acceptodds
Under review as a conference paper at ICLR 2027

MorseDistill: Spatio-Temporal Morse-Guided Distillation for Efficient Wildfire Forecasting

Abstract

Wildfires have become one of the most widespread natural hazards worldwide, posing severe threats to communities and ecosystems. Recent series of devastating wildfires have further underscored a critical need for more holistic, accurate, and efficient prediction of _when and where_ new fire events may occur. Emerging generative AI tools, particularly Transformer-based diffusion models, can capture the complex spatio-temporal dependencies exhibited by wildfires, but their architectural complexity and iterative denoising make inference computationally expensive. Knowledge distillation offers a path toward more efficient models, yet standard approaches apply uniform supervision across all events and ignore their distinct structural roles. We propose _MorseDistil_, a Morse-guided knowledge distillation framework that distills a Transformer-based teacher into a lightweight MLP student. Inspired by the concepts from low-dimensional topology, our central idea is to use the latent topological structure of the event sequence to guide knowledge transfer. Specifically, we construct a graph abstraction for spatio-temporal event representation and identify critical events through discrete Morse functions and their gradient pairings. These critical events then shape the distillation loss, concentrating teacher supervision on the most essential spatial and temporal structures. Extensive experiments on 10 wildfire benchmarks demonstrate that MorseDistill achieves superior predictive performance at substantially lower computational costs. We validate MorseDistill with respect to the state-of-the-art baselines on a diverse set of wildfire datasets, demonstrating MorseDistill's superior predictive capabilities of up to 40.6% in relative gains under reduced computational costs of up to 96.2%.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Knowledge Distillation, Wildfire Forecasting, Time-series

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_QNNhOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper address wildfire forecasting via student-teacher distillation. For this, it proposes to apply Morse theory to guide the student to focus on relevant events in space and time to add additional teacher supervision. First, a spatio-temporal graph is constructed then a Morse theory is used to simplify the graph into critical nodes. The distillation is then enforced on these detected critical nodes. The paper tested the proposed distillation method on different wildfire datasets and showed better results than different baselines.

Critical strengths
  • I think the proposed method is applicable to broader tasks beyond wildfire forecasting.
Critical weaknesses

This paper should be rejected because:

  • The paper is unclear and misses many details. Just to give few examples, the problem formulation is unclear, the input spatial and temporal resolution as well as the forecast horizon is not mentioned. What are the predictors? How is the architectural design? How is pretraining of the teacher network done?
  • Unfortunately, the experiments do not provide convincing evidence of the correctness of the proposed approach. Why should the critical events be wildfire? I also believe that the experimental setup is not standard for wildfire forecasts and the experiments were not run on benchmarks. The evaluation metric does not fit the task. The paper should report classification metrics e.g., F1 AORUC, AUPRC, ... etc. The RMSE is not a measure for wildfire forecasting.
Questions for authors
  • “Experiments on 10 wildfire benchmarks”. This is not true. Are these really benchmarks? I couldn’t find the references for these benchmarks. Where do the baselines come from and are the reported numbers from the benchmark’s papers?
  • Ablation study should report all combinations of MorseDistill components. Some components could be redundant. Why not only add weights to the critical events instead of feeding two graphs? I mean the methodological design is not motivated or validated.
  • Can’t we just give more weight to critical events? Why do we need to encode them separately? Do you have any experiments to motivate the design?
  • Did you use AI for writing the code? Scripts like transformer_st.py are AI-generated. Please make the AI-statement computable.
  • Can you show the raw graph and the simplified one from morse along with the critical events?
  • L437: “We conduct a systematic search over all hyperparameters”, the hyperparameters are identical for all models.
  • Is the spatial location of the graph fixed for each dataset? What is the spatial and temporal resolution? What is the forecast horizon? Many details are missing about the methodology. The paper only focuses on the MorseDistill. What are the input variables?

Minor:

  • L62-65: I do not think this is true. There are cheap transformers and many state-of-the-art systems e.g., weather forecast systems use transformers in an operational setting.
  • L120-129: point 2 is redundant.
  • Related works is incomplete: Wildfire is a well studied problem. You are also assuming that wildfire problem is essentially only timeseries forecast. This is not true.
  • L158-159: Diffusion is introduced without introduction. I think you should mention that you want to use generative diffusion in the introduction otherwise lines 158-159 are out of context.
  • L167: “Full and Critical Event Encoding” this is not shown in the figure. Unclear.
  • Shouldn’t the diffusion output discrete values to localize the events? How do you convert the continuous output to discrete ones. Or does the diffusion output wildfire probability at each time steps?
  • Basin-level alignment is unclear.
  • L342-343: what is the dxd_x? it is not clear?
  • L128-129: When you talk about cost, is it only about inference? Because this is misleading.
16 comments
AuthorsOpenReview ↗

Response to Reviewer QNNh (1/8)

Thank you for your constructive comments and suggestions. We have revised the paper to address all reviewers’ concerns. Here are our point-by-point responses.

Q1: The problem formulation, the input spatial, temporal resolution, forecast horizon, predictors, architectural design, pretraining.

A: We have added more detail into the paper. Given the observed history Hn={(ti,si)}i=1n\mathcal{H}_n=\{(t_i,{s}_i)\}_{i=1}^{n}, where tit_i and si{s}_i denote the time and location of the ii-th observation, the model learns the conditional distribution of (tn+1−tn,sn+1t_{n+1}-t_n,{s}_{n+1}) given Hn\mathcal{H}_n. The predictors are the times and spatial coordinates of past wildfire observations. The target is the next record in the dataset. In a hotspot sequence, two consecutive records can be detections of the same ongoing fire. Our MorseDistill and all baselines receive the same historical times and locations. We evaluate temporal accuracy by the RMSE between predicted and observed inter-event times, and spatial accuracy by the mean Euclidean distance between predicted and observed locations.

MorseDistill uses a Transformer-based diffusion teacher and a smaller student with MLP encoders, as shown in Figure 1. For each dataset, we first train the teacher on the training split to predict the time until the next recorded observation and its location. We then freeze the teacher and train the student on the same split. During distillation, both models receive the same history of observation times and locations for each training example. Each model uses two encoders. The first processes the full observation history. The second processes a chronological subsequence selected by the discrete Morse procedure. This procedure connects observations that are close in space and time and selects one representative per Morse basin at a local peak in detection density. Both encoders use sinusoidal embeddings for timestamps and a learned projection for spatial coordinates. The teacher uses two separate Transformer encoders, and the student uses two independently parameterized MLP encoders. To combine the branches, each model adds the feature vector from the subsequence encoder to the full-history feature vector at the selected observation’s original position. This retains every observation while allowing each selected event to carry information about its relationship to other selected events. The student’s training objective combines a diffusion loss on the observed next event with losses that match the teacher’s predictions and intermediate features. These distillation losses provide supervision over the full history and give additional weight to the Morse-selected observations. At inference, the student’s two MLP encoders produce a combined history representation. The student then uses this representation to guide iterative denoising and jointly predict when and where the next wildfire observation will be recorded.

Each event is represented by a continuous-valued tuple (t,lon,lat)(t,\mathrm{lon},\mathrm{lat}). The diffusion model predicts the elapsed time until the next event and its coordinates, (Δt,lon,lat)(\Delta t,\mathrm{lon},\mathrm{lat}), in continuous space and time. We use discretization only to construct the spatiotemporal Morse features that condition the model. For each dataset, we choose a spatial zoom level targeting approximately 400 cells (50–300 events per cell), and we divide time into 16 quantile bins. The forecast horizon is one event ahead. Each forecast uses the observed history through event ii to predict event i+1i+1, so the elapsed time varies across forecasts. Evaluation uses the observed history at each step and involves no autoregressive rollout.

The datasets cover different time scales. In the single-acquisition remote-sensing datasets (eMAS_fire, MASTER_fire, GOES18_smoke, GOES_smoke, and GK2A_fire), observations within a sequence span seconds to minutes. In Canada_fire and the US-state wildfire datasets, sequences can span roughly one day to one week and include observations from successive satellite revisits or fire reports. Table 1 shows mean training-sequence span for each dataset. For a sequence of nn observations, we compute its span as tn−t1t_n - t_1, the elapsed time from the first observation to the last, and average this quantity over all training sequences in that dataset.

To clarify these time scales, Table 1 reports the mean duration covered by a sequence, while the lead time for an individual forecast is the interval between two consecutive observations. For a sequence of nn observations, we compute its span as tn−t1t_n - t_1, the elapsed time from the first observation to the last, and average this quantity over all training sequences in that dataset.

Table 1

DatasetMean sequence span
eMAS32.8 sec
MASTER7.5 sec
GK2A8.6 min
Canada5.1 days
CA24 hr
AZ4.0 days
FL7.0 days
MT4.0 days
OK4.0 days
OR5.0 days
TX6.0 days
AuthorsOpenReview ↗

Response to Reviewer QNNh (2/8)

Q2: Why should the critical events be wildfire? I also believe that the experimental setup is not standard for wildfire forecasts and the experiments were not run on benchmarks. The evaluation metric does not fit the task. The paper should report classification metrics e.g., F1 AORUC, AUPRC, ... etc. The RMSE is not a measure for wildfire forecasting.

A: In this paper, we study next-event prediction in a spatio-temporal point process (STPP). Given the times and locations of past wildfire observations, the model predicts how much time will elapse before the next observation and where it will occur. In a hotspot sequence, successive observations can be detections of the same fire. Our choice of metrics follows the accepted practice in fire sciences (see, e.g., Eaturu and Vadrevu, “Evaluation of machine learning and deep learning algorithms for fire prediction in Southeast Asia” Scientific Reports 2025) as well as ML studies on the next-event evaluations such as DSTPP (Yuan et al., KDD 2023) and TEN-DM (Liu et al., ICLR 2026). Both DSTPP and TEN-DM report temporal RMSE for the interval until the next event and mean Euclidean distance for its location. These metrics measure the two quantities our model predicts. F1, AUROC, and AUPRC would be appropriate for a question such as whether a fire occurs in a specified grid cell within a given time window, and such classification task falls outside the scope of our study which focuses on predicting when and where the next wildfire observation will be recorded.

Eaturu, A. and Vadrevu, K.P., 2025. Evaluation of machine learning and deep learning algorithms for fire prediction in Southeast Asia. Scientific Reports, 15(1), p.18807.

Yuan, Y., Ding, J., Shao, C., Jin, D. and Li, Y., 2023, August. Spatio-temporal diffusion point processes. In Proceedings of the 29th ACM SIGKDD conference on knowledge discovery and data mining (pp. 3173-3184).

Liu, Y., Wang, K., Yang, C.Y., Gel, Y.R. and Chen, Y., 2026, January. TEN-DM: Topology-Enhanced Diffusion Model for Spatio-Temporal Event Prediction. The International Conference on Learning Representations (ICLR).

Q3: “Experiments on 10 wildfire benchmarks”. This is not true. Are these really benchmarks? I couldn’t find the references for these benchmarks. Where do the baselines come from and are the reported numbers from the benchmark’s papers?

A: Yes, they are really open-source benchmarks. eMAS, GK2A, and MASTER are three subsets of Fire-D (https://zenodo.org/records/15428368), a public collection of calibrated wildfire and smoke imagery with georeferenced labels maintained by NASA’s Jet Propulsion Lab. The three names identify the instruments or platforms that collected the observations. Fire-D provides the source data, and our paper defines the next-observation forecasting task used to compare methods. eMAS contains airborne multispectral imagery collected by the Enhanced MODIS Airborne Simulator aboard NASA’s ER-2 research aircraft. GK2A contains satellite imagery collected by the Advanced Meteorological Imager aboard South Korea’s Geo-KOMPSAT-2A geostationary satellite, which repeatedly observes the Korean Peninsula. MASTER contains airborne multispectral imagery collected by the MODIS/ASTER Airborne Simulator, which measures radiation across visible and infrared wavelengths.

For Canada (CAN), we use hotspot observations from the publicly available Canada Wildfire 2023 – Hotspot Data (https://www.kaggle.com/datasets/brsdincer/canada-wildfire-2023-hotspot-data).

For the United States, we use publicly available wildfire records from the US Wildfire Dataset (2014–2025) and evaluate all methods separately on six state-level subsets drawn from this collection: Arizona, California, Montana, Oklahoma, Oregon, and Texas (https://www.kaggle.com/datasets/firecastrl/us-wildfire-dataset).

All reported numbers come from running all models on the same datasets, that is, we ran all baselines ourselves under the same experimental settings and with the same computational budget as our MorseDistill.

AuthorsOpenReview ↗

Response to Reviewer QNNh (3/8)

Q4: Ablation study should report all combinations of MorseDistill components. Some components could be redundant. Why not only add weights to the critical events instead of feeding two graphs? I mean the methodological design is not motivated or validated.

A: We thank the reviewer for this suggestion. To directly examine whether the dual-branch Morse architecture is necessary, we conducted an additional ablation study comparing our dual-branch design against a simpler alternative that assigns higher weights to critical events within a single branch (Critical Weight). Both variants use the same critical-event identification procedure, allowing us to isolate the effect of explicitly modeling critical events through a separate branch. The results are summarized in Table 2. Our dual-branch design consistently outperforms critical-event reweighting across all four datasets, reducing spatial Euclidean distance by 31.09% and temporal RMSE by 8.32% on average. We have also evaluated statistical significance. Our dual-branch Morse architecture achieves statistically significant reductions in spatial Euclidean distance across all four datasets (p<0.05p < 0.05). For temporal RMSE, our method consistently achieves lower errors across all datasets, with a statistically significant improvement on CA (p=0.0085p = 0.0085). These results suggest that simply increasing the importance of critical events within a single representation is insufficient to capture their distinct structural contributions. We have added these new results to the revised version (see Appendix C.4).

Table 2 (∗^* indicates statistically significant)

DatasetDual-branch Spatial ↓Dual- branch Temporal ↓Critical weight Spatial ↓Critical weight Temporal ↓
CA1.2407±0.0166 ∗^*0.0852±0.0010 ∗^*2.4638±0.22310.1191±0.0059
AZ0.8926±0.0063 ∗^*0.3652±0.00791.4394±0.06000.3768±0.0055
OK0.8381±0.0118 ∗^*2.3260±0.01861.0260±0.02712.3394±0.0158
TX1.4509±0.0080 ∗^*0.8343±0.00521.7783±0.06570.8442±0.0123

Note that MorseDistill has one architectural component under consideration, namely the critical-event branch CC and four distillation objectives. Particularly, Lpred\mathcal{L}_{\mathrm{pred}} matches predictions, Lfeat\mathcal{L}_{\mathrm{feat}} matches full-history features, and LM-feat\mathcal{L}_{\mathrm{M\text{-}feat}} adds prediction supervision at critical events. These three objectives remain defined without CC; on the other hand, LM-feat\mathcal{L}_{\mathrm{M\text{-}feat}} matches the separately encoded critical-subsequence features HCH_C, so retaining this loss while removing CC is not a valid component combination. In essence, we note that reweighting changes the contribution of existing discrepancies to the loss. The separate branch changes the representations, as the teacher encodes the selected events together (independently of the noncritical events) and the student adds its critical-branch features at their original history positions. Consequently, reweighting full-history feature discrepancies does not implement the critical-subsequence feature objective or introduce the branch’s inference-time computation. From the theoretical stance, the branches receive observations from the same history, particularly the full sequence and its Morse-selected subsequence. The Morse construction supplies one representative per gradient basin, and the encoders do not explicitly receive the Morse quotient’s adjacency. Our additional Critical Weight ablation directly compares the separate-branch design with single-branch critical-event reweighting, using the same selection procedure.

Q5: Did you use AI for writing the code? Scripts like transformer_st.py are AI-generated. Please make the AI-statement computable.

A: We used AI to revise portions of the code for readability, including in transformer_st.py, so reviewers could more easily follow the implementation. We have updated the AI-use statement.

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

The paper studies next-event prediction in the spatio-temporal point process of wildfires. The motivation is that the optimal-accuracy Transformer with a diffusion scheme is expensive and difficult to support high-frequency rolling forecasts, while the standard distillation uniformly supervises all events, and dense areas will overwhelm small and unique fire activity clusters. The paper introduces discrete Morse theory to answer "where should the teacher supervision be concentrated": perform a one-to-one sorting on the KNN event graph according to the local spatio-temporal density, pair redundant vertices, and use the remaining key events as representatives of each gradient basin; Both the teacher with transformer and the student with MLPs are bidirectionally encoded for the full sequence and the key sequence and fused, with the loss being full sequence prediction level/feature level distillation with additional supervision on key events. The method was validated on 10 wildfire datasets; the spatial error is reduced siginficantly, and the computational costs is reduced by 96.2%.

Critical strengths

(1) Motivation is clear. The trade-off between accuracy and efficiency is the real pain point of operationalized forecasting.

(2) The existing topology with DL approach treats the topology as a representation signal. In this paper, it is transformed into a distillation supervision allocation criterion, which is a clearly distinguishable new proposition.

(3) The experimental setup is complete with 10 datasets across multiple sensors, including independent student baselines, four ablations, hyperparameter sensitivity, complexity analysis, and anonymous code.

(4) Efficiency gains should be reproducible.

Critical weaknesses

(1) The student model still uses 500-step diffusion sampling, and there is no reasoning delay in the manuscript; the abstract claim "reduces by 96.2%" actually refers to the training time per epoch. The claim of efficient inference lacks direct evidence.

(2) The definition of the teacher model is ambiguous: it does not specify which model the teacher is, nor does it describe the configuration and training method. The students completely outperformed the teacher without any explanation. This might be due to the teacher not being properly tuned.

(3) Topological necessity has not been verified: The only loss incurred is the local maximum of ρ ; Morse mechanisms such as gradient pairing, basin, quotient graph homotopy, etc. are not present in the loss; There are no non-topological counterparts such as random/uniform/top-k density/clustering representatives.

Questions for authors

See the weakness.

8 comments
AuthorsOpenReview ↗

Response to Reviewer Euyr (1/3)

Thank you very much for your positive comments on our work and for the valuable suggestions. We have revised the paper to address all reviewers’ concerns. We now respond to your points one by one as follows.

Q1: The student model still uses 500-step diffusion sampling, and there is no reasoning delay in the manuscript; the abstract claim "reduces by 96.2%" actually refers to the training time per epoch. The claim of efficient inference lacks direct evidence.

A: Thank you for your insightful comment. We have conducted additional experiments measuring the per-forecast inference latency of the teacher, and our student model under different DDIM sampling budgets (10, 20, 50, 100, 200, and 500 steps). These measurements include both history encoding and the complete iterative denoising procedure. As shown in the table below, at the same 500-step sampling budget, the student achieves a modest 1.0671–1.2945× speedup over the teacher across the ten datasets. Reducing the sampling budget yields substantially larger gains: Student@50 achieves a 3.3821–4.8631× speedup over Teacher@500 and a 3.1358–4.4880× speedup over Student@500. These results distinguish the gains from simplifying the history encoder from those obtained by reducing the number of denoising steps, showing that the latter provides the larger contribution to inference acceleration in these experiments.

Importantly, Tables 2 and 3 show that the student maintains competitive prediction performance with fewer sampling steps. Temporal prediction remains relatively stable across sampling budgets, while spatial accuracy exhibits greater dataset-dependent sensitivity. This suggests that a smaller sampling budget, such as 50 steps, provides a practical trade-off between inference efficiency and predictive accuracy. We have added these new results to the revised version (see Appendix C.2).

Table 1 (unit: ms/1k events)

DatasetTeacher@500Student@10Student@20Student@50Student@100Student@200Student@500
eMAS336.4359.4264.6780.07107.19156.37303.80
GK2A28354.394224.434808.745830.568396.5014535.7126167.60
MASTER400.7868.7275.10118.50150.74187.12371.59
CAN761.94123.51150.55176.04240.50342.82714.01
CA95.2417.3217.7421.3470.9037.9873.57
AZ173.1036.6340.6045.2465.2183.03149.48
MT222.0246.1759.7058.7380.23109.16202.26
OK279.5350.8555.4267.77107.08132.60248.42
OR306.2053.2488.7970.5292.16142.94260.56
TX346.1961.0066.8591.39113.82159.82300.54

Table 2

ModeleMAS Spatial ↓eMAS Temporal ↓GK2A Spatial ↓GK2A Temporal ↓MASTER Spatial ↓MASTER Temporal ↓CAN Spatial ↓CAN Temporal ↓CA Spatial ↓CA Temporal ↓
Student@100.2007±0.02992.5093±0.01470.1680±0.02998.4132±0.12366.8670±0.01894.5259±0.21330.8766±0.10890.5807±0.00391.2835±0.07800.0860±0.0030
Student@200.2088±0.04282.4897±0.00290.1629±0.03638.9956±0.08966.8646±0.01884.4057±0.04330.9232±0.12420.5804±0.00391.2912±0.07620.0873±0.0020
Student@500.2252±0.01362.4778±0.01070.1561±0.02019.9761±0.26616.8568±0.01684.4562±0.16260.9789±0.12280.5852±0.01211.3028±0.07350.0892±0.0032
Student@1000.2321±0.01482.5734±0.07570.1701±0.02379.6284±0.36096.8493±0.01334.4487±0.09830.9991±0.11880.5884±0.00861.3123±0.07560.0890±0.0022
Student@2000.2324±0.01732.7070±0.16360.1792±0.01649.8910±0.52276.8422±0.01094.5053±0.07941.0054±0.10810.6029±0.00511.3239±0.07370.0914±0.0028
Student@500 (paper reported)0.1806±0.01842.5125±0.02220.1769±0.01659.3061±0.06066.8536±0.00184.4160±0.03300.7396±0.03370.5775±0.00511.2407±0.01660.0852±0.0010
Teacher@5000.2070±0.03612.5785±0.09170.1910±0.01029.2694±0.08866.8430±0.02384.4451±0.06050.5163±0.03720.5927±0.01411.5957±0.05540.0932±0.0028

(Continued: please find Table 3 for Q1 in ``Response to Reviewer Euyr (2/3)'')

AuthorsOpenReview ↗

Response to Reviewer Euyr (2/3)

Table 3

ModelAZ Spatial ↓AZ Temporal ↓MT Spatial ↓MT Temporal ↓OK Spatial ↓OK Temporal ↓OR Spatial ↓OR Temporal ↓TX Spatial ↓TX Temporal ↓
Student@100.9796±0.04010.3569±0.01603.0435±0.12470.5024±0.00540.8734±0.02132.3447±0.01541.2040±0.13911.3129±0.01321.4311±0.02660.8571±0.0076
Student@201.0034±0.03780.3543±0.01143.0962±0.12760.5060±0.02290.9003±0.02432.3236±0.01491.2292±0.14441.3015±0.00781.4762±0.02030.8431±0.0112
Student@501.0134±0.04020.3609±0.00893.1096±0.13170.5084±0.03540.8924±0.02782.3127±0.01241.2303±0.14191.2918±0.01061.4983±0.02020.8371±0.0278
Student@1001.0249±0.04000.3654±0.01113.0801±0.12570.5041±0.04750.9001±0.02552.3090±0.01651.2362±0.14251.2953±0.00851.5096±0.01920.8444±0.0306
Student@2001.0221±0.03960.3648±0.01163.1162±0.13180.5216±0.05640.9023±0.03312.3317±0.01811.2458±0.13911.2944±0.01261.4965±0.01920.8730±0.0529
Student@500 (paper)0.8926±0.00630.3652±0.00792.9960±0.17200.4901±0.01450.8381±0.01182.3260±0.01861.0272±0.01431.2757±0.00141.4509±0.00800.8343±0.0052
Teacher@5001.0091±0.01250.3679±0.00862.9634±0.07790.4904±0.02870.9118±0.05672.3191±0.00371.2747±0.15331.2892±0.00911.4598±0.01740.8625±0.0313

Q2: The definition of the teacher model is ambiguous.

A: We designed the teacher by extending a diffusion-based spatiotemporal point-process model with a second Transformer encoder. The first encoder processes the time and location of every observation in history. The second processes a subset of the same observations, selected by the discrete Morse construction as representatives of their respective basins. For each selected observation, we add its representation from the second branch to its representation at the corresponding position in the full sequence. The fused sequence conditions a single diffusion denoiser (ST_diffusion_MM), and forecasts are generated using 500 DDIM sampling steps. We train the teacher from scratch for each dataset and random seed for 100 epochs using run_teacher.sh, with --alpha 0.05, --beta_topo_st 0.001, and --no_vlm. Hence, both branches use the same underlying wildfire observation history. Once trained, the teacher is frozen and supplies prediction- and feature-level supervision to the student.

We have compared the teacher with and without the Morse-selected branch. As shown in Table 4, incorporating the Morse-selected branch reduces mean spatial and temporal errors on all seven datasets. Averaging the relative reductions equally across datasets, the teacher with the Morse-selected branch achieves 24.2% lower spatial error and 6.8% lower temporal error than the teacher without this branch. Two-sided Welch’s t-tests indicate statistically significant reductions in spatial Euclidean distance on CA, AZ, OR, and TX (p<0.05p<0.05), and in temporal RMSE on CA and AZ (p<0.05p<0.05). These results support the contribution of the Morse-selected branch, with larger average improvements in spatial prediction and more modest gains in temporal prediction. We have added these new results to the revised version (see Appendix C.3).

Table 4 (∗^* indicates statistically significant)

DatasetTeacher with critical events (Spatial ↓)Teacher w/o critical events (Spatial ↓)Teacher with critical events (Temporal ↓)Teacher w/o critical events (Temporal ↓)
GK2A0.1769±0.01650.2072±0.04329.3061±0.06069.5605±0.1500
CAN0.7396±0.03370.8113±0.04050.5775±0.00510.5930±0.0157
CA1.2407±0.0166 ∗^*2.5524±0.31780.0852±0.0010 ∗^*0.1234±0.0106
AZ0.8926±0.0063 ∗^*1.4359±0.06640.3652±0.0079 ∗^*0.3828±0.0055
MT2.9960±0.17203.1694±0.17280.4901±0.01450.4922±0.0108
OR1.0272±0.0143 ∗^*1.5006±0.07501.2757±0.00141.2758±0.0153
TX1.4509±0.0080 ∗^*1.8013±0.04600.8343±0.00520.8898±0.0621

Table 2 (in the main body) distinguishes the standalone MLP student from the distilled student. Across the five datasets in that table, the Transformer teacher achieves lower mean spatial and temporal errors than the standalone student in every case. After distillation, the student improves on the teacher by an average of 13.31% in spatial error and 3.24% in temporal error. The student learns from the observed next event as well as from the teacher’s predictions and intermediate features. It can therefore correct a teacher prediction when the observed outcome supports a different forecast. The ablations support this explanation, as removing either prediction-level or feature-level supervision increases both mean errors on all five datasets.

AuthorsOpenReview ↗

Response to Reviewer Euyr (3/3)

Q3: Topological necessity has not been verified: The only loss incurred is the local maximum of ρ ; Morse mechanisms such as gradient pairing, basin, quotient graph homotopy, etc. are not present in the loss; There are no non-topological counterparts such as random/uniform/top-k density/clustering representatives.

A: We have compared Morse-based selection with random sampling, uniform sampling, top-k density selection, and clustering-based selection on five state datasets. Table 5 demonstrates that the model with Morse-based selection achieves the lowest mean spatial and temporal errors on all five. Relative to the best of the four alternatives for each state and metric, the reductions average 36.3% in spatial error and 4.3% in temporal error, weighting each state equally. In addition, Morse achieves significantly lower spatial errors in CA, AZ, OK, and OR (p=0.0002,0.0034,0.0317p = 0.0002, 0.0034, 0.0317, and 0.00120.0012, respectively), and significantly lower temporal errors in CA and AZ (p=0.0013p = 0.0013 and 0.03750.0375, respectively). The evidence is strongest for predicting location. For timing, some differences are small, particularly in Montana and Oklahoma. We have added these new results to the revised version (see Appendix C.4).

Table 5 (∗^* indicates statistically significant)

MethodCA Spatial ↓CA Temporal ↓AZ Spatial ↓AZ Temporal ↓MT Spatial ↓MT Temporal ↓OK Spatial ↓OK Temporal ↓OR Spatial ↓OR Temporal ↓
Morse1.2407±0.0166 ∗^*0.0852±0.0010 ∗^*0.8926±0.0063 ∗^*0.3652±0.0079 ∗^*2.9960±0.17200.4901±0.01450.8381±0.0118 ∗^*2.3260±0.01861.0272±0.0143 ∗^*1.2757±0.0014
Random3.2670±0.29400.1090±0.01202.2030±0.44400.3900±0.00403.9680±0.46500.5490±0.06201.3720±0.17002.3822±0.00202.3150±0.71101.2850±0.0050
Uniform2.5720±0.05500.1010±0.00001.9410±0.02200.5670±0.10905.0940±0.25900.6410±0.17501.6030±0.06702.3280±0.01503.4920±0.13001.3070±0.0270
Top-k density2.9940±0.17100.1040±0.00101.7630±0.09000.3840±0.00703.1470±0.02400.5280±0.05401.5040±0.02502.3300±0.01101.6260±0.05001.2940±0.0080
Clustering3.7240±1.63200.1670±0.11203.5570±0.22300.4130±0.02803.5540±0.30200.4910±0.01702.2930±0.11602.3730±0.02802.3920±1.33001.2840±0.0060
Reviewer_1TLKOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The authors propose a distillation based method for spatio-temporal point process (STPP) forecasting of wildfire events. A Transformer-encoder diffusion teacher is distilled into a student whose history encoder is an MLP. They employ discrete Morse functions to identify critical events. A symmetrized event graph is constructed from the fixed history of each fire detection, with its observation time and spatial coordinate and computes a Gaussian-kernel density score on it. The gradient field identifies vertex–edge pairs that are structurally redundant and identifies critical cells as the cells that can not be eliminated through this reduction. These critical events get a separate encoder branch in both teacher and student, and two extra distillation losses (prediction-level and feature-level). Experiments cover 10 wildfire datasets against 7 baselines, with ablations, hyperparameter sensitivity, and cost measurements.

Critical strengths
  • Well-motivated problem. Modeling the spatio-temporal dependencies between wildfire events, rather than treating detections independently or on a fixed grid, is relevant for hazard forecasting and comparatively underexplored in the ML literature. Knowledge distillation is a reasonable angle given the cost of recent diffusion-based STPP models.
  • Lightweight, reproducible selection criterion. The Morse construction is deterministic, using rank-based tie-breaking instead of random jitter. It has no learned parameters, and it runs in O(n log n + |E|) time on top of the kNN graph, so it adds negligible overhead to training.
  • Addresses a real weakness of uniform distillation. The paper identifies a common issue in feature- and prediction-level distillation: weighting all events equally lets densely sampled, redundant regions dominate the supervision signal, while small but structurally distinct clusters of activity are underweighted. Allocating supervision per basin, with one representative each, is a principled response that keeps sparse local activity from being drowned out.
  • Sound design choice on supervision scope. The authors explicitly state that criticality on the event graph does not imply predictive redundancy. Accordingly, MorseDistill keeps full-history prediction and feature supervision and uses the critical events only for additional targeted losses. This avoids the information loss that pure subsampling would introduce.
  • Broad empirical coverage and reproducibility. Experiments span 10 datasets from airborne, geostationary, and ground-record sources. Results are averaged over 3 seeds, and the paper includes ablations, loss-weight and k sensitivity studies, and an anonymized code repository.
Critical weaknesses
  • Evaluation protocol is poor. Building sequences from FIRMS hotspots and NOAA NCEI records is not described. This includes the event definition, how sequences are segmented, overlap, and whether splits are temporal or random. With random splits of overlapping windows from the same fire, leakage is likely. Spatial and temporal errors are reported in unexplained units. All hyperparameters are identical across the 10 datasets despite a “systematic search.” In Table 4, the default values are best or tied-best almost everywhere. It is not stated whether selection used validation or test data.
  • The spatiotemporal prediction of future events is undefined. Do the authors mean the prediction in the next minutes/ hours/ days?
  • The efficiency claim does not match the stated motivation. The introduction motivates the work with the inference cost of Transformer encoders combined with iterative denoising. However, the student keeps the conditional diffusion denoiser with 500 sampling steps (Table 3), and only the history encoder becomes an MLP. Table 6 does not report inference latency. Training-time accounting also appears to exclude two costs: pretraining the teacher, and running the teacher’s 500-step sampler
  • Most of the gains over baselines appear to come from the teacher, not from distillation. Table 2 shows that the teacher alone already beats every Table 1 baseline by large margins on several datasets. The teacher is not a standard model either. It contains a second Transformer branch over Morse-selected events, so it is a new architecture that is trained with the Morse construction. The teacher’s architecture, its training, and its origin are not described.
  • The domain claims are stronger than the experiments. The model uses no weather, fuel, or terrain covariates, which the authors acknowledge. Next-hotspot prediction from detection histories alone is a narrow proxy for “when and where new fire events may occur.” The introduction’s framing around operational risk management should be toned down.
Questions for authors

Questions for the authors

  1. What is the teacher architecture?
  2. What is the per-forecast inference latency for the teacher, the student, and the baselines, including the 500-step sampler?
  3. How are sequences and splits constructed, and is there temporal separation between train and test?
  4. Αdd an example plot of prediction with the other baselines
  5. How exactly is L_M-pred computed?
11 comments
AuthorsOpenReview ↗

Response to Reviewer 1TLK (1/5)

Thank you very much for your careful review and constructive comments. We have revised the paper to address all reviewers’ concerns. Regarding the concerns of the Reviewer 1TLK, we provide the following responses.

Q1: Building sequences from FIRMS hotspots and NOAA NCEI records is not described. This includes the event definition, how sequences are segmented, overlap, and whether splits are temporal or random. With random splits of overlapping windows from the same fire, leakage is likely. Spatial and temporal errors are reported in unexplained units. & How are sequences and splits constructed, and is there temporal separation between train and test? All hyperparameters are identical across the 10 datasets despite a “systematic search.” In Table 4, the default values are best or tied-best almost everywhere. It is not stated whether selection used validation or test data.

A: Given the observed history Hn={(ti,si)}i=1n\mathcal{H}_n=\{(t_i,{s}_i)\}_{i=1}^{n}, where tit_i and si{s}_i denote the time and location of the ii-th observation, the model learns the conditional distribution of (tn+1−tn,sn+1t_{n+1}-t_n,{s}_{n+1}) given Hn\mathcal{H}_n. The predictors are the times and spatial coordinates of past wildfire observations. The target is the next record in the dataset. In a hotspot sequence, two consecutive records can be detections of the same ongoing fire.

For all datasets, we partition the data chronologically at the sequence level, before generating any sliding-window prediction examples. The training, validation, and test sets therefore correspond to successively later time periods. Each sequence is assigned in its entirety to exactly one split, with no observations or sequences shared across splits. Prediction windows are then generated independently within each split, ensuring that both the input history and prediction target remain within that split. Thus, no window crosses a split boundary, preventing information leakage caused by overlapping windows across training, validation, and test sets. In our original experiments, we explored different hyperparameter combinations and found that the default configuration generally achieved strong performance across most datasets. We therefore adopted this configuration consistently across datasets and reported the corresponding results in the manuscript. Following the reviewer’s suggestion, we expanded the sensitivity analysis by evaluating two additional combinations of the four loss weights, as specified in Table 1. Table 2 shows that the default configuration achieves the lowest mean spatial error on 9 of the 10 datasets and the lowest mean temporal error on 5 of the 10 datasets. The additional configurations improve spatial prediction on GK2A and temporal prediction on GK2A, MASTER, CAN, MT, and OK, although these temporal improvements are modest. These results support the default configuration as a broadly competitive choice while showing that dataset-specific tuning can further improve individual metrics. We have added these results and clarified the hyperparameter selection procedure in the revised manuscript (see Appendix C.1).

Table 1

Configλ_predλ_featλ_{M-pred}λ_{M-feat}
Default (Ours, paper)0.15.00.10.1
C10.31.00.50.3
C20.51.50.20.2

Table 2

DatasetDefault Spatial ↓Default Temporal ↓C1 Spatial ↓C1 Temporal ↓C2 Spatial ↓C2 Temporal ↓
eMAS0.1806±0.01842.5125±0.02220.2781±0.04722.6020±0.07230.2436±0.03162.6127±0.1080
GK2A0.1769±0.01659.3061±0.06060.1731±0.01549.2682±0.04000.1936±0.02239.3389±0.1472
MASTER6.8536±0.00184.4160±0.03306.8543±0.01094.3845±0.06106.8571±0.01884.3802±0.0179
CAN0.7396±0.03370.5775±0.00510.9626±0.14310.5709±0.00300.7974±0.03320.5778±0.0059
CA1.2407±0.01660.0852±0.00101.2928±0.05170.0880±0.00241.2928±0.05300.0874±0.0022
AZ0.8926±0.00630.3652±0.00790.9911±0.04560.3691±0.00970.9947±0.05180.3693±0.0101
MT2.9960±0.17200.4901±0.01453.0751±0.14940.4838±0.01423.0636±0.15400.4918±0.0222
OK0.8381±0.01182.3260±0.01860.8610±0.02002.3239±0.01280.8646±0.02462.3190±0.0058
OR1.0272±0.01431.2757±0.00141.1961±0.13741.2865±0.00921.1956±0.14951.2850±0.0077
TX1.4509±0.00800.8343±0.00521.4797±0.03230.8559±0.02431.4863±0.02560.8409±0.0053
AuthorsOpenReview ↗

Response to Reviewer 1TLK (2/5)

Q2: The spatiotemporal prediction of future events is undefined. Do the authors mean the prediction in the next minutes/ hours/ days?

A: We predict the next recorded observation in each sequence. Given the times and locations of the first ii observations, the model jointly predicts the elapsed time until observation i+1i+1, Δti+1=ti+1−ti\Delta t_{i+1}=t_{i+1}-t_i, and its location. Hence, we predict the next to occur event (i.e., the forecasting horizon is 1-step ahead). Since we consider spatio-temporal point processes (STPP), the time between the most recent event and the next event is random and can range from hours to weeks. For wildfire hotspots, the next observation may be another detection of an ongoing fire and the time between the most recent and next predicted event ranges from a few hours (e.g., around 8 hours between two consecutive hotspot detections in Canada_fire) to over three months (99 days) in the longest case we observe (OK_fire), with a mean gap on the order of a day and most gaps falling well under two weeks . Also,the datasets cover different time scales. In eMAS, MASTER, and GK2A datasets, observations within a sequence span seconds to minutes. In Canada and the US-state wildfire datasets, sequences can span roughly one day to one week and include observations from successive satellite revisits or fire reports. To clarify these time scales, Table 3 reports the mean training-sequence span for each dataset. For a sequence of nn observations, we compute its span as tn−t1t_n - t_1, the elapsed time from the first observation to the last, and average this quantity over all training sequences in that dataset.

Table 3

DatasetMean sequence span
eMAS32.8 sec
MASTER7.5 sec
GK2A8.6 min
Canada5.1 days
CA24 hr
AZ4.0 days
FL7.0 days
MT4.0 days
OK4.0 days
OR5.0 days
TX6.0 days
AuthorsOpenReview ↗

Response to Reviewer 1TLK (3/5)

Q3: The student keeps the conditional diffusion denoiser with 500 sampling steps (Table 3). Training-time accounting also appears to exclude two costs: pretraining the teacher, and running the teacher’s 500-step sampler.

A: We evaluated the student using 10, 20, 50, 100, 200, and 500 denoising steps, please refer to Tables 4 and 5. At 10 steps, temporal RMSE differs from the 500-step result by less than 3% on nine of the ten datasets and improves by 9.6% on GK2A. The effect on spatial accuracy is more dataset-dependent. Relative to Student@500, Student@10 increases spatial error by only 0.2% on MASTER, 3.4% on California, and 1.6% on Montana, while reducing it by 5.0% on GK2A and 1.4% on Texas. Larger differences arise on Canada and Oregon, where spatial errors increase by 18.5% and 17.2%, respectively. Accuracy is not monotonic in the number of sampling steps: intermediate budgets achieve the lowest errors on several metrics. Overall, Student@10 requires 98% fewer denoiser evaluations than Student@500 while preserving comparable temporal accuracy, with a dataset-dependent trade-off in location prediction. These results indicate that the main accuracy benefit of a larger sampling budget lies in improved spatial prediction on a subset of datasets. We have added these new results to the revised version (see Appendix C.2).

Table 4

ModeleMAS Spatial ↓eMAS Temporal ↓GK2A Spatial ↓GK2A Temporal ↓MASTER Spatial ↓MASTER Temporal ↓CAN Spatial ↓CAN Temporal ↓CA Spatial ↓CA Temporal ↓
Student@100.2007±0.02992.5093±0.01470.1680±0.02998.4132±0.12366.8670±0.01894.5259±0.21330.8766±0.10890.5807±0.00391.2835±0.07800.0860±0.0030
Student@200.2088±0.04282.4897±0.00290.1629±0.03638.9956±0.08966.8646±0.01884.4057±0.04330.9232±0.12420.5804±0.00391.2912±0.07620.0873±0.0020
Student@500.2252±0.01362.4778±0.01070.1561±0.02019.9761±0.26616.8568±0.01684.4562±0.16260.9789±0.12280.5852±0.01211.3028±0.07350.0892±0.0032
Student@1000.2321±0.01482.5734±0.07570.1701±0.02379.6284±0.36096.8493±0.01334.4487±0.09830.9991±0.11880.5884±0.00861.3123±0.07560.0890±0.0022
Student@2000.2324±0.01732.7070±0.16360.1792±0.01649.8910±0.52276.8422±0.01094.5053±0.07941.0054±0.10810.6029±0.00511.3239±0.07370.0914±0.0028
Student@500 (paper reported)0.1806±0.01842.5125±0.02220.1769±0.01659.3061±0.06066.8536±0.00184.4160±0.03300.7396±0.03370.5775±0.00511.2407±0.01660.0852±0.0010
Teacher@5000.2070±0.03612.5785±0.09170.1910±0.01029.2694±0.08866.8430±0.02384.4451±0.06050.5163±0.03720.5927±0.01411.5957±0.05540.0932±0.0028

Table 5

ModelAZ Spatial ↓AZ Temporal ↓MT Spatial ↓MT Temporal ↓OK Spatial ↓OK Temporal ↓OR Spatial ↓OR Temporal ↓TX Spatial ↓TX Temporal ↓
Student@100.9796±0.04010.3569±0.01603.0435±0.12470.5024±0.00540.8734±0.02132.3447±0.01541.2040±0.13911.3129±0.01321.4311±0.02660.8571±0.0076
Student@201.0034±0.03780.3543±0.01143.0962±0.12760.5060±0.02290.9003±0.02432.3236±0.01491.2292±0.14441.3015±0.00781.4762±0.02030.8431±0.0112
Student@501.0134±0.04020.3609±0.00893.1096±0.13170.5084±0.03540.8924±0.02782.3127±0.01241.2303±0.14191.2918±0.01061.4983±0.02020.8371±0.0278
Student@1001.0249±0.04000.3654±0.01113.0801±0.12570.5041±0.04750.9001±0.02552.3090±0.01651.2362±0.14251.2953±0.00851.5096±0.01920.8444±0.0306
Student@2001.0221±0.03960.3648±0.01163.1162±0.13180.5216±0.05640.9023±0.03312.3317±0.01811.2458±0.13911.2944±0.01261.4965±0.01920.8730±0.0529
Student@500 (paper reported)0.8926±0.00630.3652±0.00792.9960±0.17200.4901±0.01450.8381±0.01182.3260±0.01861.0272±0.01431.2757±0.00141.4509±0.00800.8343±0.0052
Teacher@5001.0091±0.01250.3679±0.00862.9634±0.07790.4904±0.02870.9118±0.05672.3191±0.00371.2747±0.15331.2892±0.00911.4598±0.01740.8625±0.0313

One potential explanation of this phenomenon is that temporal dependencies tend to be less dispersed and can be captured with relatively fewer denoising steps, whereas spatial event distributions are often more heterogeneous, requiring finer refinement during sampling. This may explain why temporal accuracy remains relatively stable across sampling budgets, while spatial accuracy exhibits greater dataset-dependent sensitivity.

(Continued: please find the remaining answer for Q3 in ``Response to Reviewer 1TLK (4/5)'')

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.