pith. machine review for the scientific record. sign in

arxiv: 2605.06273 · v1 · submitted 2026-05-07 · 💻 cs.CV · cs.AR

Recognition: unknown

On-Orbit Real-Time Wildfire Detection Under On-Board Constraints

Matthias R\"otzer , Veronika P\"ortge , Martin Ickerott , Jayendra Praveen Kumar Chorapalli , Dimitri Scheftelowitsch , Max Bereczky , Dmitry Rashkovetsky , Sai Manoj Appalla
show 1 more author
Julia Gottfriedsen
Authors on Pith no claims yet

Pith reviewed 2026-05-08 13:32 UTC · model grok-4.3

classification 💻 cs.CV cs.AR
keywords wildfire detectionon-orbit processingself-supervised learningMWIR imagerysatellite constellationedge AIdense masked autoencodingreal-time inference
0
0 comments X

The pith

Dense masked autoencoding pretraining produces sub-megabyte models that detect sub-pixel wildfires from single-band MWIR satellite imagery at real-time speeds on orbit.

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

The paper presents an operational wildfire detection system running directly on a nine-satellite thermal constellation. It shows that pretraining compact networks with dense masked autoencoding lets them meet strict limits on model size, inference latency, and alert timing while handling single-pixel thermal anomalies in highly imbalanced data. The work replaces multi-band contextual thresholding with learned representations that operate on uncalibrated MWIR imagery. If the results hold, faster on-board alerts become possible without expanding satellite hardware.

Core claim

DenseMAE pretraining enables compact downstream models on the latency-accuracy Pareto frontier: the fastest SSL-pretrained model reaches 0.640 test AP and 0.69 event-level Fire-F1 at 65.34 ms per batch and 0.52 MB engine size without pruning. The best configuration achieves 0.699 AP and 0.744 Fire-F1 below 1 MB while outperforming a supervised baseline of 0.650 AP under the same constraints.

What carries the argument

Dense masked autoencoding (DenseMAE) pretraining on proprietary nine-satellite MWIR single-band imagery, which produces dense representations that support lightweight fine-tuned pixel-level fire detectors.

If this is right

  • Models fit the sub-megabyte footprint and sub-150 ms TensorRT FP16 inference budget on an NVIDIA Jetson Xavier NX.
  • The full pipeline targets end-to-end alerts under 10 minutes from satellite overpass to ground communication.
  • Detection works on uncalibrated single-band MWIR data at 200 m resolution despite frequent sub-pixel fires and extreme class imbalance.
  • SSL-pretrained models surpass supervised baselines when both are constrained to the same size and speed limits.

Where Pith is reading between the lines

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

  • The same pretraining approach could be tested on other single-band thermal tasks such as volcanic hot-spot monitoring.
  • Reduced dependence on ground processing may shorten response times for other time-critical satellite observations.
  • Adding limited multi-band inputs could raise accuracy further if the model size budget still permits.

Load-bearing premise

The nine-satellite proprietary MWIR dataset used for pretraining and testing represents the full range of real operational conditions and that pixel-level AP and event-level Fire-F1 reliably predict on-orbit performance.

What would settle it

A clear drop in AP or Fire-F1 when the same models are evaluated on independent public MWIR datasets collected from different satellites or geographic regions.

Figures

Figures reproduced from arXiv: 2605.06273 by Dimitri Scheftelowitsch, Dmitry Rashkovetsky, Jayendra Praveen Kumar Chorapalli, Julia Gottfriedsen, Martin Ickerott, Matthias R\"otzer, Max Bereczky, Sai Manoj Appalla, Veronika P\"ortge.

Figure 1
Figure 1. Figure 1: Random sample tiles of the MWIR band along with the manually annotated ground truth view at source ↗
Figure 2
Figure 2. Figure 2: DenseMAE pipeline. Staged convolutional encoder pre-trained with masked autoencoding on uncalibrated MWIR imagery, transferred to wildfire segmentation via a lightweight head and HR refinement. The f0 skip connection (purple) re-injects high￾resolution stem features for sub-pixel fire localisation. Architecture. While modern segmentation systems can be substantially larger, our focus is on the efficient re… view at source ↗
Figure 3
Figure 3. Figure 3: Latency–accuracy trade-off for deployment models (TensorRT FP16, 224×224, B=8). Points on the upper-left boundary are Pareto-optimal. Overall, DenseMAE representations are stable across ablations and consis￾tently transfer to downstream segmentation. Hybrid pretraining provides only marginal improvements in the best case, and we observe that applying distilla￾tion on masked regions only can degrade full-st… view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative comparison on challenging MWIR tiles (51 × 51 crops). Why refinement matters for small fires. While pixel-level AP is dominated by large fire clusters, the high-resolution refinement head is designed to improve localisation of small and isolated fire signatures that contribute few pixels. Consistent with this goal, we observe larger gains in event-level Fire-F1 ( view at source ↗
Figure 5
Figure 5. Figure 5: Comparison of active fire detections from our model and VIIRS. Panel a: MWIR measurements (in DNs) with evaluation of joint detections (green), model-only (blue), and VIIRS-only (orange). Panel b: Zoomed view of the MWIR measurements (indicated by the red rectangle). Panel c: Zoomed clustered detections. 6 Outlook While our results demonstrate the feasibility of lightweight deep learning for on-orbit wildf… view at source ↗
read the original abstract

We present a deployed system for on-orbit wildfire detection aboard a nine-satellite commercial thermal infrared constellation, operating under demanding joint constraints: sub-megabyte model footprint, sub-150 ms per-batch TensorRT FP16 inference on an NVIDIA Jetson Xavier NX, and an end-to-end alert pipeline targeting under 10 minutes from satellite overpass to fire event communication. The system operates on uncalibrated mid-wave infrared (MWIR) single-band imagery at 200 m ground sampling distance, where fires frequently appear as sub-pixel or single-pixel thermal anomalies under extreme class imbalance -- challenges not addressed by the contextual thermal-thresholding pipelines (MODIS, VIIRS) that currently dominate operational fire monitoring. We present an empirical study of lightweight dense representation learning for this regime using a proprietary nine-satellite MWIR dataset. We compare dense masked autoencoding (DenseMAE) and a hybrid DenseMAE+EMA (exponential moving average) distillation variant, and evaluate representations via linear probing and full-distribution pixel-level average precision (AP) under extreme class imbalance. DenseMAE pretraining enables compact downstream models on the latency-accuracy Pareto frontier: our fastest SSL-pretrained model achieves 0.640 test AP and 0.69 event-level Fire-F1 with 65.34 ms latency per batch and a 0.52 MB engine, without pruning or compression. The best configuration reaches 0.699 AP and 0.744 Fire-F1 below 1 MB, outperforming a supervised baseline (0.650 AP) under comparable constraints.

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

3 major / 2 minor

Summary. The paper presents a deployed on-orbit wildfire detection system for a nine-satellite MWIR constellation that uses DenseMAE (and DenseMAE+EMA) pretraining to produce compact models meeting sub-1 MB footprint and sub-150 ms TensorRT latency constraints. It reports empirical results on a proprietary dataset showing the best SSL-pretrained model reaching 0.699 pixel-level AP and 0.744 event-level Fire-F1, outperforming a supervised baseline at 0.650 AP, with the fastest variant at 0.640 AP / 0.69 Fire-F1 using a 0.52 MB engine.

Significance. If the empirical claims hold under operational conditions, the work would demonstrate a practical path for self-supervised pretraining to deliver latency-accuracy Pareto-optimal models for real-time on-board thermal anomaly detection, addressing a gap between standard contextual thresholding methods and edge-deployable deep models. The explicit focus on uncalibrated single-band MWIR, extreme imbalance, and end-to-end alert latency is a strength.

major comments (3)
  1. [Abstract] Abstract: The central performance claims (0.699 AP / 0.744 Fire-F1 beating supervised baseline at 0.650 AP) rest on a proprietary nine-satellite MWIR dataset whose size, positive-pixel ratio, event counts, train/test partitioning, and safeguards against event overlap between pretraining and downstream evaluation are not reported. Without these statistics it is impossible to determine whether the 0.049 AP delta reflects DenseMAE or dataset-specific artifacts, leakage, or evaluation choices.
  2. [Abstract] Abstract: The supervised baseline comparison provides no information on whether the baseline uses the identical backbone family, the same data augmentations, or the same training schedule as the SSL-pretrained models. This omission prevents attribution of the reported gains specifically to DenseMAE pretraining rather than other experimental factors.
  3. [Abstract] Abstract: No description is given of how pixel-level AP is computed under the stated extreme class imbalance (e.g., whether it uses standard COCO-style AP, any spatial weighting, or per-event aggregation), nor whether spatial autocorrelation in 200 m GSD MWIR imagery was mitigated; this directly affects interpretability of the 0.640–0.699 AP numbers.
minor comments (2)
  1. [Abstract] The abstract refers to 'linear probing and full-distribution pixel-level average precision' without defining the probing protocol or the exact AP implementation used.
  2. [Abstract] The latency figure (65.34 ms per batch) should specify batch size and whether it includes preprocessing or post-processing steps in the end-to-end pipeline.

Simulated Author's Rebuttal

3 responses · 1 unresolved

We thank the referee for the constructive and detailed comments. We address each major comment point-by-point below, indicating where we will revise the manuscript for improved clarity and completeness.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The central performance claims (0.699 AP / 0.744 Fire-F1 beating supervised baseline at 0.650 AP) rest on a proprietary nine-satellite MWIR dataset whose size, positive-pixel ratio, event counts, train/test partitioning, and safeguards against event overlap between pretraining and downstream evaluation are not reported. Without these statistics it is impossible to determine whether the 0.049 AP delta reflects DenseMAE or dataset-specific artifacts, leakage, or evaluation choices.

    Authors: We agree that greater transparency on dataset characteristics would aid interpretability. Due to the proprietary nature of the nine-satellite MWIR dataset, we cannot release exact numerical values for size, positive-pixel ratio, or event counts. However, we will revise the manuscript to add a dedicated paragraph in the Dataset subsection describing key properties (extreme imbalance with fire pixels <<0.1%, multi-year coverage across nine satellites), the partitioning approach (temporal splits by overpass date combined with spatial separation to eliminate event overlap), and explicit safeguards (no shared fire events or imagery between the pretraining corpus and the downstream evaluation set). These additions provide the necessary context to assess potential leakage or artifacts while respecting confidentiality. revision: partial

  2. Referee: [Abstract] Abstract: The supervised baseline comparison provides no information on whether the baseline uses the identical backbone family, the same data augmentations, or the same training schedule as the SSL-pretrained models. This omission prevents attribution of the reported gains specifically to DenseMAE pretraining rather than other experimental factors.

    Authors: We apologize for the insufficient detail in the current draft. The supervised baseline uses the identical lightweight backbone architecture, the same MWIR-specific data augmentations (random horizontal/vertical flips, rotations, and intensity scaling), and the exact same training schedule, optimizer, and hyperparameters as the fine-tuning stage of the SSL models. These equivalences are stated in Section 4 but will be made explicit with a cross-reference in the revised Results and a brief clarification in the abstract to ensure the performance delta can be attributed to DenseMAE pretraining. revision: yes

  3. Referee: [Abstract] Abstract: No description is given of how pixel-level AP is computed under the stated extreme class imbalance (e.g., whether it uses standard COCO-style AP, any spatial weighting, or per-event aggregation), nor whether spatial autocorrelation in 200 m GSD MWIR imagery was mitigated; this directly affects interpretability of the 0.640–0.699 AP numbers.

    Authors: We will expand the Evaluation Metrics section with a clear description: pixel-level AP follows the standard COCO-style computation of average precision from the precision-recall curve, treating every pixel as an independent instance with no spatial weighting or per-event aggregation (the latter is handled separately via event-level Fire-F1 using connected components). No explicit correction for spatial autocorrelation at 200 m GSD was applied, consistent with standard practice in dense prediction benchmarks; we will note this limitation and its implications for interpretation in the revised text. revision: yes

standing simulated objections not resolved
  • Exact numerical statistics for the proprietary dataset (size, positive-pixel ratio, event counts) due to confidentiality constraints.

Circularity Check

0 steps flagged

No circularity: empirical performance metrics derived from held-out evaluation

full rationale

The paper presents an empirical study of DenseMAE pretraining on a proprietary MWIR dataset, followed by downstream evaluation via linear probing and pixel-level AP/Fire-F1 on held-out test data. No derivation chain, equations, or first-principles predictions are claimed; reported numbers (e.g., 0.699 AP, 0.744 Fire-F1) are direct measurements from model training and inference under stated constraints. The central claim reduces to experimental results rather than any self-definitional mapping, fitted-input renaming, or self-citation load-bearing step. The proprietary dataset raises external verification concerns but does not create internal circularity in the reported chain.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The central claim rests on empirical performance measured on a proprietary dataset; no explicit free parameters are introduced beyond standard neural-network training choices, and no new physical entities are postulated.

axioms (2)
  • domain assumption MWIR single-band imagery at 200 m GSD contains detectable sub-pixel thermal anomalies corresponding to wildfires
    Invoked in the problem formulation and evaluation metrics.
  • domain assumption Linear probing and pixel-level AP under extreme class imbalance are appropriate proxies for operational fire detection performance
    Used to evaluate the learned representations.

pith-pipeline@v0.9.0 · 5624 in / 1482 out tokens · 82118 ms · 2026-05-08T13:32:48.062797+00:00 · methodology

discussion (0)

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

Reference graph

Works this paper leans on

24 extracted references · 17 canonical work pages

  1. [1]

    In: High Spatio- Temporal-Spectral Thermal Remote Sensing, pp

    Cawse-Nicholson, K., Abrams, M., Buongiorno, M.F., Freeborn, D., Gamet, P., Gottfriedsen, J., Goullioud, R., Halverson, G., Holmes, T., Hook, S., et al.: Current and Future High-Resolution Multi-Channel Thermal Imagers. In: High Spatio- Temporal-Spectral Thermal Remote Sensing, pp. 37–60. CRC Press (2026) 3

    2026

  2. [2]

    Journal of Geophysical Research: Atmospheres119(2), 803–816 (2014).https://doi.org/https://doi.org/10

    Csiszar, I., Schroeder, W., Giglio, L., Ellicott, E., Vadrevu, K.P., Justice, C.O., Wind, B.: Active fires from the Suomi NPP Visible Infrared Imaging Radiometer Suite: Product status and first evaluation results. Journal of Geophysical Research: Atmospheres119(2), 803–816 (2014).https://doi.org/https://doi.org/10. 1002/2013JD020453,https://agupubs.online...

    2014

  3. [3]

    International Journal of Wild- land Fire25(1), 48–61 (Dec 2015).https://doi.org/10.1071/wf15090,http: //dx.doi.org/10.1071/WF1509013

    Dickinson, M.B., Hudak, A.T., Zajkowski, T., Loudermilk, E.L., Schroeder, W., Ellison, L., Kremens, R.L., Holley, W., Martinez, O., Paxton, A., Bright, B.C., O’Brien, J.J., Hornsby, B., Ichoku, C., Faulring, J., Gerace, A., Peterson, D., Mauceri, J.: Measuring radiant emissions from entire prescribed fires with ground, airborne and satellite sensors – RxC...

    work page doi:10.1071/wf15090 2012

  4. [4]

    Inter- national Journal of Remote Sensing17(2), 419–424 (Jan 1996).https://doi

    Flasse, S.P., Ceccato, P.: A contextual algorithm for AVHRR fire detection. Inter- national Journal of Remote Sensing17(2), 419–424 (Jan 1996).https://doi. org/10.1080/01431169608949018,https://www.tandfonline.com/doi/full/ 10.1080/014311696089490183

    work page doi:10.1080/01431169608949018 1996

  5. [5]

    Fire6(5) 16 M

    Ghali, R., Akhloufi, M.A.: Deep Learning Approaches for Wildland Fires Using Satellite Remote Sensing Data: Detection, Mapping, and Prediction. Fire6(5) 16 M. Rötzer et al. (2023).https://doi.org/10.3390/fire6050192,https://www.mdpi.com/2571- 6255/6/5/1923

    work page doi:10.3390/fire6050192 2023

  6. [6]

    https://doi.org/10.5067/VIIRS/VJ114IMG.002,https://doi.org/10.5067/ VIIRS/VJ114IMG.002, accessed: 2025-12-22 13

    Giglio, L.: Viirs/jpss1 active fires 6-min l2 swath 375m v002 (vj114img) (2024). https://doi.org/10.5067/VIIRS/VJ114IMG.002,https://doi.org/10.5067/ VIIRS/VJ114IMG.002, accessed: 2025-12-22 13

    work page doi:10.5067/viirs/vj114img.002 2024

  7. [7]

    Remote Sensing of Environment178, 31–41 (Jun 2016).https://doi.org/10.1016/j.rse.2016.02.054,https://linkinghub

    Giglio, L., Schroeder, W., Justice, C.O.: The collection 6 MODIS active fire detec- tion algorithm and fire products. Remote Sensing of Environment178, 31–41 (Jun 2016).https://doi.org/10.1016/j.rse.2016.02.054,https://linkinghub. elsevier.com/retrieve/pii/S00344257163008273

    work page doi:10.1016/j.rse.2016.02.054 2016

  8. [8]

    Holder, C.J., Shafique, M.: On Efficient Real-Time Semantic Segmentation: A Sur- vey (2022),https://arxiv.org/abs/2206.086056

    work page arXiv 2022

  9. [9]

    Hong,Z.,Tang,Z.,Pan,H.,Zhang,Y.,Zheng,Z.,Zhou,R.,Ma,Z.,Zhang,Y.,Han, Y., Wang, J., Yang, S.: Active Fire Detection Using a Novel Convolutional Neural Network Based on Himawari-8 Satellite Images. Frontiers in Environmental Sci- enceV olume 10 - 2022(2022).https://doi.org/10.3389/fenvs.2022.794028, https://www.frontiersin.org/journals/environmental- scienc...

    work page doi:10.3389/fenvs.2022.794028 2022

  10. [10]

    Remote Sensing of En- vironment237, 111600 (2020).https://doi.org/https://doi.org/10.1016/ j.rse.2019.111600,https://www.sciencedirect.com/science/article/pii/ S003442571930620013

    Li, F., Zhang, X., Kondragunta, S., Schmidt, C.C., Holmes, C.D.: A prelimi- nary evaluation of GOES-16 active fire product using Landsat-8 and VIIRS ac- tive fire data, and ground-based prescribed fire records. Remote Sensing of En- vironment237, 111600 (2020).https://doi.org/https://doi.org/10.1016/ j.rse.2019.111600,https://www.sciencedirect.com/science...

    work page arXiv 2020

  11. [11]

    Re- mote Sensing Applications: Society and Environment41, 101868 (2026).https: //doi.org/https://doi.org/10.1016/j.rsase.2026.101868,https://www

    Liu, Z., Zhang, Q., Zhang, B., Zhu, J.: Validation of Himawari-8/9 10-minute wildfire products: Comparisons with MODIS and VIIRS from 2015 to 2023. Re- mote Sensing Applications: Society and Environment41, 101868 (2026).https: //doi.org/https://doi.org/10.1016/j.rsase.2026.101868,https://www. sciencedirect.com/science/article/pii/S235293852600001713

    work page doi:10.1016/j.rsase.2026.101868 2015

  12. [12]

    Remote Sensing of Environment160, 144– 155 (2015).https://doi.org/https://doi.org/10.1016/j.rse.2015.01.010, https://www.sciencedirect.com/science/article/pii/S003442571500019X13

    Oliva, P., Schroeder, W.: Assessment of VIIRS 375m active fire detection prod- uct for direct burned area mapping. Remote Sensing of Environment160, 144– 155 (2015).https://doi.org/https://doi.org/10.1016/j.rse.2015.01.010, https://www.sciencedirect.com/science/article/pii/S003442571500019X13

    work page doi:10.1016/j.rse.2015.01.010 2015

  13. [13]

    Remote Sensing14(4) (2022).https://doi.org/10.3390/rs14040992, https://www.mdpi.com/2072-4292/14/4/9923

    Rostami, A., Shah-Hosseini, R., Asgari, S., Zarei, A., Aghdami-Nia, M., Homay- ouni,S.:ActiveFireDetectionfromLandsat-8ImageryUsingDeepMultipleKernel Learning. Remote Sensing14(4) (2022).https://doi.org/10.3390/rs14040992, https://www.mdpi.com/2072-4292/14/4/9923

    work page doi:10.3390/rs14040992 2022

  14. [14]

    Algorithm Theoretical Basis Document Version 2.5, NOAA/NESDIS/STAR (2012) 3

    Schmidt, C.C., Hoffman, J., Prins, E., Lindstrom, S.: GOES-R Advanced Baseline Imager (ABI) Algorithm Theoretical Basis Document for Fire/Hot Spot Characterization. Algorithm Theoretical Basis Document Version 2.5, NOAA/NESDIS/STAR (2012) 3

    2012

  15. [15]

    National Oceanic and Atmospheric Administration: Washington, DC, USA (2020) 13

    Schroeder, W., Giglio, L., Csiszar, I., Tsidulko, M.: Algorithm theoretical basis document for NOAA NDE VIIRS I-band (375m) active fire. National Oceanic and Atmospheric Administration: Washington, DC, USA (2020) 13

    2020

  16. [16]

    Remote Sensing of Environment143, 85–96 (Mar 2014).https://doi.org/10

    Schroeder, W., Oliva, P., Giglio, L., Csiszar, I.A.: The New VIIRS 375 m ac- tive fire detection data product: Algorithm description and initial assessment. Remote Sensing of Environment143, 85–96 (Mar 2014).https://doi.org/10. 1016/j.rse.2013.12.008,https://linkinghub.elsevier.com/retrieve/pii/ S00344257130044833

    2014

  17. [17]

    Seneviratne, S., Zhang, X., Adnan, M., Badi, W., Dereczynski, C., Di Luca, A., Ghosh, S., Iskandar, I., Kossin, J., Lewis, S., Otto, F., Pinto, I., Satoh, M., Vicente- Serrano, S., Wehner, M., Zhou, B.: Weather and climate extreme events in a Real-Time On-Orbit Wildfire Segmentation 17 changing climate. In: Masson-Delmotte, V., Zhai, P., Pirani, A., Conno...

    work page doi:10.1017/9781009157896.013 2021

  18. [18]

    Journal of Sensors2022(1), 8044390 (2022).https://doi.org/https://doi.org/10.1155/2022/8044390, https://onlinelibrary.wiley.com/doi/abs/10.1155/2022/80443903

    Seydi, S.T., Saeidi, V., Kalantar, B., Ueda, N., Halin, A.A.: Fire-Net: A Deep Learning Framework for Active Forest Fire Detection. Journal of Sensors2022(1), 8044390 (2022).https://doi.org/https://doi.org/10.1155/2022/8044390, https://onlinelibrary.wiley.com/doi/abs/10.1155/2022/80443903

    work page doi:10.1155/2022/8044390 2022

  19. [19]

    In: Proceedings of the 36th International Conference on Machine Learn- ing (ICML)

    Tan, M., Le, Q.V.: Efficientnet: Rethinking model scaling for convolutional neural networks. In: Proceedings of the 36th International Conference on Machine Learn- ing (ICML). Proceedings of Machine Learning Research, vol. 97, pp. 6105–6114. PMLR (2019) 4

    2019

  20. [20]

    Remote Sensing14(1) (2022).https://doi

    Wang, Z., Yang, P., Liang, H., Zheng, C., Yin, J., Tian, Y., Cui, W.: Semantic Segmentation and Analysis on Sensitive Parameters of Forest Fire Smoke Using Smoke-Unet and Landsat-8 Imagery. Remote Sensing14(1) (2022).https://doi. org/10.3390/rs14010045,https://www.mdpi.com/2072-4292/14/1/453

    work page doi:10.3390/rs14010045 2022

  21. [21]

    Remote Sensing of Environment267, 112694 (Dec 2021).https://doi.org/10.1016/j.rse.2021

    Wooster, M.J., Roberts, G.J., Giglio, L., Roy, D.P., Freeborn, P.H., Boschetti, L., Justice, C., Ichoku, C., Schroeder, W., Davies, D., Smith, A.M., Setzer, A., Csiszar, I., Strydom, T., Frost, P., Zhang, T., Xu, W., De Jong, M.C., Johnston, J.M., Ellison, L., Vadrevu, K., Sparks, A.M., Nguyen, H., McCarty, J., Tanpipat, V., Schmidt, C., San-Miguel-Ayanz,...

    work page doi:10.1016/j.rse.2021 2021

  22. [22]

    Science of Remote Sensing13, 100366 (2026).https://doi.org/https://doi.org/10

    Xu, W., Wooster, M.J., He, J., Meraner, A., Gomez-Dans, J., Liu, Z., Trigo, I.F., Dutra, E.: Major improvements in spaceborne early fire detection and small-fire FRP retrieval with the meteosat third generation flexible combined imager. Science of Remote Sensing13, 100366 (2026).https://doi.org/https://doi.org/10. 1016/j.srs.2026.100366,https://www.scienc...

    work page arXiv 2026

  23. [23]

    Remote Sensing14(17) (2022).https://doi.org/10.3390/ rs14174347,https://www.mdpi.com/2072-4292/14/17/43473

    Zhao, Y., Ban, Y.: GOES-R Time Series for Early Detection of Wildfires with Deep GRU-Network. Remote Sensing14(17) (2022).https://doi.org/10.3390/ rs14174347,https://www.mdpi.com/2072-4292/14/17/43473

    2022

  24. [24]

    UNet++: A nested U-Net architecture for medical image segmentation,

    Zhou, Z., Siddiquee, M.M.R., Tajbakhsh, N., Liang, J.: Unet++: A nested u-net ar- chitecture for medical image segmentation. arXiv preprint arXiv:1807.10165 (2018) 4

    work page arXiv 2018