer_yJF9OpenReview ↗

Response to authors revisions

I welcome the authors attempt to improve their manuscript. However, the initial state of the paper cannot be corrected in a rebuttal stage. Even if it was corrected it would require a completely new review. I encourage the authors to keep on improving their work and resubmit when ready. I maintain my rating.

Comments on the paper

1 comment

In these replies, R1, R2 and R3 mean Reviewer [4DP8], Reviewer [f22S] and Reviewer [yJF9].

We thank the three reviewers. This note lists the main changes. Each reviewer thread gives the details.

After the deadline, we audited our pipeline from end to end. The audit and the reviews showed several errors. We report all of them here, also the errors that no review mentions.

Revised PDF. We uploaded a revised PDF with a new structure:

  1. §1 states the task, the input, the data and the metric before the method, and lists the findings.
  2. §3 gives the problem formulation and the data. §4 gives the method and the baselines.
  3. §5 gives the results: the main comparison, the rate dependence on three sets of fires, the change of lattice and the mechanism controls.
  4. §6 lists the limitations and gives the conclusion.
  5. Appendix A is a glossary. Appendix B gives the new experiments and the corrections.

New measurements

  1. Attention in place of the GRU (R1). We replaced only the per-cell GRU of GLOW with per-cell self-attention. We used the GLOW training contract and three seeds.
  2. Further checks of this test (R1). We retrained attention with two other training settings. We also retrained GLOW in the environment of the attention runs.
  3. Frame and channel ablations (R1). At inference, we replaced the past frames, or only their burn channels, with copies of the latest frame.
  4. Fires outside the main pool (R2). We evaluated the verified GLOW checkpoints on 48 held-out fires that are not in the main pool.
  5. Rebuilt dataset. We rebuilt the dataset from the same satellite archive. It has a 12 h lattice, a causal time feature, detection filters and 96 held-out fires. On it, we trained GLOW and ConvLSTM with one training loop.

Corrections to the submitted paper

  1. L234–236 gave an untested reason. We tested it, and the reason is not correct. We removed the sentence (reply to R1).
  2. The submitted Table 3 uses the historical GLOW checkpoints, but its caption says "same checkpoints as Table 1". Its pool also contains all 27 main-pool fires. Table 8 of the revised PDF corrects it (reply to R2).
  3. The submitted §5.3 uses OR_2017_39 as a fast Oregon fire that ConvLSTM wins. With the verified checkpoints, GLOW wins it (0.807 against 0.774). Its rate (3.9 cells per step) is in the mid stratum. We removed this example.
  4. The time feature t/(T−1) uses the total duration of the fire. Forecasts cannot know this value. Alone, it gives a pooled ROC-AUC of 0.762 on WA at h = 5. A causal replacement at evaluation time keeps the signs of the per-fire strata. The rebuilt dataset does not use the feature. §6 lists it as the first limitation.
  5. The paper calls the 6 h → 12 h intervention "preregistered". We logged the prediction before the run, but we did not freeze the protocol. We removed the word for this intervention. Our reply to R2 gives its corrected statistics.
  6. The submitted §5.4 says that seven interventions "all failed" on fast fires. One of them, a ConvLSTM local branch, reduces the deficit (reply to R2).
  7. The satellite is Suomi-NPP, not NOAA-20.
  8. With the threshold of 0.5, few predicted cells enter the burn state in rollout. Thus the rollout input stays close to the last observed burn state. §3.1 now states this.

What we claim now

  • GLOW's score relative to ConvLSTM decreases as the spread rate increases. The main pool and the rebuilt dataset show this.
  • A slow-fire lead of GLOW appears only in the main pool. We removed the general slow-fire claim.
  • A coarser lattice for the same fires shifts the comparison. Per-step travel alone does not explain the shift.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.