pith. machine review for the scientific record. sign in

arxiv: 2604.18606 · v1 · submitted 2026-04-10 · 📡 eess.SP · cs.AI

Recognition: no theorem link

Thermal Anomaly Detection using Physics Aware Neuromorphic Networks: Comparison between Raw and L1C Sentinel-2 Data

Stephen Smith , Cormac Purcell , Gabriele Meoni , Roberto Del Prete , Zdenka Kuncic
Authors on Pith no claims yet

Pith reviewed 2026-05-10 18:07 UTC · model grok-4.3

classification 📡 eess.SP cs.AI
keywords thermal anomaly detectionneuromorphic networksSentinel-2raw L0 dataonboard processingbushfire detectionphysics-aware AI
0
0 comments X

The pith

A physics-aware neuromorphic network detects thermal anomalies in raw Sentinel-2 data with performance close to ground-processed products while meeting real-time onboard constraints.

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

The paper sets out to show that thermal anomalies such as those from bushfires or volcanoes can be identified directly from decompressed raw Level-0 sensor measurements instead of relying on computationally heavy ground preprocessing. It does this by introducing the Physics-Aware Neuromorphic Network, a lightweight model designed to cope with domain shifts, sensor drift, and scarce labels in raw data. The work compares results on raw measurements against standard L1C products and measures processing speed against satellite acquisition timing. A sympathetic reader would care because earlier onboard detection could shorten the delay between event occurrence and warning, limiting rapid escalation of damage.

Core claim

The Physics-Aware Neuromorphic Network achieves a Matthews Correlation Coefficient of 0.809 on decompressed raw L0 measurements and 0.875 on L1C products, with mean software processing latency of 2.44 seconds per granule below the 3.6-second Sentinel-2 acquisition window, and projects substantially lower latency and memory use when realized on neuromorphic hardware.

What carries the argument

The Physics-Aware Neuromorphic Network (PANN), a lightweight architecture that incorporates physical neural network principles to process raw sensor data directly for thermal anomaly classification.

If this is right

  • Thermal anomaly detection becomes possible without the full ground preprocessing chain, shortening the time from observation to alert.
  • Processing latency remains under the satellite's acquisition interval, enabling continuous real-time operation.
  • Memory footprint stays within realistic onboard limits for both software and projected hardware versions.
  • The same lightweight approach could support other event-detection tasks that currently require heavy preprocessing.

Where Pith is reading between the lines

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

  • Reducing reliance on ground stations could lower overall system latency and bandwidth demands for high-volume Earth observation missions.
  • If hardware execution times scale as projected, power budgets on small satellites might allow continuous anomaly monitoring without dedicated thermal sensors.
  • The gap between raw and processed accuracy (0.809 vs 0.875) suggests targeted calibration of the network on additional raw samples could close most of the difference without changing the architecture.

Load-bearing premise

The limited set of labeled training samples is representative of the full range of domain shifts, sensor drift, and radiometric inconsistencies that occur in real raw L0 satellite data.

What would settle it

Running the trained PANN on live raw Sentinel-2 L0 streams from an actual satellite or neuromorphic hardware testbed and comparing its Matthews Correlation Coefficient and latency against the reported software figures would confirm or refute the onboard feasibility claim.

Figures

Figures reproduced from arXiv: 2604.18606 by Cormac Purcell, Gabriele Meoni, Roberto Del Prete, Stephen Smith, Zdenka Kuncic.

Figure 1
Figure 1. Figure 1: Visualisation of the simulated nanowire network de [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Diagram of the model pipeline. Starting with the input image (left), the image is fist normalised and tiled into smaller [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 4
Figure 4. Figure 4: Effect of varying the threshold value on the differ [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 3
Figure 3. Figure 3: Confusion matrix for the raw (a) and L1C (b) datasets, [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 5
Figure 5. Figure 5: Three examples for fires (a) and volcanoes (b) for both the raw and L1C datasets. Each example shows an RGB image, using [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
read the original abstract

Damage caused by bushfires and volcanic eruptions escalates rapidly when detection is delayed, making fast and reliable early warning capabilities essential. Recent Earth Observation (EO) approaches have shown that thermal anomaly detection can be performed directly on decompressed Level-0 (L0) sensor data, avoiding computationally expensive preprocessing chains. However, direct exploitation of raw data remains challenging due to domain shift, sensor drift, radiometric inconsistencies, and the scarcity of labelled training samples. To address these challenges, this work proposes a Physics-Aware Neuromorphic Network (PANN) framework for onboard thermal anomaly detection. The proposed lightweight architecture, inspired by physical neural network principles and neuromorphic computing paradigms, is evaluated using two Sentinel-2 datasets: decompressed L0 with additional metadata (i.e. raw) and Level-1C (L1C). The PANN achieves a Matthews Correlation Coefficient (MCC) of $0.809$ on raw measurements, compared to $0.875$ when using ground-processed L1C products. The mean processing latency per L0 granule is $2.44 \pm 0.09~\mathrm{s}$, which is below the Sentinel-2 acquisition time of $3.6~\mathrm{s}$, demonstrating the feasibility of real-time, onboard processing. Furthermore, the projected execution time for the corresponding neuromorphic hardware instantiation is substantially lower at $0.1290 \pm 0.0002~\mathrm{s}$. Memory usage, including all necessary programs and packages, remains within realistic onboard constraints, with requirements of $0.673 \pm 0.007~\mathrm{Gb}$ for the software PANN and $0.393 \pm 0.004~\mathrm{Gb}$ for the estimated hardware realisation. Overall, these results indicate that PANN offers a promising pathway toward low-latency and resource-efficient onboard EO processing for thermal event detection.

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 a Physics-Aware Neuromorphic Network (PANN) framework for thermal anomaly detection directly on decompressed raw Level-0 Sentinel-2 data, addressing challenges of domain shift, sensor drift, radiometric inconsistencies, and limited labeled samples. It reports an MCC of 0.809 on raw L0 measurements versus 0.875 on ground-processed L1C products, a mean software processing latency of 2.44 ± 0.09 s per L0 granule (below the 3.6 s acquisition time), and projected neuromorphic hardware metrics of 0.129 s execution time and 0.393 Gb memory footprint, concluding that this enables feasible real-time onboard EO processing for early warning of bushfires and volcanic eruptions.

Significance. If the reported performance generalizes, this would represent a meaningful advance in onboard Earth observation by demonstrating lightweight, physics-informed neuromorphic processing that bypasses full ground preprocessing chains while meeting real-time latency constraints, with potential impact on disaster response applications.

major comments (2)
  1. [Results] Results section: The MCC of 0.809 on raw L0 data is reported without any information on the total number of labeled samples, train/test split ratios, cross-validation procedure, or explicit out-of-distribution testing (e.g., temporal or cross-granule splits) to address the domain shifts and radiometric inconsistencies highlighted in the abstract. This detail is load-bearing for the central onboard feasibility claim.
  2. [Methods] Methods section: No ablation study or quantitative analysis is provided to isolate the contribution of the physics-aware components in mitigating sensor drift or radiometric inconsistencies relative to a standard neuromorphic baseline, leaving the effectiveness of the 'physics-aware' design unverified for the stated challenges.
minor comments (1)
  1. [Abstract] Abstract: The phrase 'decompressed L0 with additional metadata (i.e. raw)' is unclear regarding exactly which metadata fields are provided and how they are incorporated into the PANN input.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive and detailed review. The comments identify important gaps in experimental documentation and validation that we will address in the revised manuscript.

read point-by-point responses

  1. Referee: [Results] Results section: The MCC of 0.809 on raw L0 data is reported without any information on the total number of labeled samples, train/test split ratios, cross-validation procedure, or explicit out-of-distribution testing (e.g., temporal or cross-granule splits) to address the domain shifts and radiometric inconsistencies highlighted in the abstract. This detail is load-bearing for the central onboard feasibility claim.

    Authors: We agree that these experimental details are essential for substantiating the robustness of the reported MCC of 0.809 on raw L0 data, particularly given the domain-shift and radiometric challenges noted in the abstract. The current manuscript does not provide the total number of labeled samples, train/test split ratios, cross-validation procedure, or explicit out-of-distribution testing results. In the revised version we will add a dedicated paragraph in the Results section (and update the Methods section as needed) that reports the dataset size, the train/test split ratios used, the cross-validation strategy, and quantitative results from temporal and cross-granule out-of-distribution splits. These additions will directly support the onboard feasibility claim. revision: yes

  2. Referee: [Methods] Methods section: No ablation study or quantitative analysis is provided to isolate the contribution of the physics-aware components in mitigating sensor drift or radiometric inconsistencies relative to a standard neuromorphic baseline, leaving the effectiveness of the 'physics-aware' design unverified for the stated challenges.

    Authors: We concur that an ablation study isolating the physics-aware elements would strengthen the manuscript. The present work compares PANN performance on raw L0 versus L1C data but does not include a direct quantitative comparison against a standard (non-physics-aware) neuromorphic baseline with metrics on sensor drift or radiometric inconsistency. In the revised manuscript we will add an ablation study in the Methods and Results sections that compares the full PANN architecture to a standard neuromorphic network baseline and reports the differential performance under simulated sensor drift and radiometric variations. revision: yes

Circularity Check

0 steps flagged

No circularity in derivation chain

full rationale

The paper reports empirical performance metrics obtained by training and evaluating a proposed PANN architecture on two Sentinel-2 datasets (raw L0 and L1C). The MCC values (0.809 and 0.875), latency figures (2.44 s software, 0.129 s projected hardware), and memory requirements are direct experimental outcomes from dataset runs rather than quantities derived from equations or parameters that loop back to the inputs. No self-definitional steps, fitted-input predictions, load-bearing self-citations, or uniqueness theorems appear in the abstract or described claims. The physics-aware inspiration is stated as a design principle, but the reported results remain independent validations against held-out data and do not reduce to tautological redefinitions of the training distribution.

Axiom & Free-Parameter Ledger

1 free parameters · 2 axioms · 0 invented entities

The central claim rests on trained neural network parameters fitted to limited labeled samples and on unverified projections for neuromorphic hardware performance; no independent evidence is provided for generalization under real sensor conditions.

free parameters (1)
  • PANN network weights and hyperparameters
    Trained on available labeled thermal anomaly samples to achieve the reported MCC values; these are fitted parameters central to the performance claims.
axioms (2)
  • domain assumption Raw L0 data with metadata can be processed by the lightweight PANN without prohibitive domain shift effects after training.
    Invoked when claiming feasibility on raw measurements despite noted challenges of sensor drift and radiometric inconsistencies.
  • ad hoc to paper Neuromorphic hardware will deliver the projected 0.129 s execution time and 0.393 Gb memory footprint.
    Used to support the real-time onboard conclusion but presented as an estimate without empirical validation in the abstract.

pith-pipeline@v0.9.0 · 5668 in / 1362 out tokens · 53184 ms · 2026-05-10T18:07:46.610438+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

30 extracted references · 30 canonical work pages · 1 internal anchor

  1. [1]

    Human exposure and sensitivity to globally extreme wildfire events.Nature ecology & evolution, 1(3): 0058, 2017

    David MJS Bowman, Grant J Williamson, John T Abat- zoglou, Crystal A Kolden, Mark A Cochrane, and Alis- tair MS Smith. Human exposure and sensitivity to globally extreme wildfire events.Nature ecology & evolution, 1(3): 0058, 2017. 1

    work page 2017

  2. [2]

    Vegetation fires in the anthropocene.Nature Reviews Earth & Environment, 1(10):500–515, 2020

    David MJS Bowman, Crystal A Kolden, John T Abatzoglou, Fay H Johnston, Guido R van der Werf, and Mike Flannigan. Vegetation fires in the anthropocene.Nature Reviews Earth & Environment, 1(10):500–515, 2020. 1

    work page 2020

  3. [3]

    Memristors for the curious outsiders.Technologies, 6(4):118, 2018

    Francesco Caravelli and Juan Pablo Carbajal. Memristors for the curious outsiders.Technologies, 6(4):118, 2018. 2

    work page 2018

  4. [4]

    Self-Organising Memristive Networks as Physical Learning Systems

    Francesco Caravelli, Gianluca Milano, Adam Z Stieg, Carlo Ricciardi, Simon Anthony Brown, and Zdenka Kuncic. Self- organising memristive networks as physical learning sys- tems.arXiv:2509.00747, 2025. 2

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  5. [5]

    Opportunities and challenges of on-board ai-based image recognition for small satellite earth observation missions.Advances in space research, 75(9):6734–6751, 2025

    Bharadwaj Chintalapati, Arthur Precht, Sougata Hanra, Rene Laufer, Marcus Liwicki, and Jens Eickhoff. Opportunities and challenges of on-board ai-based image recognition for small satellite earth observation missions.Advances in space research, 75(9):6734–6751, 2025. 2

    work page 2025

  6. [6]

    Enhancing maritime situational awareness through end-to-end onboard raw data analysis.IEEE Journal of Selected Topics in Ap- plied Earth Observations and Remote Sensing, 2025

    Roberto Del Prete, Manuel Salvoldi, Domenico Bar- retta, Nicolas Long ´ep´e, Gabriele Meoni, Arnon Karnieli, Maria Daniela Graziano, and Alfredo Renga. Enhancing maritime situational awareness through end-to-end onboard raw data analysis.IEEE Journal of Selected Topics in Ap- plied Earth Observations and Remote Sensing, 2025. 2

    work page 2025

  7. [7]

    Sentinel-2: Esa’s optical high-resolution mission for gmes operational services.Remote sensing of Environment, 120:25–36, 2012

    Matthias Drusch, Umberto Del Bello, S ´ebastien Carlier, Olivier Colin, Veronica Fernandez, Ferran Gascon, Bianca Hoersch, Claudia Isola, Paolo Laberinti, Philippe Martimort, et al. Sentinel-2: Esa’s optical high-resolution mission for gmes operational services.Remote sensing of Environment, 120:25–36, 2012. 1, 2

    work page 2012

  8. [8]

    Atmospheric and environmental impacts of volcanic particu- lates.Elements, 6(4):235–240, 2010

    Adam J Durant, Costanza Bonadonna, and Claire J Horwell. Atmospheric and environmental impacts of volcanic particu- lates.Elements, 6(4):235–240, 2010. 1

    work page 2010

  9. [9]

    Global wildland fire season severity in the 21st century.Forest Ecol- ogy and Management, 294:54–61, 2013

    Mike Flannigan, Alan S Cantin, William J De Groot, Mike Wotton, Alison Newbery, and Lynn M Gowman. Global wildland fire season severity in the 21st century.Forest Ecol- ogy and Management, 294:54–61, 2013. 1

    work page 2013

  10. [10]

    Sentinel-2 products specification document.Rapport technique, pages 4–7, 2015

    A Gatti and A Bertolini. Sentinel-2 products specification document.Rapport technique, pages 4–7, 2015. 1, 2

    work page 2015

  11. [11]

    Wildfire and prescribed burning impacts on air quality in the united states.Journal of the Air & Waste Management Association, 70(6):583–615, 2020

    Daniel A Jaffe, Susan M O’Neill, Narasimhan K Larkin, Amara L Holder, David L Peterson, Jessica E Halofsky, and Ana G Rappold. Wildfire and prescribed burning impacts on air quality in the united states.Journal of the Air & Waste Management Association, 70(6):583–615, 2020. 1

    work page 2020

  12. [12]

    Wilds: A benchmark of in-the- wild distribution shifts

    Pang Wei Koh, Shiori Sagawa, Henrik Marklund, Sang Michael Xie, Marvin Zhang, Akshay Balsubra- mani, Weihua Hu, Michihiro Yasunaga, Richard Lanas Phillips, Irena Gao, et al. Wilds: A benchmark of in-the- wild distribution shifts. InInternational conference on machine learning, pages 5637–5664. PMLR, 2021. 2

    work page 2021

  13. [13]

    Neuromorphic nanowire networks: Principles, progress and future prospects for neuro-inspired information processing.Advances in Physics: X, 6(1):1894234, 2021

    Zdenka Kuncic and Tomonobu Nakayama. Neuromorphic nanowire networks: Principles, progress and future prospects for neuro-inspired information processing.Advances in Physics: X, 6(1):1894234, 2021. 2

    work page 2021

  14. [14]

    Shine, Adrian Diaz-Alvarez, Adam Stieg, James Gimzewski, and Tomonobu Nakayama

    Zdenka Kuncic, Omid Kavehei, Ruomin Zhu, Alon Loef- fler, Kaiwei Fu, Joel Hochstetter, Mike Li, James M. Shine, Adrian Diaz-Alvarez, Adam Stieg, James Gimzewski, and Tomonobu Nakayama. Neuromorphic information process- ing with nanowire networks. In2020 IEEE International Symposium on Circuits and Systems (ISCAS), pages 1–5,

    work page

  15. [15]

    Energy ef- ficiency analysis of spiking neural networks for space appli- cations.arXiv preprint arXiv:2505.11418, 2025

    Paolo Lunghi, Stefano Silvestrini, Dominik Dold, Gabriele Meoni, Alexander Hadjiivanov, and Dario Izzo. Energy ef- ficiency analysis of spiking neural networks for space appli- cations.arXiv preprint arXiv:2505.11418, 2025. 2

    work page arXiv 2025

  16. [16]

    V olcanic hot-spot detection using sentinel-2: A comparison with modis–mirova thermal data series.Remote Sensing, 12 (5):820, 2020

    Francesco Massimetti, Diego Coppola, Marco Laiolo, S´ebastien Valade, Corrado Cigolini, and Maurizio Ripepe. V olcanic hot-spot detection using sentinel-2: A comparison with modis–mirova thermal data series.Remote Sensing, 12 (5):820, 2020. 3 8

    work page 2020

  17. [17]

    In-orbit demonstration of a re-trainable machine learning payload for processing optical imagery.Scientific Reports, 13(1):10391,

    Gonzalo Mateo-Garcia, Josh Veitch-Michaelis, Cormac Pur- cell, Nicolas Longepe, Simon Reid, Alice Anlind, Fredrik Bruhn, James Parr, and Pierre Philippe Mathieu. In-orbit demonstration of a re-trainable machine learning payload for processing optical imagery.Scientific Reports, 13(1):10391,

    work page

  18. [18]

    Gabriele Meoni, Roberto Del Prete, Federico Serva, Alix De Beusscher, Olivier Colin, and Nicolas Long´ep´e. Unlock- ing the use of raw multispectral earth observation imagery for onboard artificial intelligence.IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 17:12521–12537, 2024. 2, 3

    work page 2024

  19. [19]

    E2e: Onboard satellite real-time classification of thermal hotspots events on optical raw data.Astrodynamics, pages 1–17, 2025

    Gabriele Meoni, Roberto Del Prete, Lucia Ancos-Villa, Enrique Albalate-Prieto, David Rijlaarsdam, Jose Luis Espinosa-Aranda, Nicolas Long ´ep´e, Maria Daniela Graziano, and Alfredo Renga. E2e: Onboard satellite real-time classification of thermal hotspots events on optical raw data.Astrodynamics, pages 1–17, 2025. 2

    work page 2025

  20. [20]

    Wildfire induces changes in receiving wa- ters: A review with considerations for water quality manage- ment.Water Resources Research, 58(9):e2021WR030699,

    MJ Paul, SD LeDuc, MG Lassiter, LC Moorhead, PD Noyes, and SG Leibowitz. Wildfire induces changes in receiving wa- ters: A review with considerations for water quality manage- ment.Water Resources Research, 58(9):e2021WR030699,

    work page

  21. [21]

    Ravæn: unsupervised change detection of extreme events using ml on-board satel- lites.Scientific reports, 12(1):16939, 2022

    V ´ıt R˚uˇziˇcka, Anna Vaughan, Daniele De Martini, James Fulton, Valentina Salvatelli, Chris Bridges, Gonzalo Mateo- Garcia, and Valentina Zantedeschi. Ravæn: unsupervised change detection of extreme events using ml on-board satel- lites.Scientific reports, 12(1):16939, 2022. 4

    work page 2022

  22. [22]

    Superglue: Learning feature matching with graph neural networks

    Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4938–4947, 2020. 2

    work page 2020

  23. [23]

    Training-free ai for earth observation change detection using physics aware neuromorphic networks.Scientific Reports, 15(1):35292, 2025

    Stephen Smith, Cormac Purcell, and Zdenka Kuncic. Training-free ai for earth observation change detection using physics aware neuromorphic networks.Scientific Reports, 15(1):35292, 2025. 2, 4

    work page 2025

  24. [24]

    Stieg, Audrius V

    Adam Z. Stieg, Audrius V . Avizienis, Henry O. Sillin, Cristina Martin-Olmos, Masakazu Aono, and James K. Gimzewski. Emergent criticality in complex turing b- type atomic switch networks.Advanced Materials, 24(2): 286–293, 2011. 3

    work page 2011

  25. [25]

    A deep learning approach for early wild- fire detection from hyperspectral satellite images

    Nguyen Thanh Toan, Phan Thanh Cong, Nguyen Quoc Viet Hung, and Jun Jo. A deep learning approach for early wild- fire detection from hyperspectral satellite images. In2019 7th International Conference on Robot Intelligence Technol- ogy and Applications (RiTA), pages 38–45. IEEE, 2019. 1

    work page 2019

  26. [26]

    Towards onboard thermal hotspots segmentation with raw multispectral satel- lite imagery.International Journal of Applied Earth Obser- vation and Geoinformation, 146:105095, 2026

    Cristopher Castro Traba, David Rijlaarsdam, Jian Guo, Roberto Del Prete, and Gabriele Meoni. Towards onboard thermal hotspots segmentation with raw multispectral satel- lite imagery.International Journal of Applied Earth Obser- vation and Geoinformation, 146:105095, 2026. 2

    work page 2026

  27. [27]

    Brain-inspired computing with self-assembled networks of nano-objects.Journal of Physics D: Applied Physics, 57(50):503001, 2024

    Alexander Vahl, Gianluca Milano, Zdenka Kuncic, Simon A Brown, and Paolo Milani. Brain-inspired computing with self-assembled networks of nano-objects.Journal of Physics D: Applied Physics, 57(50):503001, 2024. 2

    work page 2024

  28. [28]

    Current status of landsat program, science, and applications.Remote sensing of environment, 225:127–147, 2019

    Michael A Wulder, Thomas R Loveland, David P Roy, Christopher J Crawford, Jeffrey G Masek, Curtis E Wood- cock, Richard G Allen, Martha C Anderson, Alan S Belward, Warren B Cohen, et al. Current status of landsat program, science, and applications.Remote sensing of environment, 225:127–147, 2019. 1

    work page 2019

  29. [29]

    Information dynamics in neuromorphic nanowire networks.Scientific Reports, 11(1):13047, 2021

    Ruomin Zhu, Joel Hochstetter, Alon Loeffler, Adrian Diaz- Alvarez, Tomonobu Nakayama, Joseph T Lizier, and Zdenka Kuncic. Information dynamics in neuromorphic nanowire networks.Scientific Reports, 11(1):13047, 2021. 2

    work page 2021

  30. [30]

    Online dy- namical learning and sequence memory with neuromorphic nanowire networks.Nature Communications, 14(1):6697,

    Ruomin Zhu, Sam Lilak, Alon Loeffler, Joseph Lizier, Adam Stieg, James Gimzewski, and Zdenka Kuncic. Online dy- namical learning and sequence memory with neuromorphic nanowire networks.Nature Communications, 14(1):6697,

    work page