acceptodds
Under review as a conference paper at ICLR 2027

Set Prediction for Next-Day Active Fire Forecasting

Abstract

Predicting where active-fire clusters will occur over the next 24 hours is a sparse localisation problem with a variable number of targets. We introduce WISP, a query-based framework that formulates next-day active-fire forecasting as point-set prediction rather than dense raster prediction. WISP combines 48 hours of meteorological and fire-history inputs with satellite-derived vegetation, land descriptors, and weather covariates over the prediction horizon to produce a confidence-ranked set of future fire-cluster centres. The model integrates multi-source spatio-temporal encoding with a query decoder trained through bipartite Hungarian matching. We use asymmetric classification–localisation weighting in matching and loss, and investigate how this weighting and query capacity affect localisation, ranking, and redundant predictions. We construct a globally distributed benchmark spanning 2014–2024, with targets derived from the next-day union of VIIRS active-fire observations on a nominal 375 m reference grid. Under ERA5 reanalysis-conditioned evaluation, experiments reveal a trade-off between cluster recall and prediction-set compactness, with continuing fire activity easier to predict than newly observed activity. These findings support set prediction as a viable formulation for next-day active-fire cluster forecasting and highlight the importance of evaluating recall alongside false alarms and duplicate predictions.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Wildfire forecasting, Earth observation, set prediction, spatio-temporal forecasting, sparse event 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_jqV6OpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes WISP, which formulates next‑day wildfire forecasting as a sparse set‑prediction task. It constructs WISP‑benchmark, a global high‑resolution wildfire dataset, and predicts fire‑cluster centers via a query‑based decoder together with Hungarian matching. The model achieves decent performance even without future meteorological inputs. However, predicting newly ignited fires with no historical fire signals remains a major challenge.

Critical strengths

This work reformulates wildfire forecasting as DETR‑style sparse set prediction. Hungarian matching naturally accommodates the sparse, unordered nature of fire events and mitigates severe foreground‑background class imbalance common in pixel‑wise dense prediction. Its proposed gated multi‑scale condition injection flexibly fuses heterogeneous multi‑source information including historical observations, terrain, fire records and future weather, yielding competitive cluster recall and localization performance on global high‑resolution wildfire forecasting.

Critical weaknesses
  1. The paper only states that the set‑prediction model outputs cluster centers, yet it does not analyze what practical pain points discrete predicted points can solve that dense risk maps cannot for end‑users. Real‑world wildfires are areal burning zones; predicting cluster centers is merely an engineering compromise rather than representing physically complete wildfires. This inherent limitation of the task is under‑discussed in the motivation section.
  2. Although the ablation variant WISP‑OP is provided to exclude future weather inputs, experiments with noise perturbations simulating real‑world forecast meteorology are missing. The paper targets next‑day operational wildfire forecasting, yet the core model is evaluated under ideal input conditions.
  3. What the gating modules learn after training remains unexplored. The activation distributions of gates for different input sources (meteorology, terrain, fire history) are not statistically analyzed, so the dominant driving conditions cannot be identified. Only black‑box ablation is provided, showing performance drops without gates, with little insight into internal mechanisms.
  4. Baseline performance is highly sensitive to readout adapters for converting dense outputs into point predictions. It is ambiguous whether poor baseline results stem from intrinsic model limitations or sub‑optimal readout post‑processing. The unified readout code is not released, and ablations over alternative readout strategies are absent.
Questions for authors
  1. Insufficient justification of practical value: The paper only demonstrates the model’s ability to output fire‑cluster centers, without analyzing practical operational benefits compared with conventional dense risk maps. It downplays the fact that cluster‑center prediction is an engineering approximation and cannot capture full areal wildfire extent.
  2. Evaluation under ideal inputs, lacking real‑world robustness tests: The main WISP model is evaluated using ground‑truth future meteorology which is unavailable in practice. While WISP‑OP is presented, no noise‑perturbation experiments mimic real‑world forecast errors, leaving real‑world deployment performance unvalidated.
  3. Black‑box ablation for gating modules with limited mechanistic understanding: Only performance degradation upon gate removal is reported. Gate activations across meteorology, terrain and fire‑history inputs are not quantified, hurting model interpretability.
  4. Concerns about baseline‑comparison fairness: Dense baselines require external adapters to produce point‑format predictions. It is hard to disentangle model‑intrinsic weaknesses from artifacts introduced by post‑processing. No ablations across multiple readout strategies are conducted, and the relevant code is not publicly available.
Reviewer_AGUZOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes a DETR-like framework for detecting points that correspond to likely wildfire locations in the near future. Inputs are a range of weather (past observations and near future predictions) and land structure variables. A new formulation of the problem, predicting a set of points, is proposed and a new evaluation dataset is presented.

Critical strengths
  • An effective solution to the problem would likely have positive societal benefits.
  • The presentation is relatively clear, although I would have benefited from additional visuals to provide an overview of the WISP feature architecture.
  • The methods seem reasonable, although nothing sticks out as particularly novel and it's all very much focused on being tuned for this application (and this particular dataset).
Critical weaknesses
  • The main weakness is that this work is very narrowly targeted at a niche application and there doesn't seem to be much that would be of interest to anyone outside of this particular domain.
  • The predictions (Fig 2) of the baseline methods seem bad enough that it seems likely that the implementation didn't provide a truly fair setup. It seems possible that fairly small modifications to the inputs and structures of those methods might result in significant performance improvements. The strength of baselines is critical because this work is proposing a new evaluation dataset and even new evaluation task, that prior methods weren't tuned for. So, it calls into question how much real innovation there is in the methods or whether it's just weak baselines.
  • I remain unconvinced that the set prediction setup is the right one. It seems like for many applications you'd want to have per-pixel predictions and then process those to factor in various other constraints. I can see the motivation w.r.t. tip-and-cue for a higher fidelity satellite, but it's unclear how this integrates into such a workflow because it's focused on relatively small areas. How would you integrate these into a system that help task a satellite at a global scale? Are there new metrics or design criteria that would influence that?
  • L064: I don't fully understand the motivation for the "fixed budget of ranked fire-query hypotheses". My sense is that you'd want to set the budget high enough so that it always discovered all the likely fire locations. If it is a fixed budget due to satellite image tasking, then I think you'd want to integrate the evaluation metric with the constraints on tasking, potential risks, and information gain with the additional imagery.
  • L318: It's a minor issue, but using ERA5 reanalysis isn't a realistic operational choice. I under stand the limitation w.r.t. building a dataset with sufficient temporal coverage, but I'd want to see how such a model performs when using actual forecasts as opposed to post-hoc reanalysis weather data.
  • I would have liked to see more details on exactly how the baselines were trained.
Questions for authors

L343: What exactly do you mean by "without assuming identical inputs or training budgets"?

Reviewer_WFNGOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

WISP predicts next-day fire clusters as locations with confidence scores. It combines historical weather, vegetation, land information, and fire observations. A matching objective trains the predicted locations against observed cluster centres.

The paper also introduces a benchmark with samples from worldwide regions. The results suggest useful cluster recovery, including without future weather. However, missing evaluation details limit confidence in the reported advantage.

Critical strengths
  1. The prediction format fits the task.

A set of locations is a sensible representation for sparse fire clusters. The matching objective directly supervises the locations and their confidence scores. This gives the method a clear connection to its prediction target. [Section 4, pp. 4–6]

The benchmark could also support research beyond this model. It contains 4,085 samples spanning 2014–2024 across 14 GFED regions. Weather, vegetation, land information, and fire history provide several useful sources of evidence. [Section 3, p. 3]

  1. The results show gains without future weather.

WISP reaches 49.73% retained-cluster recall, compared with 41.89% for ClimaX. WISP-OP reaches 45.12% without weather inputs from the forecast period. It also exceeds the listed neural baselines on all five reported metrics. [Table 2, p. 7]

I give the WISP-OP result substantial weight. It shows that access to future weather does not explain all the reported gains. The authors clearly identify WISP’s future weather as reanalysis. They also acknowledge the need to test weather forecasts available at prediction time. [Section 5, pp. 6–8]

  1. The analysis exposes useful model behavior.

The component tests examine conditioning gates, auxiliary paths, temporal aggregation, the backbone, and interactions between queries. Removing auxiliary paths reduces weighted coverage from 54.28% to 33.45%. These tests help explain which parts support recovery within the proposed model. [Table 3, p. 8]

The regime analysis separates new activity, continuing activity, and scenes without observed future fire. It openly reports difficult cases and false alarms. Tests with different query counts and additional examples provide further context. [Table 4, p. 9; Appendices G–I, pp. 19–23]

Critical weaknesses
  1. The data split remains the main concern.

The split description is my main sticking point. Section 3 describes source cubes spanning fifteen days and gives the final split sizes. Appendix E explains preprocessing, but does not state whether splits separate cubes, fire events, regions, or time periods. [Section 3, p. 3; Appendix E, pp. 16–18]

Related windows could contain the same fire or nearby stages of its development. Without grouping rules and overlap checks, I cannot judge how independent the test samples are. I do not regard leakage as established. The available split description leaves this risk unresolved.

The extraction rules and quality filters also need clearer descriptions. They determine which fire events and empty scenes enter the benchmark. This matters for interpreting both generalization and the reported false alarms.

  1. The comparisons need stronger controls and uncertainty estimates.

The paper explicitly allows different inputs and training budgets across methods. It also converts dense predictions into points using different extraction rules. Those choices can directly affect the location metrics. The description does not provide enough detail to assess their effect. [Section 5, pp. 6–8]

The DETR-head comparison is useful, but its architecture and training setup need a clearer definition. A comparison with common inputs, a shared backbone, and comparable tuning budgets would better isolate the design’s contribution. The authors appropriately acknowledge this attribution limit. [Table 3 and Section 5.1, p. 8]

The main result tables report point estimates without uncertainty intervals. I could not find results across several training seeds. This does not prove that the authors used one run. It does leave uncertainty about small differences and model rankings. [Tables 2–4 and 7]

  1. Recovery scores leave important forecast errors unresolved.

The main cluster recall considers at most the strongest Q target clusters. The default Q is ten. Top10Rec also focuses on the strongest ten targets. MassCov uses the strongest target mass for the matched count as its denominator. It does not use all target mass. [Appendix F, p. 19]

These measures describe useful aspects of a limited prediction budget. However, they need recall over all clusters and counts of excluded targets. Average precision gives another perspective: Hawkes scores 26.13%, compared with WISP’s 23.31%. The paper acknowledges this different ranking. [Table 2, p. 7]

The main recall tolerance is 5.25 km. A 375 m target grid does not itself establish 375 m location accuracy. Finer tolerances and a distribution of location errors would clarify precision. [Section 5, p. 7]

False alarms also deserve more attention. WISP produces 7.54 predictions per extinguished scene and 0.75 per quiescent scene. Both subsets lack observed future fire. Results across several thresholds would show the tradeoff between recovery and false alarms. [Table 4, p. 9]

Questions for authors

How did you separate source cubes and fire events across training, validation, and test sets? What inputs, tuning budgets, and point-extraction rules did each comparison method use? How does recall over all clusters change at finer distance tolerances? Can you report variation across training runs and false alarms at several thresholds?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.