pith. machine review for the scientific record. sign in

arxiv: 2605.10298 · v1 · submitted 2026-05-11 · 💻 cs.LG

Recognition: 2 theorem links

· Lean Theorem

Set Prediction for Next-Day Active Fire Forecasting

Diogenis Antonopoulos, Georgios Athanasiou, Ioannis Papoutsis, Nuno Carvalhais, Stijn Hantson, Xin Yu, Yuchen Bai

Authors on Pith no claims yet

Pith reviewed 2026-05-12 04:16 UTC · model grok-4.3

classification 💻 cs.LG
keywords wildfire forecastingset predictionHungarian matchingactive fire detectionquery-based modelsmachine learningglobal benchmark
0
0 comments X

The pith

Wildfire forecasting can be reframed as predicting a fixed-size ranked set of fire cluster centers from weather and satellite data.

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

The paper shows that next-day active fire locations can be forecasted directly as a sparse point set rather than as probability values on a coarse grid. From 48 hours of meteorology, vegetation indices, land cover, and prior fires, the model outputs a ranked collection of potential ignition points on a 375 m global grid. It trains this output end-to-end by matching predicted points to observed fires with the Hungarian algorithm and an asymmetric loss that separates classification from localization. On a worldwide held-out test set the best variant reaches 38.2 percent average precision, accounts for 53.4 percent of fire radiative power mass, and places 54.1 percent of observed clusters within 5 km. The result matters because it supplies localized event predictions that can feed early-warning systems and carbon-emission calculations without an intermediate gridding step.

Core claim

WISP reformulates next-day active fire forecasting as point-set prediction: a query-based network ingests 48-hour covariate stacks and produces a fixed-size ranked set of future fire cluster centres; end-to-end training uses Hungarian matching with asymmetric classification-localization weighting to reconcile the conflicting demands placed on the classification score.

What carries the argument

Query-based set predictor trained with Hungarian matching and asymmetric classification-localization weighting.

If this is right

  • Forecasts become lists of discrete locations rather than raster danger maps, enabling direct use in dispatch and emissions models.
  • A new global hourly multi-source benchmark is released for the set-prediction formulation of wildfire forecasting.
  • Performance metrics of 38.2 percent AP, 53.4 percent FRP-weighted coverage, and 54.1 percent localization within 5 km are established as reference numbers for future work.
  • Sparse set prediction is shown to be a workable alternative to grid-based regression for high-resolution, low-density event forecasting.

Where Pith is reading between the lines

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

  • The same architecture could be tested on other sparse geospatial events such as lightning strikes or new disease outbreaks.
  • Replacing the fixed set size with a learned cardinality head would remove the need to guess how many fires will occur on a given day.
  • Coupling the set output with uncertainty estimates per point would indicate which predicted ignitions are most reliable for operational use.

Load-bearing premise

The fixed set size together with asymmetric weighting in the matching loss adequately balances assignment, ranking, and query activation without introducing bias in regions where fires are rare.

What would settle it

A new test collection of fire events, drawn from an unseen continent or fire season, in which the model recovers less than 40 percent of high-FRP clusters within 5 km while its reported average precision stays above 35 percent.

Figures

Figures reproduced from arXiv: 2605.10298 by Diogenis Antonopoulos, Georgios Athanasiou, Ioannis Papoutsis, Nuno Carvalhais, Stijn Hantson, Xin Yu, Yuchen Bai.

Figure 1
Figure 1. Figure 1: Overview of WISP. set prediction to oriented objects in aerial imagery through a point-axis representation [27]; and MapTR [28] represents vectorized map elements as structured point sets with hierarchical bipartite matching. These examples show that set prediction is well suited to sparse spatial outputs. WISP brings this sparse localization paradigm to next-day active fire forecasting. 3 WISP: Wildfire I… view at source ↗
Figure 2
Figure 2. Figure 2: Qualitative set-prediction example on one test entity. (a) Ground-truth (GT) 24-hour active [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: The valid-region and spatial-jitter diagnostic during training process. [PITH_FULL_IMAGE:figures/full_fig_p016_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Burning-pixel frequency maps for the train, validation, and test splits. For each split, the [PITH_FULL_IMAGE:figures/full_fig_p016_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Evolution of WISP query predictions on the first validation entity during the first 120 training [PITH_FULL_IMAGE:figures/full_fig_p017_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Additional qualitative gallery for WISP variant v1. [PITH_FULL_IMAGE:figures/full_fig_p020_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Additional qualitative gallery for WISP variant v2. [PITH_FULL_IMAGE:figures/full_fig_p020_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Additional qualitative gallery for WISP variant v3. [PITH_FULL_IMAGE:figures/full_fig_p021_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Additional qualitative gallery for WISP variant v4. [PITH_FULL_IMAGE:figures/full_fig_p021_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Additional qualitative gallery for WISP variant v5. [PITH_FULL_IMAGE:figures/full_fig_p022_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Additional qualitative gallery for WISP variant v6. [PITH_FULL_IMAGE:figures/full_fig_p022_11.png] view at source ↗
read the original abstract

Accurate next-day active fire forecasts can support early warning, disaster response, forest risk assessment, and downstream estimation of fire-related carbon emissions. Existing machine learning approaches to wildfire forecasting typically predict wildfire danger or fire probability on kilometre-scale daily grids, which is useful for regional warning but does not directly represent localized fire events. We propose Wildfire Ignition Set Predictor (WISP), a query-based model that reformulates next-day active fire forecasting as point-set prediction. From 48 hours of covariates including meteorology, satellite vegetation products, static land, and fire history, WISP predicts a fixed-size ranked set of future active fire cluster centres on a 375 m grid across globally distributed regions. The model is trained end-to-end with Hungarian matching; to address the conflicting roles of the classification score in assignment, ranking, and query activation, we use asymmetric classification-localization weighting in matching and loss. We further construct a globally distributed, hourly, multi-source benchmark for this task. On a held-out test set spanning fire regions worldwide, the best WISP variant achieves 38.2% average precision (AP) for ranked fire-centre detections, covers 53.4% of fire cluster mass weighted by fire radiative power (FRP), and localizes 54.1% of observed clusters within 5 km. These results establish sparse set prediction as a viable formulation for high-resolution wildfire forecasting and provide a benchmark for future work in this regime.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 1 minor

Summary. The paper proposes Wildfire Ignition Set Predictor (WISP), a query-based set prediction model that takes 48 hours of multi-source covariates (meteorology, vegetation, land, fire history) and outputs a fixed-size ranked set of next-day active fire cluster centers on a 375 m global grid. It is trained end-to-end using Hungarian matching with asymmetric classification-localization weighting. On a held-out global test set the best variant reports 38.2% AP for ranked detections, 53.4% FRP-weighted cluster-mass coverage, and 54.1% of observed clusters localized within 5 km, positioning sparse set prediction as a viable formulation for high-resolution wildfire forecasting and supplying a new benchmark.

Significance. If the reported metrics are reproducible, the work is significant because it reframes wildfire forecasting from per-grid probability maps to direct prediction of localized events, which better matches operational needs for early warning and carbon-emission estimation. The construction of a globally distributed, hourly, multi-source benchmark is a concrete, reusable contribution that future work can build upon. End-to-end training with bipartite matching and the explicit handling of the classification score's multiple roles via asymmetric weighting are technically interesting design choices for sparse point-set prediction.

major comments (2)
  1. [Abstract] Abstract: the central performance claims (38.2% AP, 53.4% FRP-weighted coverage, 54.1% localization) are stated for a held-out global test set, yet the manuscript supplies no description of data-construction details, temporal/spatial exclusion rules, train/test split criteria, or error bars. These omissions are load-bearing because they prevent verification that the metrics reflect genuine generalization rather than post-hoc choices or leakage.
  2. [Model training description] Model training description: the asymmetric classification-localization weighting inside the Hungarian matching is introduced to resolve the classification score's conflicting roles in assignment, ranking, and query activation. However, no ablation across fire-density strata or analysis of query-activation bias in globally sparse regimes (most 375 m cells contain zero fires) is provided. Without such evidence the weighting's sufficiency for preventing under- or over-activation remains unverified and directly affects the reliability of the reported AP and coverage numbers.
minor comments (1)
  1. [Abstract] The abstract and methods would benefit from stating the exact number of queries, the precise form of the asymmetric weighting coefficients, and the architecture backbone (e.g., transformer depth or encoder type) so that the experimental setup is fully reproducible from the text.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback on our manuscript. We address each major comment below and indicate the corresponding revisions.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the central performance claims (38.2% AP, 53.4% FRP-weighted coverage, 54.1% localization) are stated for a held-out global test set, yet the manuscript supplies no description of data-construction details, temporal/spatial exclusion rules, train/test split criteria, or error bars. These omissions are load-bearing because they prevent verification that the metrics reflect genuine generalization rather than post-hoc choices or leakage.

    Authors: We acknowledge that the abstract, due to its length constraints, does not detail the data construction process. The full manuscript describes the multi-source covariate assembly, global 375 m gridding, and the construction of the held-out test set in the methods section, using temporal separation and spatial exclusions to avoid leakage. Error bars from multiple runs are reported in the results. To make the abstract more self-contained, we will add a brief clause summarizing the held-out test set construction and direct readers to the methods for full details on splits and error computation. This constitutes a partial revision. revision: partial

  2. Referee: [Model training description] Model training description: the asymmetric classification-localization weighting inside the Hungarian matching is introduced to resolve the classification score's conflicting roles in assignment, ranking, and query activation. However, no ablation across fire-density strata or analysis of query-activation bias in globally sparse regimes (most 375 m cells contain zero fires) is provided. Without such evidence the weighting's sufficiency for preventing under- or over-activation remains unverified and directly affects the reliability of the reported AP and coverage numbers.

    Authors: We agree that additional analysis would strengthen the claims regarding the asymmetric weighting. The weighting scheme is introduced in the model training section specifically to address the sparsity of fires and the multiple roles of the classification score. In the revised manuscript we will include an ablation varying the classification-localization weights and report results stratified by fire-density regimes (high-, medium-, and low-activity regions), along with query activation statistics, to verify robustness against under- or over-activation in globally sparse settings. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical held-out metrics from end-to-end trained set predictor

full rationale

The paper's central claims are empirical performance numbers (38.2% AP, 53.4% FRP-weighted coverage, 54.1% localization within 5 km) obtained on a held-out global test set after training WISP with Hungarian matching and asymmetric classification-localization weighting. No equations, predictions, or first-principles derivations are presented that reduce these metrics to quantities fitted on the evaluation data itself. The derivation chain consists of a standard query-based architecture plus a training procedure whose outputs are evaluated externally; this is self-contained against benchmarks and contains no self-definitional, fitted-input-renamed-as-prediction, or self-citation-load-bearing steps.

Axiom & Free-Parameter Ledger

1 free parameters · 0 axioms · 1 invented entities

The central claim rests on the viability of set prediction for sparse events and on the quality of the newly constructed benchmark; the model itself introduces many implicit hyperparameters and architectural choices typical of deep learning but no explicit free parameters or invented physical entities are named in the abstract.

free parameters (1)
  • model hyperparameters and architecture details
    Standard deep-learning training choices (query count, loss weights, network depth) that are not enumerated in the abstract but are required for the reported performance.
invented entities (1)
  • WISP query-based set predictor no independent evidence
    purpose: To output ranked fire cluster centers directly
    New model formulation introduced in the paper; no independent evidence outside the reported metrics is provided.

pith-pipeline@v0.9.0 · 5583 in / 1307 out tokens · 41816 ms · 2026-05-12T04:16:15.218351+00:00 · methodology

discussion (0)

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

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

Reference graph

Works this paper leans on

41 extracted references · 41 canonical work pages

  1. [1]

    Increasing frequency and intensity of the most extreme wildfires on earth,

    C. X. Cunningham, G. J. Williamson, and D. M. Bowman, “Increasing frequency and intensity of the most extreme wildfires on earth,”Nature ecology & evolution, vol. 8, no. 8, pp. 1420–1425, 2024

    work page 2024

  2. [2]

    Global data-driven prediction of fire activity,

    F. Di Giuseppe, J. McNorton, A. Lombardi, and F. Wetterhall, “Global data-driven prediction of fire activity,”Nature Communications, vol. 16, p. 2918, Apr. 2025

    work page 2025

  3. [3]

    Mesogeos: A multi-purpose dataset for data-driven wildfire modeling in the mediterranean,

    S. Kondylatos, I. Prapas, G. Camps-Valls, and I. Papoutsis, “Mesogeos: A multi-purpose dataset for data-driven wildfire modeling in the mediterranean,”Advances in Neural Information Processing Systems, vol. 36, pp. 50661–50676, 2023

    work page 2023

  4. [4]

    Wildfire danger prediction and understanding with deep learning,

    S. Kondylatos, I. Prapas, M. Ronco, I. Papoutsis, G. Camps-Valls, M. Piles, M.-Á. Fernández- Torres, and N. Carvalhais, “Wildfire danger prediction and understanding with deep learning,” Geophysical Research Letters, vol. 49, no. 17, p. e2022GL099368, 2022. e2022GL099368 2022GL099368

    work page 2022

  5. [5]

    Location-aware adaptive normalization: A deep learning approach for wildfire danger forecasting,

    M. H. Shams Eddin, R. Roscher, and J. Gall, “Location-aware adaptive normalization: A deep learning approach for wildfire danger forecasting,”IEEE Transactions on Geoscience and Remote Sensing, vol. 61, pp. 1–18, 2023

    work page 2023

  6. [6]

    Uncertainty-aware deep learning for wildfire danger forecasting,

    S. Kondylatos, G. Camps-Valls, and I. Papoutsis, “Uncertainty-aware deep learning for wildfire danger forecasting,” no. arXiv:2509.25017, 2025. arXiv:2509.25017 [cs]

    work page arXiv 2025

  7. [7]

    Deploying initial attack resources for wildfire suppression: spatial coordination, budget constraints, and capacity constraints,

    Y . Lee, J. S. Fried, H. J. Albers, and R. G. Haight, “Deploying initial attack resources for wildfire suppression: spatial coordination, budget constraints, and capacity constraints,”Canadian Journal of Forest Research, vol. 43, p. 56–65, Jan. 2013

    work page 2013

  8. [8]

    Understanding imbalanced semantic segmentation through neural collapse,

    Z. Zhong, J. Cui, Y . Yang, X. Wu, X. Qi, X. Zhang, and J. Jia, “Understanding imbalanced semantic segmentation through neural collapse,” in2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 19550–19559, 2023

    work page 2023

  9. [9]

    Focal loss for dense object detection,

    T.-Y . Lin, P. Goyal, R. Girshick, K. He, and P. Dollár, “Focal loss for dense object detection,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 42, no. 2, pp. 318–327, 2020

    work page 2020

  10. [10]

    Carion, F

    N. Carion, F. Massa, G. Synnaeve, N. Usunier, A. Kirillov, and S. Zagoruyko,End-to-End Object Detection with Transformers, vol. 12346 ofLecture Notes in Computer Science, p. 213–229. Cham: Springer International Publishing, 2020

    work page 2020

  11. [11]

    Rethinking counting and localization in crowds: A purely point-based framework,

    Q. Song, C. Wang, Z. Jiang, Y . Wang, Y . Tai, C. Wang, J. Li, F. Huang, and Y . Wu, “Rethinking counting and localization in crowds: A purely point-based framework,” inProceedings of the IEEE/CVF international conference on computer vision, pp. 3365–3374, 2021

    work page 2021

  12. [12]

    The hungarian method for the assignment problem,

    H. W. Kuhn, “The hungarian method for the assignment problem,”Naval Research Logistics (NRL), vol. 52, 1955

    work page 1955

  13. [13]

    The New VIIRS 375m active fire detection data product: Algorithm description and initial assessment,

    W. Schroeder, P. Oliva, L. Giglio, and I. A. Csiszar, “The New VIIRS 375m active fire detection data product: Algorithm description and initial assessment,”Remote Sensing of Environment, vol. 143, pp. 85–96, 2014

    work page 2014

  14. [14]

    The era5 global reanaly- sis,

    H. Hersbach, B. Bell, P. Berrisford, S. Hirahara, A. Horányi, J. Muñoz-Sabater, J. Nicolas, C. Peubey, R. Radu, D. Schepers, A. Simmons, C. Soci, S. Abdalla, X. Abellan, G. Balsamo, P. Bechtold, G. Biavati, J. Bidlot, M. Bonavita, G. De Chiara, P. Dahlgren, D. Dee, M. Dia- mantakis, R. Dragani, J. Flemming, R. Forbes, M. Fuentes, A. Geer, L. Haimberger, S...

    work page 1999

  15. [15]

    Quality assessment of proba-v lai, fapar and fcover collection 300 m products of copernicus global land service,

    B. Fuster, J. Sánchez-Zapero, F. Camacho, V . García-Santos, A. Verger, R. Lacaze, M. Weiss, F. Baret, and B. Smets, “Quality assessment of proba-v lai, fapar and fcover collection 300 m products of copernicus global land service,”Remote Sensing, vol. 12, no. 6, 2020. 11

    work page 2020

  16. [16]

    Geomorphons — a pattern recognition approach to classifica- tion and mapping of landforms,

    J. Jasiewicz and T. F. Stepinski, “Geomorphons — a pattern recognition approach to classifica- tion and mapping of landforms,”Geomorphology, vol. 182, pp. 147–156, 2013

    work page 2013

  17. [17]

    Height above the nearest drainage – a hydrologically relevant new terrain model,

    A. Nobre, L. Cuartas, M. Hodnett, C. Rennó, G. Rodrigues, A. Silveira, M. Waterloo, and S. Saleska, “Height above the nearest drainage – a hydrologically relevant new terrain model,” Journal of Hydrology, vol. 404, no. 1, pp. 13–29, 2011

    work page 2011

  18. [18]

    Next day wildfire spread: A machine learning dataset to predict wildfire spreading from remote-sensing data,

    F. Huot, R. L. Hu, N. Goyal, T. Sankar, M. Ihme, and Y .-F. Chen, “Next day wildfire spread: A machine learning dataset to predict wildfire spreading from remote-sensing data,”IEEE Transactions on Geoscience and Remote Sensing, vol. 60, pp. 1–13, 2022

    work page 2022

  19. [19]

    Estimating next day’s forest fire risk via a complete machine learning methodology,

    A. Apostolakis, S. Girtsou, G. Giannopoulos, N. S. Bartsotas, and C. Kontoes, “Estimating next day’s forest fire risk via a complete machine learning methodology,”Remote Sensing, vol. 14, no. 5, 2022

    work page 2022

  20. [20]

    A global probability-of-fire (pof) forecast,

    J. R. McNorton, F. Di Giuseppe, E. Pinnington, M. Chantry, and C. Barnard, “A global probability-of-fire (pof) forecast,”Geophysical Research Letters, vol. 51, no. 12, p. e2023GL107929, 2024. e2023GL107929 2023GL107929

    work page 2024

  21. [21]

    Data-driven wildfire spread modeling of european wildfires using a spatiotemporal graph neural network,

    M. Rösch, M. Nolde, T. Ullmann, and T. Riedlinger, “Data-driven wildfire spread modeling of european wildfires using a spatiotemporal graph neural network,”Fire, vol. 7, no. 6, 2024

    work page 2024

  22. [22]

    Wildfire spreading prediction using multimodal data and deep neural network approach,

    D. Shadrin, S. Illarionova, F. Gubanov, K. Evteeva, M. Mironenko, I. Levchunets, R. Belousov, and E. Burnaev, “Wildfire spreading prediction using multimodal data and deep neural network approach,”Scientific Reports, vol. 14, p. 2606, Jan. 2024

    work page 2024

  23. [23]

    Integration of a deep-learning- based fire model into a global land surface model,

    R. Son, T. Stacke, V . Gayler, J. E. M. S. Nabel, R. Schnur, L. Alonso, C. Requena-Mesa, A. J. Winkler, S. Hantson, S. Zaehle, U. Weber, and N. Carvalhais, “Integration of a deep-learning- based fire model into a global land surface model,”Journal of Advances in Modeling Earth Systems, vol. 16, no. 1, p. e2023MS003710, 2024. e2023MS003710 2023MS003710

    work page 2024

  24. [24]

    Deformable {detr}: Deformable transformers for end-to-end object detection,

    X. Zhu, W. Su, L. Lu, B. Li, X. Wang, and J. Dai, “Deformable {detr}: Deformable transformers for end-to-end object detection,” inInternational Conference on Learning Representations, 2021

    work page 2021

  25. [25]

    DAB-DETR: Dy- namic anchor boxes are better queries for DETR,

    S. Liu, F. Li, H. Zhang, X. Yang, X. Qi, H. Su, J. Zhu, and L. Zhang, “DAB-DETR: Dy- namic anchor boxes are better queries for DETR,” inInternational Conference on Learning Representations, 2022

    work page 2022

  26. [26]

    Detection transformer with stable matching,

    S. Liu, T. Ren, J. Chen, Z. Zeng, H. Zhang, F. Li, H. Li, J. Huang, H. Su, J. Zhu, and L. Zhang, “Detection transformer with stable matching,” in2023 IEEE/CVF International Conference on Computer Vision (ICCV), pp. 6468–6477, 2023

    work page 2023

  27. [27]

    Z. Zhao, Q. Xue, Y . He, Y . Bai, X. Wei, and Y . Gong,Projecting Points to Axes: Oriented Object Detection via Point-Axis Representation, vol. 15086 ofLecture Notes in Computer Science, p. 161–179. Cham: Springer Nature Switzerland, 2025

    work page 2025

  28. [28]

    Maptr: Structured modeling and learning for online vectorized hd map construction,

    B. Liao, S. Chen, X. Wang, T. Cheng, Q. Zhang, W. Liu, and C. Huang, “Maptr: Structured modeling and learning for online vectorized hd map construction,” Jan. 2023. arXiv:2208.14437 [cs]

    work page arXiv 2023

  29. [29]

    Earthformer: Exploring space-time transformers for earth system forecasting,

    Z. Gao, X. Shi, H. Wang, Y . Zhu, Y . B. Wang, M. Li, and D.-Y . Yeung, “Earthformer: Exploring space-time transformers for earth system forecasting,” inAdvances in Neural Information Processing Systems(S. Koyejo, S. Mohamed, A. Agarwal, D. Belgrave, K. Cho, and A. Oh, eds.), vol. 35, pp. 25390–25403, Curran Associates, Inc., 2022

    work page 2022

  30. [30]

    Highway Networks

    R. K. Srivastava, K. Greff, and J. Schmidhuber, “Highway networks,” Nov. 2015. arXiv:1505.00387 [cs]

    work page Pith review arXiv 2015

  31. [31]

    Film: visual reasoning with a general conditioning layer,

    E. Perez, F. Strub, H. de Vries, V . Dumoulin, and A. Courville, “Film: visual reasoning with a general conditioning layer,” inProceedings of the Thirty-Second AAAI Conference on Artificial Intelligence and Thirtieth Innovative Applications of Artificial Intelligence Con- ference and Eighth AAAI Symposium on Educational Advances in Artificial Intelligence...

    work page 2018

  32. [32]

    Efficientdet: Scalable and efficient object detection,

    M. Tan, R. Pang, and Q. V . Le, “Efficientdet: Scalable and efficient object detection,” in2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), (Seattle, WA, USA), p. 10778–10787, IEEE, 2020

    work page 2020

  33. [33]

    Multi-decadal trends and variability in burned area from the fifth version of the global fire emissions database (gfed5),

    Y . Chen, J. Hall, D. van Wees, N. Andela, S. Hantson, L. Giglio, G. R. van der Werf, D. C. Morton, and J. T. Randerson, “Multi-decadal trends and variability in burned area from the fifth version of the global fire emissions database (gfed5),”Earth System Science Data, vol. 15, no. 11, pp. 5227–5259, 2023

    work page 2023

  34. [34]

    Oksuz, B

    K. Oksuz, B. C. Cam, E. Akbas, and S. Kalkan,Localization Recall Precision (LRP): A New Performance Metric for Object Detection, vol. 11211 ofLecture Notes in Computer Science, p. 521–537. Cham: Springer International Publishing, 2018

    work page 2018

  35. [35]

    Spatiotemporal wildfire modeling through point processes with moderate and extreme marks,

    J. Koh, F. Pimont, J.-L. Dupuy, and T. Opitz, “Spatiotemporal wildfire modeling through point processes with moderate and extreme marks,”The Annals of Applied Statistics, vol. 17, no. 1, pp. 560 – 582, 2023

    work page 2023

  36. [36]

    Point-process based bayesian modeling of space–time struc- tures of forest fire occurrences in mediterranean france,

    T. Opitz, F. Bonneu, and E. Gabriel, “Point-process based bayesian modeling of space–time struc- tures of forest fire occurrences in mediterranean france,”Spatial Statistics, vol. 40, p. 100429,

    work page

  37. [37]

    Space-Time Modeling of Rare Events and Environmental Risks: METMA Conference

    work page

  38. [38]

    Towards accurate one-stage object detection with ap-loss,

    K. Chen, J. Li, W. Lin, J. See, J. Wang, L. Duan, Z. Chen, C. He, and J. Zou, “Towards accurate one-stage object detection with ap-loss,” in2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 5114–5122, 2019

    work page 2019

  39. [39]

    Searching parameterized ap loss for object detection,

    T. Chenxin, Z. Li, X. Zhu, G. Huang, Y . Liu, and j. dai, “Searching parameterized ap loss for object detection,” inAdvances in Neural Information Processing Systems(M. Ranzato, A. Beygelzimer, Y . Dauphin, P. Liang, and J. W. Vaughan, eds.), vol. 34, pp. 22021–22033, Curran Associates, Inc., 2021

    work page 2021

  40. [40]

    Rank-detr for high quality object detection,

    Y . Pu, W. Liang, Y . Hao, Y . YUAN, Y . Yang, C. Zhang, H. Hu, and G. Huang, “Rank-detr for high quality object detection,” inAdvances in Neural Information Processing Systems(A. Oh, T. Naumann, A. Globerson, K. Saenko, M. Hardt, and S. Levine, eds.), vol. 36, pp. 16100–16113, Curran Associates, Inc., 2023

    work page 2023

  41. [41]

    Deep high-resolution representation learning for visual recognition,

    J. Wang, K. Sun, T. Cheng, B. Jiang, C. Deng, Y . Zhao, D. Liu, Y . Mu, M. Tan, X. Wang, W. Liu, and B. Xiao, “Deep high-resolution representation learning for visual recognition,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 43, p. 3349–3364, Oct. 2021. 13 A Feature List Table 4: Input features used by WISP, grouped by model featu...

    work page 2021