Uncertainty-Aware Wildfire Smoke Density Classification from Satellite Imagery via CBAM-Augmented EfficientNet with Evidential Deep Learning

Ranjith Chodavarapu

arxiv: 2605.15894 · v1 · pith:XQLXBDPCnew · submitted 2026-05-15 · 💻 cs.CV · cs.AI

Uncertainty-Aware Wildfire Smoke Density Classification from Satellite Imagery via CBAM-Augmented EfficientNet with Evidential Deep Learning

Ranjith Chodavarapu This is my paper

Pith reviewed 2026-05-20 19:08 UTC · model grok-4.3

classification 💻 cs.CV cs.AI

keywords wildfire smoke classificationsatellite imageryevidential deep learninguncertainty quantificationCBAM attentionEfficientNetepistemic uncertaintyaleatoric uncertainty

0 comments

The pith

A CBAM-augmented EfficientNet with evidential deep learning classifies satellite wildfire smoke into light moderate or heavy while estimating epistemic and aleatoric uncertainty directly.

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

The paper builds a model that takes satellite image patches of wildfires and assigns each to one of three smoke-density classes. It augments a pre-trained EfficientNet-B3 backbone with a CBAM attention module and replaces the usual classifier with an evidential deep learning head that outputs Dirichlet concentration parameters. From those parameters the model produces both the class probabilities and two separate uncertainty values: vacuity for epistemic uncertainty and dissonance for aleatoric uncertainty. This setup lets the system flag ambiguous transitional cases and report rising uncertainty when image quality drops, all in a single forward pass. Emergency response and air-quality applications would benefit because they receive not only a severity label but an explicit reliability signal for each prediction.

Core claim

On 16298 real satellite patches the CBAM-augmented EfficientNet-B3 with evidential head reaches 93.8 percent weighted accuracy and 91.1 percent unweighted accuracy with an expected calibration error of 0.0274. Retaining only the 50 percent of patches with lowest uncertainty lifts accuracy to 96.7 percent. As image quality is degraded uncertainty rises monotonically, vacuity serves as a scan-quality indicator, and the Moderate class records the highest mean vacuity of 0.187, confirming that the model recognizes transitional smoke boundaries as the most epistemically uncertain. CBAM attention maps localize to structurally distinctive regions and t-SNE embeddings show clear separation between L

What carries the argument

Evidential deep learning head that predicts Dirichlet concentration parameters to produce class probabilities together with vacuity (epistemic) and dissonance (aleatoric) uncertainty values in one forward pass.

If this is right

The model reaches 93.8 percent weighted test accuracy and 0.0274 expected calibration error on real satellite patches.
Selective prediction on the most certain 50 percent of patches raises accuracy to 96.7 percent.
Uncertainty values increase monotonically as image quality degrades, allowing vacuity to serve as a practical scan-quality measure.
The Moderate class exhibits the highest mean epistemic uncertainty of 0.187, matching its transitional character.
CBAM attention maps focus on structurally distinctive scene regions while t-SNE shows clear separation of Light and Heavy classes.

Where Pith is reading between the lines

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

Operational wildfire systems could route high-vacuity Moderate patches to human review or additional sensors first.
The monotonic uncertainty response to degradation could be used to automatically discard or re-acquire poor-quality satellite passes.
The same single-pass uncertainty decomposition might transfer to other satellite tasks that involve gradual transitions, such as cloud or flood mapping.
Selective prediction could be combined with cost-sensitive routing so that low-certainty alerts trigger cheaper verification steps.

Load-bearing premise

The evidential head correctly extracts epistemic uncertainty as vacuity and aleatoric uncertainty as dissonance from the Dirichlet parameters, and the Moderate class labels genuinely mark transitional smoke conditions expected to carry the highest epistemic uncertainty.

What would settle it

Progressively degrade the satellite patches by adding noise or lowering resolution and check whether reported vacuity increases at each step; separately compare mean vacuity across the three classes to test whether Moderate remains highest.

Figures

Figures reproduced from arXiv: 2605.15894 by Ranjith Chodavarapu.

**Figure 2.** Figure 2: Uncertainty decomposition. Left: vacuity ( [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗

**Figure 3.** Figure 3: Selective prediction: accuracy (blue) improves from 91.1% to 96.7% as uncertain patches are filtered. [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

**Figure 4.** Figure 4: t-SNE of EfficientNet+CBAM features (3,000 patches). Light (blue) and Heavy (red) occupy well [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗

**Figure 5.** Figure 5: Per-class accuracy, vacuity, and dissonance. The moderate class (orange) exhibits the highest [PITH_FULL_IMAGE:figures/full_fig_p011_5.png] view at source ↗

**Figure 6.** Figure 6: Confusion matrix (counts and normalised). Light–Light (0.97) and Heavy–Heavy (0.95) are correctly [PITH_FULL_IMAGE:figures/full_fig_p014_6.png] view at source ↗

**Figure 7.** Figure 7: CBAM spatial attention maps for Light, Moderate, and Heavy severity patches. Attention localises to [PITH_FULL_IMAGE:figures/full_fig_p015_7.png] view at source ↗

**Figure 8.** Figure 8: Spatial epistemic uncertainty maps per severity class (sliding window inference, stride=32, win [PITH_FULL_IMAGE:figures/full_fig_p016_8.png] view at source ↗

read the original abstract

Rapid and accurate wildfire smoke severity assessment from satellite images is essential for emergency response, air quality modeling, and human health risk management. Existing deep learning approaches treat smoke detection as a binary task, producing point estimates without any measure of prediction confidence. We propose a probabilistic framework to categorize a satellite patch into Light, Moderate, and Heavy severity classes and to provide decomposed epistemic and aleatoric uncertainty in a single forward pass. Our architecture uses the backbone of a pre-trained EfficientNet-B3 and a CBAM module with an evidential deep learning head that predicts Dirichlet concentration parameters, directly estimating vacuity (epistemic) and dissonance (aleatoric) without Monte Carlo sampling. Evaluated on 16,298 real satellite patches derived from the Wildfire Detection dataset, our model achieves 93.8% weighted test accuracy (91.1% unweighted) with ECE=0.0274. Selective prediction retaining the most certain 50% of patches achieves 96.7% accuracy. As image quality degrades, uncertainty increases monotonically, and vacuity is a practical scan quality measure. The Moderate class represents transitional smoke conditions that exhibit the highest epistemic uncertainty (mean vacuity = 0.187), confirming the model correctly identifies ambiguous smoke boundary regions. CBAM spatial attention maps localize to structurally distinctive scene regions, and t-SNE demonstrates the clear cluster separation of Light and Heavy smoke.

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.

Desk Editor's Note private letter to a colleague

Straightforward application of EfficientNet plus CBAM and evidential head to smoke severity classification that delivers usable accuracy and uncertainty numbers but rests its key interpretation on the moderate labels being genuinely transitional.

read the letter

The core of this paper is a practical pipeline: EfficientNet-B3 backbone, CBAM attention, and an evidential head that outputs Dirichlet parameters for light/moderate/heavy smoke on satellite patches. On the 16k-patch test set it reaches 93.8% weighted accuracy, 91.1% unweighted, ECE of 0.0274, and 96.7% accuracy when keeping only the top 50% most certain predictions. Uncertainty rises monotonically with image degradation and the moderate class shows the highest mean vacuity at 0.187. Attention maps and t-SNE separation are also reported. That is the actual contribution: a working system with decomposed uncertainty in one forward pass for an environmental monitoring task. The numbers are concrete and the selective-prediction gain is the kind of thing that matters for deployment. The evidential formulation follows the standard approach from prior work, so the novelty is in the domain application and the empirical checks rather than new theory. The main soft spot is the interpretation that moderate patches represent transitional boundary cases and therefore should carry the highest epistemic uncertainty. If the dataset labels were derived from heuristics or thresholds that do not cleanly map to density transitions, then the vacuity result loses its supporting story even while raw accuracy and calibration stay intact. Without ablations or details on label generation it is hard to judge how much the CBAM or evidential head actually drive the gains versus the backbone alone. This paper is for applied researchers in remote sensing or wildfire response who need uncertainty-aware classifiers they can run on real imagery. A reader looking for a deployable baseline with selective prediction and basic uncertainty diagnostics will find usable results here. It is coherent enough and the empirical claims are specific enough that it deserves a serious referee rather than a desk reject; the uncertainty decomposition and selective-prediction numbers are worth checking in review even if the label interpretation needs tightening.

Referee Report

1 major / 2 minor

Summary. The paper introduces a CBAM-augmented EfficientNet-B3 backbone with an evidential deep learning head that outputs Dirichlet concentration parameters for three-class (Light/Moderate/Heavy) wildfire smoke density classification from satellite patches. It reports 93.8% weighted and 91.1% unweighted accuracy, ECE=0.0274 on 16,298 real patches, 96.7% accuracy under selective prediction on the top 50% most certain samples, monotonic uncertainty growth under degradation, and highest mean vacuity (0.187) for the Moderate class, interpreted as successful capture of epistemic uncertainty at smoke boundaries. CBAM attention maps and t-SNE visualizations support interpretability.

Significance. If the central empirical results and uncertainty decomposition hold after addressing label validation, the work offers a computationally efficient, single-pass method for uncertainty-aware smoke severity assessment that could support real-time emergency response and air-quality applications. The decomposed vacuity/dissonance outputs and selective-prediction gains are practically useful; the approach avoids Monte Carlo sampling while providing attention-based localization. Credit is due for the concrete accuracy/ECE/selective numbers on a sizable real dataset and the monotonic degradation test.

major comments (1)

[Abstract and Results (per-class uncertainty analysis)] Abstract and Results section on per-class uncertainty: the claim that Moderate patches exhibit the highest epistemic uncertainty (mean vacuity = 0.187) and thereby confirm that the model 'correctly identifies ambiguous smoke boundary regions' is load-bearing for the uncertainty-aware contribution. This interpretation requires that the derived Moderate labels genuinely correspond to transitional density states rather than annotation heuristics, imbalance artifacts, or subjective thresholds. Without additional validation (e.g., expert re-labeling of Moderate samples or controlled boundary-case experiments), the reported monotonic uncertainty increase and selective-prediction improvement lose their supporting interpretation for epistemic capture, even if raw accuracy and ECE remain high.

minor comments (2)

[Methods] Training protocol and hyper-parameter details for the evidential head (e.g., annealing schedule for Dirichlet parameters, loss weighting) are only sketched; explicit values or a reproducibility checklist would strengthen the manuscript.
[Results (visualizations)] Figure 3 (t-SNE) and attention-map figures would benefit from quantitative cluster-separation metrics (e.g., silhouette score) to support the qualitative claim of 'clear cluster separation'.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the thoughtful and detailed review. The concern about the evidential support for interpreting Moderate-class vacuity as evidence of epistemic capture at true transitional boundaries is well-taken. We address this point directly below and indicate the revisions we will make.

read point-by-point responses

Referee: [Abstract and Results (per-class uncertainty analysis)] Abstract and Results section on per-class uncertainty: the claim that Moderate patches exhibit the highest epistemic uncertainty (mean vacuity = 0.187) and thereby confirm that the model 'correctly identifies ambiguous smoke boundary regions' is load-bearing for the uncertainty-aware contribution. This interpretation requires that the derived Moderate labels genuinely correspond to transitional density states rather than annotation heuristics, imbalance artifacts, or subjective thresholds. Without additional validation (e.g., expert re-labeling of Moderate samples or controlled boundary-case experiments), the reported monotonic uncertainty increase and selective-prediction improvement lose their supporting interpretation for epistemic capture, even if raw accuracy and ECE remain high.

Authors: We agree that the interpretive claim in the abstract and results section is stronger than the supporting evidence for label provenance allows. The Moderate labels are taken directly from the source Wildfire Detection dataset and were produced by its original annotators using visual density heuristics; we do not possess independent expert re-annotation or controlled boundary experiments. In the revised manuscript we will (i) add a concise description of label provenance in the Dataset subsection, (ii) replace the word “confirming” with “consistent with the hypothesis that” in both the abstract and the per-class uncertainty paragraph, and (iii) insert a short limitations paragraph acknowledging possible subjectivity or threshold effects in the Moderate class. The numerical results—weighted accuracy, ECE, selective-prediction gains, and monotonic uncertainty growth under degradation—do not rely on this interpretation and will remain unchanged. We therefore view the revision as partial: textual clarification and tempering of language rather than new empirical validation. revision: partial

Circularity Check

0 steps flagged

No significant circularity; empirical evaluation on external dataset with standard evidential formulation

full rationale

The paper's central results consist of training a CBAM-augmented EfficientNet-B3 with an evidential deep learning head on the Wildfire Detection dataset and reporting test accuracy, ECE, selective prediction performance, and uncertainty trends on held-out patches. The evidential head follows the standard Dirichlet concentration parameterization from prior literature rather than deriving it internally. No load-bearing step reduces a claimed prediction or uniqueness result to a fitted parameter or self-citation by construction. The interpretation of Moderate-class vacuity is an empirical observation on the given labels, not a definitional equivalence. The derivation chain is therefore self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

1 free parameters · 0 axioms · 0 invented entities

The central claims rest on the standard assumptions of supervised deep learning (i.i.d. training and test patches, accurate human-provided severity labels) plus the evidential deep learning framework's mapping from Dirichlet parameters to uncertainty quantities; no new axioms or invented entities are introduced.

free parameters (1)

Dirichlet concentration parameters
Output by the evidential head and optimized during training on the 16,298 patches to produce the reported accuracy and uncertainty values.

pith-pipeline@v0.9.0 · 5788 in / 1392 out tokens · 62000 ms · 2026-05-20T19:08:40.699828+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

21 extracted references · 21 canonical work pages

[1]

A. Aaba. Wildfire detection from satellite images. Kaggle dataset, 2022. URL https://www.kaggle. com/datasets/abdelghaniaaba/wildfire-detection-from-satellite-images

work page 2022
[2]

Amini, W

A. Amini, W. Schwarting, A. Soleimany, and D. Rus. Deep evidential regression.Advances in neural information processing systems, 33:14927–14937, 2020. 11

work page 2020
[3]

Audebert, B

N. Audebert, B. Le Saux, and S. Lef`evre. Beyond rgb: Very high resolution urban remote sensing with multimodal deep networks.ISPRS journal of photogrammetry and remote sensing, 140:20–32, 2018

work page 2018
[4]

R. Ba, C. Chen, J. Yuan, W. Song, and S. Lo. Smokenet: Satellite smoke scene detection using convolutional neural network with spatial and channel-wise attention.Remote Sensing, 11(14), 2019. ISSN 2072-4292. doi: 10.3390/rs11141702. URLhttps://www.mdpi.com/2072-4292/11/14/1702

work page doi:10.3390/rs11141702 2019
[5]

J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei. Imagenet: A large-scale hierarchical image database. In2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248–255, 2009. doi: 10.1109/CVPR.2009.5206848

work page doi:10.1109/cvpr.2009.5206848 2009
[6]

Gal and Z

Y . Gal and Z. Ghahramani. Dropout as a bayesian approximation: Representing model uncertainty in deep learning. Ininternational conference on machine learning, pages 1050–1059. PMLR, 2016

work page 2016
[7]

Ghali and M

R. Ghali and M. A. Akhloufi. Deep learning approaches for wildland fires using satellite remote sensing data: Detection, mapping, and prediction.Fire, 6(5), 2023. ISSN 2571-6255. doi: 10.3390/fire6050192. URLhttps://www.mdpi.com/2571-6255/6/5/192

work page doi:10.3390/fire6050192 2023
[8]

F. C. Ghesu, B. Georgescu, A. Mansoor, Y . Yoo, E. Gibson, R. Vishwanath, A. Balachandran, J. M. Balter, Y . Cao, R. Singh, et al. Quantifying and leveraging predictive uncertainty for medical image assessment. Medical Image Analysis, 68:101855, 2021

work page 2021
[9]

J. Hu, L. Shen, and G. Sun. Squeeze-and-excitation networks. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 7132–7141, 2018

work page 2018
[10]

Hutchison, B

K. Hutchison, B. Iisager, T. Kopp, and J. Jackson. Distinguishing aerosols from clouds in global, multi- spectral satellite data with automated cloud classification algorithms.Journal of Atmospheric and Oceanic Technology - J ATMOS OCEAN TECHNOL, 25, 04 2008. doi: 10.1175/2007JTECHA1004.1

work page doi:10.1175/2007jtecha1004.1 2008
[11]

P. Jain, S. C. Coogan, S. G. Subramanian, M. Crowley, S. Taylor, and M. D. Flannigan. A review of machine learning applications in wildfire science and management.Environmental Reviews, 28(4):478–505, 2020. ISSN 1208-6053. doi: https://doi.org/10.1139/er-2020-0019. URL https://www.sciencedirect. com/science/article/pii/S1208605320000374

work page doi:10.1139/er-2020-0019 2020
[12]

M. W. Jones, A. Smith, R. Betts, J. G. Canadell, I. C. Prentice, and C. L. Qu ´er´e. Climate change increases the risk of wildfires. Technical report, Tyndall Centre for Climate Change Research, 2020. URL http://www.jstor.org/stable/resrep51248

work page 2020
[13]

Y . J. Kaufman, C. O. Justice, L. P. Flynn, J. D. Kendall, E. M. Prins, L. Giglio, D. E. Ward, W. P. Menzel, and A. W. Setzer. Potential global fire monitoring from eos-modis.Journal of Geophysical Research: Atmospheres, 103(D24):32215–32238, 1998. doi: https://doi.org/10.1029/98JD01644. URL https://agupubs.onlinelibrary.wiley.com/doi/abs/10.1029/98JD01644

work page doi:10.1029/98jd01644 1998
[14]

Laden, J

F. Laden, J. Schwartz, F. E. Speizer, and D. W. Dockery. Reduction in fine particulate air pollution and mortality: Extended follow-up of the harvard six cities study.American Journal of Respiratory and Critical Care Medicine, 173(6):667–672, 03 2006. doi: 10.1164/rccm.200503-443OC. URL https: //doi.org/10.1164/rccm.200503-443OC

work page doi:10.1164/rccm.200503-443oc 2006
[15]

Lakshminarayanan, A

B. Lakshminarayanan, A. Pritzel, and C. Blundell. Simple and scalable predictive uncertainty estimation using deep ensembles.Advances in neural information processing systems, 30, 2017

work page 2017
[16]

Sensoy, L

M. Sensoy, L. Kaplan, and M. Kandemir. Evidential deep learning to quantify classification uncertainty. Advances in neural information processing systems, 31, 2018. 12

work page 2018
[17]

Tan and Q

M. Tan and Q. Le. Efficientnet: Rethinking model scaling for convolutional neural networks. InInterna- tional conference on machine learning, pages 6105–6114. PMLR, 2019

work page 2019
[18]

C. J. Tucker. Red and photographic infrared linear combinations for monitoring vegetation.Remote Sensing of Environment, 8(2):127–150, 1979. ISSN 0034-4257. doi: https://doi.org/10.1016/0034-4257(79)90013-0. URLhttps://www.sciencedirect.com/science/article/pii/0034425779900130

work page doi:10.1016/0034-4257(79)90013-0 1979
[19]

S. Woo, J. Park, J.-Y . Lee, and I. S. Kweon. Cbam: Convolutional block attention module. InProceedings of the European conference on computer vision (ECCV), pages 3–19, 2018

work page 2018
[20]

Zhang, I

C. Zhang, I. Sargent, X. Pan, H. Li, A. Gardiner, J. Hare, and P. M. Atkinson. Joint deep learning for land cover and land use classification.Remote Sensing of Environment, 221:173–187, 2019. ISSN 0034-

work page 2019
[21]

URL https://www.sciencedirect.com/ science/article/pii/S0034425718305236

doi: https://doi.org/10.1016/j.rse.2018.11.014. URL https://www.sciencedirect.com/ science/article/pii/S0034425718305236. 13 A Supplementary Figures This appendix provides additional visualizations. Figure 6 shows the full confusion matrix. Figure 7 shows CBAM attention maps per severity class. Figure 8 shows spatial uncertainty decomposition maps. Figure...

work page doi:10.1016/j.rse.2018.11.014 2018

[1] [1]

A. Aaba. Wildfire detection from satellite images. Kaggle dataset, 2022. URL https://www.kaggle. com/datasets/abdelghaniaaba/wildfire-detection-from-satellite-images

work page 2022

[2] [2]

Amini, W

A. Amini, W. Schwarting, A. Soleimany, and D. Rus. Deep evidential regression.Advances in neural information processing systems, 33:14927–14937, 2020. 11

work page 2020

[3] [3]

Audebert, B

N. Audebert, B. Le Saux, and S. Lef`evre. Beyond rgb: Very high resolution urban remote sensing with multimodal deep networks.ISPRS journal of photogrammetry and remote sensing, 140:20–32, 2018

work page 2018

[4] [4]

R. Ba, C. Chen, J. Yuan, W. Song, and S. Lo. Smokenet: Satellite smoke scene detection using convolutional neural network with spatial and channel-wise attention.Remote Sensing, 11(14), 2019. ISSN 2072-4292. doi: 10.3390/rs11141702. URLhttps://www.mdpi.com/2072-4292/11/14/1702

work page doi:10.3390/rs11141702 2019

[5] [5]

J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei. Imagenet: A large-scale hierarchical image database. In2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248–255, 2009. doi: 10.1109/CVPR.2009.5206848

work page doi:10.1109/cvpr.2009.5206848 2009

[6] [6]

Gal and Z

Y . Gal and Z. Ghahramani. Dropout as a bayesian approximation: Representing model uncertainty in deep learning. Ininternational conference on machine learning, pages 1050–1059. PMLR, 2016

work page 2016

[7] [7]

Ghali and M

R. Ghali and M. A. Akhloufi. Deep learning approaches for wildland fires using satellite remote sensing data: Detection, mapping, and prediction.Fire, 6(5), 2023. ISSN 2571-6255. doi: 10.3390/fire6050192. URLhttps://www.mdpi.com/2571-6255/6/5/192

work page doi:10.3390/fire6050192 2023

[8] [8]

F. C. Ghesu, B. Georgescu, A. Mansoor, Y . Yoo, E. Gibson, R. Vishwanath, A. Balachandran, J. M. Balter, Y . Cao, R. Singh, et al. Quantifying and leveraging predictive uncertainty for medical image assessment. Medical Image Analysis, 68:101855, 2021

work page 2021

[9] [9]

J. Hu, L. Shen, and G. Sun. Squeeze-and-excitation networks. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 7132–7141, 2018

work page 2018

[10] [10]

Hutchison, B

K. Hutchison, B. Iisager, T. Kopp, and J. Jackson. Distinguishing aerosols from clouds in global, multi- spectral satellite data with automated cloud classification algorithms.Journal of Atmospheric and Oceanic Technology - J ATMOS OCEAN TECHNOL, 25, 04 2008. doi: 10.1175/2007JTECHA1004.1

work page doi:10.1175/2007jtecha1004.1 2008

[11] [11]

P. Jain, S. C. Coogan, S. G. Subramanian, M. Crowley, S. Taylor, and M. D. Flannigan. A review of machine learning applications in wildfire science and management.Environmental Reviews, 28(4):478–505, 2020. ISSN 1208-6053. doi: https://doi.org/10.1139/er-2020-0019. URL https://www.sciencedirect. com/science/article/pii/S1208605320000374

work page doi:10.1139/er-2020-0019 2020

[12] [12]

M. W. Jones, A. Smith, R. Betts, J. G. Canadell, I. C. Prentice, and C. L. Qu ´er´e. Climate change increases the risk of wildfires. Technical report, Tyndall Centre for Climate Change Research, 2020. URL http://www.jstor.org/stable/resrep51248

work page 2020

[13] [13]

Y . J. Kaufman, C. O. Justice, L. P. Flynn, J. D. Kendall, E. M. Prins, L. Giglio, D. E. Ward, W. P. Menzel, and A. W. Setzer. Potential global fire monitoring from eos-modis.Journal of Geophysical Research: Atmospheres, 103(D24):32215–32238, 1998. doi: https://doi.org/10.1029/98JD01644. URL https://agupubs.onlinelibrary.wiley.com/doi/abs/10.1029/98JD01644

work page doi:10.1029/98jd01644 1998

[14] [14]

Laden, J

F. Laden, J. Schwartz, F. E. Speizer, and D. W. Dockery. Reduction in fine particulate air pollution and mortality: Extended follow-up of the harvard six cities study.American Journal of Respiratory and Critical Care Medicine, 173(6):667–672, 03 2006. doi: 10.1164/rccm.200503-443OC. URL https: //doi.org/10.1164/rccm.200503-443OC

work page doi:10.1164/rccm.200503-443oc 2006

[15] [15]

Lakshminarayanan, A

B. Lakshminarayanan, A. Pritzel, and C. Blundell. Simple and scalable predictive uncertainty estimation using deep ensembles.Advances in neural information processing systems, 30, 2017

work page 2017

[16] [16]

Sensoy, L

M. Sensoy, L. Kaplan, and M. Kandemir. Evidential deep learning to quantify classification uncertainty. Advances in neural information processing systems, 31, 2018. 12

work page 2018

[17] [17]

Tan and Q

M. Tan and Q. Le. Efficientnet: Rethinking model scaling for convolutional neural networks. InInterna- tional conference on machine learning, pages 6105–6114. PMLR, 2019

work page 2019

[18] [18]

C. J. Tucker. Red and photographic infrared linear combinations for monitoring vegetation.Remote Sensing of Environment, 8(2):127–150, 1979. ISSN 0034-4257. doi: https://doi.org/10.1016/0034-4257(79)90013-0. URLhttps://www.sciencedirect.com/science/article/pii/0034425779900130

work page doi:10.1016/0034-4257(79)90013-0 1979

[19] [19]

S. Woo, J. Park, J.-Y . Lee, and I. S. Kweon. Cbam: Convolutional block attention module. InProceedings of the European conference on computer vision (ECCV), pages 3–19, 2018

work page 2018

[20] [20]

Zhang, I

C. Zhang, I. Sargent, X. Pan, H. Li, A. Gardiner, J. Hare, and P. M. Atkinson. Joint deep learning for land cover and land use classification.Remote Sensing of Environment, 221:173–187, 2019. ISSN 0034-

work page 2019

[21] [21]

URL https://www.sciencedirect.com/ science/article/pii/S0034425718305236

doi: https://doi.org/10.1016/j.rse.2018.11.014. URL https://www.sciencedirect.com/ science/article/pii/S0034425718305236. 13 A Supplementary Figures This appendix provides additional visualizations. Figure 6 shows the full confusion matrix. Figure 7 shows CBAM attention maps per severity class. Figure 8 shows spatial uncertainty decomposition maps. Figure...

work page doi:10.1016/j.rse.2018.11.014 2018