Pith. sign in

REVIEW 3 major objections 6 minor 19 references

Fighting Fires from Space: Leveraging Vision Transformers for Enhanced Wildfire Detection and Characterization

T0 review · 3 major / 6 minor · reviewed 2026-08-16 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read A Swin-Unet vision transformer detects wildfires from Landsat-8 imagery with an IoU of 89.93%, slightly beating the best published CNN baseline, while a custom CNN U-Net remains the top model at 93.58%.

desk verdict A useful but flawed benchmark: the headline claim that a ViT beats the CNN baseline evaporates once you notice Table 1 mixes two different IoU definitions. read the letter →

arxiv 2504.13776 v1 pith:ND6WLDPB submitted 2025-04-18 cs.CV eess.IV

Aman Agarwal , James Gearon , Raksha Rank , Etienne Chenevert This is my paper · ORCID
classification cs.CVeess.IV
keywords wildfiredetectionvisiontransformerssemanticsegmentationLandsat-8U-NetremotesensingSwin-Unetdeeplearning
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper asks whether Vision Transformers—networks built on self-attention rather than convolution—can match or beat convolutional networks for wildfire segmentation in satellite imagery. It reports that one transformer variant, Swin-Unet, reaches an IoU (pixel-overlap score) of 89.93% on a manually annotated Landsat-8 test set, edging the best previously published CNN baseline at 89.00%. The authors' own CNN U-Net scores 93.58% IoU, the highest of any model, so the paper's conclusion is mixed: transformers are competitive and efficient to train, but a carefully tuned CNN is still the best detector on this data. This matters because early wildfire detection from satellites needs models that train quickly and use context beyond local image patches.

What carries the argument

The comparison hinges on three segmentation models that share a U-Net-style encoder-decoder skeleton. The CNN U-Net uses strided convolutions for downsampling and dilated convolutions (rates 1, 1, 2, 3) to widen its receptive field. TransUNet inserts a transformer encoder between CNN layers, while Swin-Unet replaces the CNN core with shifted-window self-attention, which computes attention within local windows that shift between layers to capture both fine detail and long-range context. This attention mechanism is what gives the ViT models their global view, and the reported results show it transfers to fire detection about as well as convolution does, with Swin's hierarchical windows faring better than TransUNet's plain transformer.

What would settle it

Re-run the best published CNN baseline and Swin-Unet on the same test images using one shared evaluation script with multiple random seeds and identical crops; if the IoU gap disappears or reverses across seeds, the claimed outperformance is not robust. Additionally, train Swin-Unet with spectral-imagery pre-training; if it still fails to exceed the custom U-Net's 93.58%, the paper's proposed remedy for the ViT gap is called into question.

Watch

Extended reading notes

Core claim

The central claim is an empirical ranking on one fixed dataset. On the manually annotated test images, Swin-Unet produces an IoU of 89.93%, which is 0.93 percentage points above the best baseline U-Net (89.00%); TransUNet does not surpass the baseline (87.49%). The paper's own U-Net reaches 93.58% IoU and also has the highest precision, leading the authors to state that well-tuned CNNs remain the best technique for wildfire detection even though ViTs are comparably capable. They attribute the ViT shortfall mainly to a mismatch between the everyday RGB images used for pre-training and the multispectral satellite images in the task.

Load-bearing premise

The comparison assumes the previously published CNN scores were obtained under the same evaluation protocol—same test split, same 224x224 center crop, same preprocessing—as the new models; if the baselines used different crops, resolutions, or splits, the reported improvement may be an artifact of evaluation rather than a real architectural gain.

Editorial extensions

If this is right

  • If the reported numbers hold, transformer-based segmentation is a credible alternative to CNNs for active wildfire detection, with accuracy close enough to be usable in practice.
  • The custom U-Net's 93.58% IoU indicates that convolutional architecture details—dilated and strided convolutions—still move performance more than the choice of attention versus convolution on this dataset.
  • Swin-Unet's margin over TransUNet suggests that hierarchical, shifted-window attention is better suited to satellite fire imagery than a single transformer encoder.
  • Because the paper describes ViTs as cheaper to train and easier to run in parallel, parity with CNNs would already make them attractive for near-real-time monitoring of large satellite archives.
  • The paper's own diagnosis points to a concrete next step: pre-training the transformer on spectral or remote-sensing imagery rather than natural RGB photos could close or reverse the gap to the custom U-Net.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Our inference: the 0.93 percentage-point advantage over the baseline is small enough that it could fall within run-to-run training variance, so the superiority of Swin-Unet over CNNs should be treated as provisional until replicated across multiple seeds and training runs.
  • Our inference: the more durable takeaway is operational, not architectural—if transformers really train faster, equal accuracy would translate into faster model updates as fresh satellite data arrives, which matters more for fire response than a fraction of an IoU point.
  • Our inference: the same comparison could be run on other multispectral fire datasets (for example, Sentinel-2 imagery) and on the intersection and voting masks from the same source; a consistent small Swin-Unet edge across sensors would argue for a real architectural effect rather than a quirk of one test set.
  • Our inference: the pre-training-mismatch explanation is directly testable by fine-tuning Swin-Unet from weights learned on aerial or multispectral imagery; if its IoU climbs past 93.58%, the paper's stated path to improvement gains support.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. The paper applies three semantic segmentation models—a custom U-Net, TransUNet, and Swin-Unet—to the Pereira et al. Landsat-8 wildfire detection dataset, training on algorithmically labeled data and testing on the manually annotated test set. It reports precision, recall, F-score, and IoU for each model and claims that Swin-Unet exceeds the best published CNN baseline by 0.93 percentage points, while the authors' own U-Net achieves the best overall performance. The manuscript includes a code repository and explicitly acknowledges that the Mask2Former experiment failed. The central comparative claim, however, is undermined by an internal inconsistency in the reported IoU values and by the lack of a same-protocol baseline comparison.

Significance. If the reported comparison were valid, the contribution would be a modest but useful empirical study: two ViT-based segmenters applied to a public wildfire benchmark, with the honest and non-obvious finding that a well-tuned CNN remains the best model. The paper ships a code link and builds on a previously published dataset, which are strengths. The conclusion that ViTs are competitive with CNNs for this task is plausible. However, the headline quantitative claim currently rests on an internally inconsistent IoU metric applied to the in-house models and on published baseline numbers obtained under a different protocol. The significance of the paper is therefore contingent on a corrected evaluation.

major comments (3)
  1. [§4, Table 1] The IoU column for the three in-house models is not the foreground Jaccard index implied by the same rows' precision and recall values. Under the standard binary Jaccard formula IoU = PR/(P+R−PR), Swin-Unet's P=88.28 and R=92.30 yield approximately 82.2% (reported 89.93%), TransUNet's P=88.46 and R=86.88 yield approximately 78.0% (reported 87.49%), and the authors' U-Net's P=93.37 and R=93.96 yield approximately 88.1% (reported 93.58%). The F-scores satisfy F=2PR/(P+R), so precision, recall, and F-score are mutually consistent; only the IoU column is anomalous for all in-house models, while the Pereira baseline rows satisfy the foreground Jaccard relation. This indicates that the in-house IoU was computed with a different aggregation, likely class-averaged foreground/background IoU, making the headline comparison of Swin-Unet 89.93% against the baseline 89.00% a comparison of two different metrics. Please define the IoU computation precisely, report foreground Jaccard index for all models, and confirm whether the Pereira baseline values are foreground Jaccard.
  2. [§3.3 and §4] The claim that Swin-Unet outperforms the Pereira et al. baselines assumes that the published metrics are directly comparable to the authors' evaluation pipeline. Section 3.3 describes training on 224×224 random crops, validation on a 224×224 center crop, AdamW, Dice loss, batch size 64, and 30–50 epochs; the Pereira baselines were not retrained or re-evaluated under this protocol. Differences in preprocessing, crop location, resolution, loss function, optimizer, and epoch count can therefore confound the reported 0.93 percentage-point advantage. To support the comparative claim, the baseline models should be re-evaluated under exactly the same preprocessing and evaluation protocol, or the claims should be restricted to comparisons among the in-house models.
  3. [§4] No repeated runs or error bars are reported. With a claimed advantage of only 0.93 percentage points, the difference is within the range of typical seed-to-seed and data-augmentation variability for segmentation models. Please report mean and standard deviation over at least three random seeds and, ideally, a paired comparison over the same test images, so the reader can judge whether the ViT-versus-CNN difference is statistically meaningful.
minor comments (6)
  1. [Abstract and §4] The abstract reports that the ViT outperforms the baseline by 0.92%, while §4 reports 0.93%; these numbers should be reconciled.
  2. [§3.3] The paper says models were trained for '30-50 epochs' but does not give the exact epoch count, final learning rate, or number of training images per model; include these details for reproducibility.
  3. [§4] IoU is defined only as 'checks for the overlap between predicted and ground truth masks'; the exact formula (foreground Jaccard versus mean IoU) and any binarization threshold must be stated.
  4. [Figure 5] The statement that the outputs are visually 'almost indistinguishable' is qualitative; if kept, support it with per-image quantitative IoU or error maps.
  5. [§2] The claim that ViTs 'require substantially less computational resources' is not evaluated in this paper and should be qualified or removed to avoid overstating the motivation.
  6. [References] Reference formatting is inconsistent, including 'UN2 [2022]' and 'Dewagan et al.'; these should be cleaned up.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the central claim is an empirical benchmark against a held-out manually annotated test set, with no fitted parameter renamed as a prediction and no load-bearing self-citation chain.

full rationale

The paper's central claim is an empirical comparison of trained segmentation models on a separate manually annotated test set (Section 3.1: "We use the algorithmically generated data masks for training and validating the models and keep the manually annotated data exclusively for testing"). No parameter is fitted to the headline metric; the IoU/F-score values in Table 1 are reported model outputs. The only author self-citation (Agarwal et al. [2021] for strided convolutions) is a design recommendation, explicitly not load-bearing because the authors state "We were unable to prove the specific cause of improvement via any ablation study due to time constraints." The skeptical observation that the in-house IoU column is numerically inconsistent with the reported precision/recall under foreground Jaccard is a potential metric-definition or correctness problem; it does not make the comparison circular, because the evaluation is still measurement against held-out data. Baseline numbers are imported from an external publication, not from a self-citation chain. No self-definitional, fitted-input-as-prediction, uniqueness-imported, or ansatz-smuggling pattern is present.

Assumptions & free parameters 5 free parameters · 3 assumptions · 0 invented entities

The paper introduces no new theoretical entities or parameters. All unknowns are standard ML hyperparameters chosen by hand. The key load-bearing assumptions are empirical: the quality of the training labels and the comparability of the baseline metrics.

free parameters (5)
  • learning_rate = 1e-3
    Initial learning rate chosen by hand with a one-cycle cosine schedule. Results are likely sensitive to this choice.
  • batch_size = 64
    Chosen by hand; affects optimization trajectory and batch statistics.
  • training_epochs = 30-50
    Variable range; no explicit early stopping criterion, which may affect final model quality.
  • unet_dilation_rates = [1,1,2,3]
    Hand-picked dilation schedule for the custom U-Net; the authors attribute performance gains to this without an ablation study.
  • augmentation_probabilities = 0.5 (flip, rotation), 0.2 (shear, erase)
    Chosen by hand; standard augmentations but probabilities are not tuned.
assumptions (3)
  • domain assumption Algorithmically generated training masks from Schroeder et al., Murphy et al., and Kumar and Roy are reliable enough to learn a model that transfers to manual ground truth.
    Training is supervised by these automated masks; if they are noisy or biased, the learned model may not generalize to the manually annotated test labels.
  • domain assumption The manually annotated test set is the same as in Pereira et al. and is evaluated consistently.
    Direct comparability of the reported metrics depends on this; the paper does not specify that the test preprocessing matches the baseline evaluation.
  • domain assumption Pre-trained weights from ImageNet (RGB) provide a useful initialization for 3-channel spectral inputs.
    The authors note this may be suboptimal, but it is used for Swin-Unet and TransUNet, and it is a background assumption for transfer learning.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Fighting Fires from Space: Leveraging Vision Transformers for Enhanced Wildfire Detection and Characterization." pith.science (2026). https://pith.science/paper/ND6WLDPB

@misc{pith2026250413776,
  author       = {Pith},
  title        = {Pith review of: Fighting Fires from Space: Leveraging Vision Transformers for Enhanced Wildfire Detection and Characterization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ND6WLDPB}},
  note         = {Machine review of arXiv:2504.13776}
}
read the original abstract

Wildfires are increasing in intensity, frequency, and duration across large parts of the world as a result of anthropogenic climate change. Modern hazard detection and response systems that deal with wildfires are under-equipped for sustained wildfire seasons. Recent work has proved automated wildfire detection using Convolutional Neural Networks (CNNs) trained on satellite imagery are capable of high-accuracy results. However, CNNs are computationally expensive to train and only incorporate local image context. Recently, Vision Transformers (ViTs) have gained popularity for their efficient training and their ability to include both local and global contextual information. In this work, we show that ViT can outperform well-trained and specialized CNNs to detect wildfires on a previously published dataset of LandSat-8 imagery. One of our ViTs outperforms the baseline CNN comparison by 0.92%. However, we find our own implementation of CNN-based UNet to perform best in every category, showing their sustained utility in image tasks. Overall, ViTs are comparably capable in detecting wildfires as CNNs, though well-tuned CNNs are still the best technique for detecting wildfire with our UNet providing an IoU of 93.58%, better than the baseline UNet by some 4.58%.

Figures

Figures reproduced from arXiv: 2504.13776 by the authors.

Figure 1
Figure 1. Publication counts containing the keyword “Wildfire Detection” in recent years [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Fire as visualized from three different bands of a Landsat image. Part (a), (b), and (c) show a range of bands [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. A high level representation of our UNet model architecture. Each transition stage applied two dilated [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Various data augmentation methods applied to the input images and masks. Each transformation was applied [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Inference results on some of the test-set images for each model. Visually, the results look almost indistin [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Inference result example for the Mask2Former instance detection model. This is one of its better results. [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

19 extracted references · 10 canonical work pages

  1. [1]

    Active fire detection in landsat-8 imagery: A large-scale dataset and a deep-learning study

    Gabriel Henrique de Almeida Pereira , Andre Minoro Fusioka, Bogdan Tomoyuki Nassu, and Rodrigo Minetto. Active fire detection in landsat-8 imagery: A large-scale dataset and a deep-learning study. ISPRS Journal of Photogrammetry and Remote Sensing, 178: 0 171--186, 2021. ISSN 0924-2716. doi:https://doi.org/10.1016/j.isprsjprs.2021.06.002. URL https://www....

    doi:10.1016/j.isprsjprs.2021.06.002 2021

  2. [2]

    https://www.grida.no/publications/769, 2022

    Spreading like wildfire: The rising threat of extraordinary landscape fires. https://www.grida.no/publications/769, 2022

    work page 2022

  3. [3]

    Cottrell, and Mai H

    Anshuman Dewangan, Yash Pande, Hans-Werner Braun, Frank Vernon, Ismael Perez, Ilkay Altintas, Garrison W. Cottrell, and Mai H. Nguyen. Figlib & smokeynet: Dataset and deep learning model for real-time wildland fire smoke detection. Remote Sensing, 14 0 (4): 0 1007, Feb 2022. ISSN 2072-4292. doi:10.3390/rs14041007. URL http://dx.doi.org/10.3390/rs14041007

    work page doi:10.3390/rs14041007 2022

  4. [4]

    An image is worth 16x16 words: Transformers for image recognition at scale

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020

    arXiv 2010

  5. [5]

    Vision transformers for remote sensing image classification

    Yakoub Bazi, Laila Bashmal, Mohamad M Al Rahhal, Reham Al Dayil, and Naif Al Ajlan. Vision transformers for remote sensing image classification. Remote Sensing, 13 0 (3): 0 516, 2021

    work page 2021

  6. [6]

    Wildfire detection from multisensor satellite imagery using deep semantic segmentation

    Dmitry Rashkovetsky, Florian Mauracher, Martin Langer, and Michael Schmitt. Wildfire detection from multisensor satellite imagery using deep semantic segmentation. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 14: 0 7001--7016, 2021. doi:10.1109/JSTARS.2021.3093625

    work page arXiv 2021

  7. [7]

    A review on early forest fire detection systems using optical remote sensing

    Panagiotis Barmpoutis, Periklis Papaioannou, Kosmas Dimitropoulos, and Nikos Grammalidis. A review on early forest fire detection systems using optical remote sensing. Sensors, 20 0 (22), 2020. ISSN 1424-8220. doi:10.3390/s20226442. URL https://www.mdpi.com/1424-8220/20/22/6442

    doi:10.3390/s20226442 2020

  8. [8]

    Active fire detection using landsat-8/oli data

    Wilfrid Schroeder, Patricia Oliva, Louis Giglio, Brad Quayle, Eckehard Lorenz, and Fabiano Morelli. Active fire detection using landsat-8/oli data. Remote sensing of environment, 185: 0 210--220, 2016

    work page 2016

Show all 19 references
  1. [9]

    Hotmap: Global hot target detection at moderate spatial resolution

    Sam W Murphy, Carlos Roberto de Souza Filho, Rob Wright, Giovanni Sabatino, and Rosa Correa Pabon. Hotmap: Global hot target detection at moderate spatial resolution. Remote Sensing of Environment, 177: 0 78--88, 2016

    2016

  2. [10]

    Global operational land imager landsat-8 reflectance-based active fire detection algorithm

    Sanath Sathyachandran Kumar and David P Roy. Global operational land imager landsat-8 reflectance-based active fire detection algorithm. International Journal of Digital Earth, 11 0 (2): 0 154--178, 2018

    2018

  3. [11]

    Training data-efficient image transformers & distillation through attention

    Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Herv \'e J \'e gou. Training data-efficient image transformers & distillation through attention. In International Conference on Machine Learning, pages 10347--10357. PMLR, 2021

    2021

  4. [12]

    Swin transformer: Hierarchical vision transformer using shifted windows

    Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012--10022, 2021

    2021

  5. [13]

    Cvt: Introducing convolutions to vision transformers

    Haiping Wu, Bin Xiao, Noel Codella, Mengchen Liu, Xiyang Dai, Lu Yuan, and Lei Zhang. Cvt: Introducing convolutions to vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22--31, 2021

    2021

  6. [14]

    Schwing, Alexander Kirillov, and Rohit Girdhar

    Bowen Cheng, Ishan Misra, Alexander G. Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation, 2021. URL https://arxiv.org/abs/2112.01527

    2021 arXiv

  7. [15]

    U-net: Convolutional networks for biomedical image segmentation

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention, pages 234--241. Springer, 2015

    2015

  8. [16]

    Transunet: Transformers make strong encoders for medical image segmentation

    Jieneng Chen, Yongyi Lu, Qihang Yu, Xiangde Luo, Ehsan Adeli, Yan Wang, Le Lu, Alan L Yuille, and Yuyin Zhou. Transunet: Transformers make strong encoders for medical image segmentation. arXiv preprint arXiv:2102.04306, 2021

    2021 arXiv

  9. [17]

    Transformers in computer vision

    Nikolas Adaloglou. Transformers in computer vision. https://theaisummer.com/, 2021

    2021

  10. [18]

    Swin-unet: Unet-like pure transformer for medical image segmentation

    Hu Cao, Yueyue Wang, Joy Chen, Dongsheng Jiang, Xiaopeng Zhang, Qi Tian, and Manning Wang. Swin-unet: Unet-like pure transformer for medical image segmentation. arXiv preprint arXiv:2105.05537, 2021

    2021 arXiv

  11. [19]

    Dilated volumetric network: an enhanced fully convolutional network for volumetric prostate segmentation from magnetic resonance imaging

    Aman Agarwal, Aditya Mishra, Madhushree Basavarajaiah, Priyanka Sharma, and Sudeep Tanwar. Dilated volumetric network: an enhanced fully convolutional network for volumetric prostate segmentation from magnetic resonance imaging. Pattern Recognition and Image Analysis, 31 0 (2)...

    2021

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.