Pith. sign in

REVIEW 3 major objections 6 minor 54 references

Towards Autonomous Aircraft Surveillance from Nanosatellites through On-Board Inference and Generative Data Augmentation

T0 review · 3 major / 6 minor · reviewed 2026-07-31 · grok-4.5

Pith's one-line read On-board inference and generative balancing of rare aircraft classes turn a nanosatellite from a passive image dump into a real-time airborne surveillance node.

desk verdict Solid SWaP-first engineering that joins known pieces and ships code; offline mAP/F1 gains look real enough, but the on-orbit claim and the unablated pseudo-label loop keep it conditional. read the letter →

arxiv 2607.28470 v1 pith:SHWQYIO3 submitted 2026-07-30 cs.AI cs.CV

classification cs.AIcs.CV
keywords SatelliteAircraftSurveillanceEdgeComputingSmallObjectDetectionDataAugmentationLoRACubeSatOn-boardInferenceClassImbalance
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

Nanosatellites cannot downlink the raw imagery they collect, and public aircraft datasets are too scarce and skewed for standard detectors to learn rare classes such as helicopters. This paper argues that the two problems must be solved together: run a detector on the satellite itself, and grow the minority class with synthetic images before training. A diffusion model fine-tuned with low-rank adaptation produces helicopter scenes that an intermediate detector auto-labels; those samples are mixed with ordinary augmentations and used to train a tiny one-stage detector. The balanced set lifts overall detection quality and especially the helicopter score, while the quantized model fits the edge accelerator’s on-chip memory and is projected to run at tens of frames per second in orbit. If the workflow holds, a 6U CubeSat can filter empty frames on board and downlink only compact detection alerts instead of terabytes of pixels.

What carries the argument

A hardware-first pipeline: FLUX+LoRA synthesis of the helicopter class, automatic pseudo-labelling by an intermediate detector, merge with classical augmentation, then selection and INT8 export of YOLO11n so the entire model resides in the Edge TPU’s 8 MB SRAM.

What would settle it

Compile the INT8 model onto a real Coral Edge TPU (or equivalent flight unit), run SAHI-sliced 1024×1024 inference, and report measured frames per second and power draw; if throughput falls well below the 25–30 FPS projection or the model must spill off-chip, the real-time autonomy claim fails.

Watch

Extended reading notes

Core claim

Combining generative minority-class augmentation (FLUX fine-tuned with LoRA, then pseudo-labelled) with a SWaP-constrained INT8 detector sized for a Coral Edge TPU raises global mAP@50 from 77.9% to 82.2% and helicopter F1 from 0.683 to 0.811 on HRPlanesV2, while the quantized model fits the 8 MB on-die SRAM and is projected to deliver 25–30 FPS on a 6U CubeSat—supporting autonomous on-orbit aircraft surveillance instead of bent-pipe collection.

Load-bearing premise

The claimed real-time on-orbit speed is only a projection scaled from a ground GPU, not a measurement on flight-grade hardware under vacuum or radiation.

Editorial extensions

If this is right

  • A 6U CubeSat can downlink only detection metadata and small patches instead of saturating the radio with raw imagery.
  • Decision latency for aircraft cues drops from hours or days of ground processing to near-real-time on orbit.
  • Minority aircraft classes that classical augmentation alone cannot rescue become trainable via diffusion+LoRA and pseudo-labels.
  • The same nano detector can be extended to multi-type airbase inventories for ISR/GEOINT order-of-battle estimation.
  • Energy spent on the edge accelerator is offset by large savings in radio-frequency transmission within a CubeSat power budget.

Reading between the lines

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

  • If pseudo-labelling quality is the hidden bottleneck, a second round of human spot-checks or active learning on synthetic failures could raise helicopter F1 further without new real captures.
  • The same LoRA-token recipe could be retargeted to other rare orbital classes (ships, vehicles, debris) wherever open datasets are skewed.
  • Exporting the nano model to small UAVs would give a shared detection stack across space and airborne hosts without redesigning the training pipeline.
  • Until Hardware-in-the-Loop and radiation results exist, mission planners should treat the 25–30 FPS figure as a design target rather than a flight guarantee.
Share X Bluesky LinkedIn Reddit HN

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 proposes a joint workflow for autonomous aircraft surveillance from a 6U CubeSat that pairs SWaP-aware on-board inference with generative minority-class augmentation. A Coral Edge TPU constraint (INT8, 8 MB on-die SRAM) drives architecture selection among SSD MobileNet V3, RT-DETR-L and YOLO11n on HRPlanesV2; YOLO11n at 1024 px is retained. Class imbalance (military:helicopter ≈ 4:1) is addressed by classical Albumentations plus a FLUX+LoRA pipeline that synthesises ~2226 helicopter images across four environments, which an intermediate YOLO auto-labels and merges into a balanced set. Reported gains are global mAP@50 77.9%→82.2% and helicopter F1 0.683→0.811; the quantised model (~2.5 MB) is projected at 25–30 FPS with SAHI. The authors contrast this with bent-pipe collection and release code, LoRA adapters and weights.

Significance. If the detection gains and deployability claims hold, the work is a useful systems-level contribution for edge AI in New Space ISR/GEOINT: it jointly treats downlink scarcity and minority-class scarcity under an explicit 8 MB SRAM rule, rather than treating either in isolation. Strengths that should be credited include the hardware-first selection criterion, the open repository (weights, ComfyUI workflows, LoRA adapters), the five-seed stability check (mean mAP@50 0.805), SAHI packaging for full-scene GSD, and an honest limitations section that flags software-only FPS projection and residual domain shift. The empirical ladder unbalanced→classical→generative is clearly motivated and largely reproducible offline.

major comments (3)
  1. [Section 4.2.3, Section 5.3.2, Figure 11] Section 4.2.3 and §5.3.2: the headline helicopter F1 lift (0.683→0.811) and the final balanced training set rest on pseudo-labels produced by an intermediate YOLO trained only on the classically augmented partition. No human audit, confidence filter, label-precision/recall estimate, or ablation that trains on synthetic images with withheld/noisy labels is reported. Figure 11 and the accompanying text already note residual domain shift and elevated background FN/FP; without a quality check on the ~2226 auto-labelled boxes, a non-trivial fraction of the minority-class gain may be an artefact of the labelling loop rather than improved feature learning on real held-out imagery. A small audited subset or a controlled noise ablation is needed to underwrite the central empirical claim.
  2. [Abstract, Section 5.3, Figure 9] Abstract, §5.3 and Figure 9 disagree on the final helicopter F1 (0.811 in abstract/conclusions vs. 0.79 at conf=0.392 on the F1-confidence curve). Global mAP@50 82.2% for the generative regime is stated in the abstract and conclusions but is not tabulated beside the unbalanced (77.90%) and classical (81.62%) figures with the same metric suite (per-class P/R/F1, mAP@50-95). Please reconcile the F1 numbers and add a single comparison table so the incremental contribution of FLUX+LoRA over classical augmentation is unambiguous.
  3. [Section 5.4, Abstract, Section 6.2] Section 5.4: the 25–30 FPS on-orbit figure is obtained only by scaling measured T4 latency with an Edge TPU throughput-to-power ratio; no compiled-model latency, power, or thermal measurement on a Coral device (even benchtop) is given. The authors correctly flag this in §6.2, yet the abstract and contribution list still present the projection as supporting “real-time, autonomous” decision support. Either demote the FPS claim to an explicitly labelled estimate everywhere it appears, or add at least a host-side Edge TPU bench measurement of the exported INT8+SAHI pipeline so the real-time claim is not carried solely by the scaling axiom.
minor comments (6)
  1. [Section 5.2, Figures 5–7] Table 5 and Figures 5–7: figure captions and axis labels are partially swapped relative to content (e.g., Fig. 5 is described as mAP vs. speed but plots parameters/disk weight; Fig. 6 is described as parametric footprint but plots mAP and FPS). Align captions with the plotted quantities.
  2. [Section 5.1, Table 4] Section 5.1 states training at 640×640 in Table 4 while the retained operating point is 1024 px; clarify which resolution was used for the five-seed study and for each reported mAP.
  3. [References] Several Spanish access-date strings remain in the references (“Último acceso: …”). Normalise to the journal’s English bibliography style.
  4. [Section 2.1] Section 2.1: “European Space Agency European Space Agency (ESA)” is duplicated; fix the acronym expansion.
  5. [Introduction, Data availability] Contribution bullet on open release and the Data availability statement use slightly different GitHub URLs (with/without .git). Make them identical.
  6. [Figure 10] Figure 10 caption cites an example “14 vs 13” that does not match the panel headers (e.g., 11|12, 8|9). Align caption with the figure.

Circularity Check

1 steps flagged · score 1.0 of 10

No load-bearing circular derivation; empirical mAP/F1 gains are measured on held-out real data, and the mild pseudo-label loop is standard semi-supervised practice rather than a tautology.

  1. fitted input called prediction [Section 4.2.3 (Pseudo-labelling and class balancing); cf. §5.3.2]
    "An intermediate YOLOv11 detector, previously trained on the classically augmented set, is used to auto-label the synthetic output. The resulting labels are merged with the classically augmented samples, producing a final balanced dataset of more than 53000 instances at approximately 18000 instances per class."

    Synthetic helicopter boxes used to train the final detector are produced by a model trained only on classical data, so training targets are model-dependent rather than independent human GT. This is mild and non-load-bearing: final mAP/F1 are still reported on unaltered real val/test partitions, so the reported lift is not equal to the pseudo-label fit by construction—only the training signal is partly circular. Standard semi-supervised practice; does not force the headline metrics.

full rationale

The paper’s central claims are empirical engineering results, not first-principles predictions. Architecture selection is constrained by an external hardware budget (Coral Edge TPU 8 MB SRAM after INT8), then YOLO11n is trained and scored on HRPlanesV2 partitions that the authors state are left untouched by augmentation. Classical vs. generative regimes are compared by the same metrics on that held-out real test distribution; nothing forces mAP@50 77.9%→82.2% or helicopter F1 0.683→0.811 by construction. The generative stage uses an intermediate detector only to auto-annotate synthetic images before they enter training—if those pseudo-labels were systematically wrong, real-test metrics would be expected to fall, not rise tautologically—so this is a validity/noise concern, not self-definitional circularity. The 25–30 FPS on-orbit figure is an explicit scaling projection from T4 latency (acknowledged as software-only in §6.2), not a fitted parameter renamed as a prediction of the same quantity. There is no uniqueness theorem, self-citation chain, or ansatz smuggled in as external fact. Score 1 only for the weak, non-load-bearing dependence of training labels on a model trained on the same classical pipeline; the evaluation chain itself remains independent.

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

The central claim rests on standard ML and CubeSat engineering assumptions plus several hand-chosen training and deployment knobs. No new physical entity is postulated. Load-bearing domain assumptions include the 8 MB on-die SRAM hard limit, sufficiency of ~1.5 m GSD for aircraft morphology, and validity of T4-to-Edge-TPU latency scaling. Free parameters are ordinary ML choices (seed, epochs, LoRA strength, confidence operating point, synthetic count) that affect reported scores but are documented.

free parameters (5)
  • Training seed and epoch budget = seed 93; 20 epochs
    Seed 93 chosen as representative after a five-seed sweep; final models trained 20 epochs only—affects reported mAP/F1.
  • LoRA trigger/strength and sampler settings = strength 0.85 / 27 steps / denoise 0.95
    Custom token and strengths (model 0.85, clip 1.0), 27 steps, CFG=1, denoise 0.95 control synthetic distribution quality.
  • Operating confidence threshold = 0.394 (final F1 peak)
    F1-maximizing threshold shifts across regimes (e.g. 0.394 on final model) and defines tactical precision/recall trade-off.
  • Synthetic helicopter count and four environment prompts = ~2226 images; 4 environments
    About 2226 synthetic images across industrial, rural, coastal, arid scenes set the minority-class balance target (~18k/class).
  • SAHI patch size and overlap = 1024 px, 15% overlap
    1024×1024 slices with 15% overlap determine full-scene throughput and duplicate handling on orbit.
assumptions (5)
  • domain assumption After INT8 quantisation the detector must fit entirely in the Coral Edge TPU’s 8 MB on-die SRAM or FPS collapses due to CPU offload.
    Stated as the central design rule in Sections 3.3 and 4.3; drives discarding RT-DETR-L.
  • domain assumption COTS optics at ~500 km yield ~1.5 m GSD so standard aircraft retain usable morphology at 1024-pixel native training resolution.
    Section 3.4; justifies detector input size and small-object framing.
  • ad hoc to paper T4-measured latency scaled by Edge TPU throughput-to-power ratio predicts on-orbit FPS.
    Section 5.4 projection underpinning 25–30 FPS; not an industry-standard equivalence without device measurement.
  • domain assumption Pseudo-labels from an intermediate detector on synthetic images are accurate enough to improve real-test minority-class metrics when merged with classical augmentation.
    Section 4.2.3 pipeline; standard but unverified label precision/recall on the synthetic set.
  • standard math One-stage detector training and mAP/F1 evaluation methodology (Ultralytics/COCO-style) is a valid quality measure for tactical aircraft alerts.
    Sections 4.3 and 5; conventional CV metrics.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Towards Autonomous Aircraft Surveillance from Nanosatellites through On-Board Inference and Generative Data Augmentation." pith.science (2026). https://pith.science/paper/SHWQYIO3

@misc{pith2026260728470,
  author       = {Pith},
  title        = {Pith review of: Towards Autonomous Aircraft Surveillance from Nanosatellites through On-Board Inference and Generative Data Augmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SHWQYIO3}},
  note         = {Machine review of arXiv:2607.28470}
}
read the original abstract

Airborne surveillance from low Earth orbit is hindered by two interconnected bottlenecks: nanosatellites have a limited downlink budget, yet the conventional approach still transmits terabytes of raw imagery to the ground for processing, and open satellite datasets for aircraft are scarce and severely class-imbalanced. These limitations either delay timely decision-making or prevent standard detectors from learning robust representations of rare aircraft classes. In this paper, a workflow that combines on-board inference with generative data augmentation is proposed to address both limitations jointly. Inference is executed on a 6U CubeSat equipped with a low-power edge tensor accelerator, while a diffusion model fine-tuned through low-rank adaptation generates synthetic minority-class imagery. This synthetic output is automatically annotated, pseudo-labelled, by an intermediate detector and merged with classically augmented samples. The results show that the balanced dataset increases global mean average precision from 77.9% to 82.2%, with the minority class rising from F1=0.683 to F1=0.811, and that the quantised detector fits the on-chip memory and projects 25-30 frames per second on orbit. This approach contrasts with the conventional bent-pipe architecture, in which the satellite acts as a passive data collector. Therefore, the computational tests support the proposed workflow as a decision-support tool for real-time, autonomous airborne surveillance from nanosatellites.

Figures

Figures reproduced from arXiv: 2607.28470 by the authors.

Figure 1
Figure 1. Block diagram of the space-to-ground data-downlink architecture for the 6U [PITH_FULL_IMAGE:figures/full_fig_p014_1.png] view at source ↗
Figure 2
Figure 2. Distribution of instances (bounding boxes) per category in the original training [PITH_FULL_IMAGE:figures/full_fig_p016_2.png] view at source ↗
Figure 3
Figure 3. Distribution of the relative area of the bounding boxes with respect to the total [PITH_FULL_IMAGE:figures/full_fig_p016_3.png] view at source ↗
Figures from the paper (11 more)
Figure 4
Figure 4. Figure 4: Block diagram of the generative data-augmentation pipeline. The limited real [PITH_FULL_IMAGE:figures/full_fig_p019_4.png]
Figure 5
Figure 5. Figure 5: Direct comparison between mean average precision (mAP@50) and baseline [PITH_FULL_IMAGE:figures/full_fig_p023_5.png]
Figure 6
Figure 6. Figure 6: Parametric footprint of each architecture against the 8 MB SRAM budget of the [PITH_FULL_IMAGE:figures/full_fig_p024_6.png]
Figure 7
Figure 7. Figure 7: Operational design space. The bubble size is proportional to the model weight [PITH_FULL_IMAGE:figures/full_fig_p024_7.png]
Figure 8
Figure 8. Figure 8: Evolution of mAP@50 across the three training regimes: unbalanced baseline, [PITH_FULL_IMAGE:figures/full_fig_p025_8.png]
Figure 9
Figure 9. Figure 9: F1 curve of the final YOLO11n model trained on the FLUX+LoRA-balanced [PITH_FULL_IMAGE:figures/full_fig_p026_9.png]
Figure 10
Figure 10. Figure 10: Qualitative evaluation of apparent false positives. The inference results show [PITH_FULL_IMAGE:figures/full_fig_p029_10.png]
Figure 11
Figure 11. Figure 11: Combined confusion matrix of the final balanced model, displaying both nor [PITH_FULL_IMAGE:figures/full_fig_p030_11.png]
Figure 12
Figure 12. Figure 12: Normalised confusion matrix of the extended 22-class YOLO11n model. The [PITH_FULL_IMAGE:figures/full_fig_p031_12.png]
Figure 13
Figure 13. Figure 13: Inference mosaic on an airbase scene, illustrating the operational value of the [PITH_FULL_IMAGE:figures/full_fig_p032_13.png]
Figure 14
Figure 14. Figure 14: Synthetic helicopter imagery generated by the FLUX+LoRA pipeline across [PITH_FULL_IMAGE:figures/full_fig_p033_14.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

54 extracted references · 1 canonical work pages

  1. [1]

    M. I. Skolnik, et al., Introduction to radar systems, volume 3, McGraw- hill New York, 1980

  2. [2]

    Strohmeier, M

    M. Strohmeier, M. Schäfer, V. Lenders, I. Martinovic, Realities and challenges of nextgen air traffic management: the case of ads-b, IEEE communications magazine 52 (2014) 111–118. 37

  3. [3]

    M. N. Sweeting, Modern small satellites-changing the economics of space, Proceedings of the IEEE 106 (2018) 343–361

  4. [4]

    Poghosyan, A

    A. Poghosyan, A. Golkar, Cubesat evolution: Analyzing cubesat capa- bilities for conducting science missions, Progress in Aerospace Sciences 88 (2017) 59–83

  5. [5]

    Furano, G

    G. Furano, G. Meoni, A. Dunne, D. Moloney, V. Ferlet-Cavrois, A. Tavoularis, J. Byrne, L. Buckley, M. Psarakis, K.-O. Voss, et al., Towards the use of artificial intelligence on the edge in space systems: Challenges and opportunities, IEEE Aerospace and Electronic Systems Magazine 35 (2020) 44–56

  6. [6]

    Tchikladze, HRPlanesV2 computer vision model, https://universe.roboflow.com/alban-tchikladze-kxzqe/ hrplanesv2-7alnq, 2023

    A. Tchikladze, HRPlanesV2 computer vision model, https://universe.roboflow.com/alban-tchikladze-kxzqe/ hrplanesv2-7alnq, 2023. Último acceso: 26 de mayo de 2026

  7. [7]

    B. F. Labs, Flux.1: Text-to-image generation models, 2024. URL: https://blackforestlabs.ai/

  8. [8]

    E. J. Hu, Y. Shen, P. Wallis, Z. Allen-Zhu, Y. Li, S. Wang, L. Wang, W. Chen, Lora: Low-rank adaptation of large language models, 2021. URL:https://arxiv.org/abs/2106.09685.arXiv:2106.09685

Show all 54 references
  1. [9]

    Último acceso: 26 de mayo de 2026

    European Space Agency, Deimos Space, Ciseres: Ai-powered satellites for rapid disaster response,https://www.esa.int/Applications/ Connectivity_and_Secure_Communications/Ciseres_AI-powered_ satellites_for_rapid_disaster_response, 2024. Último acceso: 26 de mayo de 2026

  2. [10]

    Cass, Taking ai to the edge: Google’s tpu now comes in a maker- friendly package, IEEE Spectrum 56 (2019) 16–17

    S. Cass, Taking ai to the edge: Google’s tpu now comes in a maker- friendly package, IEEE Spectrum 56 (2019) 16–17

  3. [11]

    Prokscha, M

    R. Prokscha, M. Schneider, A. Höß, Efficient edge deployment demon- strated on yolov5 and coral edge tpu, in: Industrial Artificial Intelligence Technologies and Applications, River Publishers, 2023, pp. 141–155

  4. [12]

    M. A. Chanoui, I. Khalil, M. Sbihi, Z. E. A. Alaoui Ismaili, Z. Guen- noun, Trends and applications of on-board image processing for earth observation nanosatellites: a systematic review, International Journal of Aeronautical and Space Sciences 26 (2025) 1989–2019. 38

  5. [13]

    Mystkowska, M

    G. Mystkowska, M. Monopoli, P. Nannipieri, L. Zulberti, D. M. Co- dinachs, L. Fanucci, Hardware platforms enabling edge ai for space applications: A critical review, IEEE Access (2025)

  6. [14]

    G.Lentaris, V.Leon, C.Sakos, D.Soudris, A.Tavoularis, A.Costantino, C. B. Polo, Performance and radiation testing of the coral tpu co- processor for ai onboard satellites, in: 2023 European Data Handling & Data Processing Conference (EDHPC), IEEE, 2023, pp. 1–4

  7. [15]

    S. C. Magalhães, F. N. dos Santos, P. Machado, A. P. Moreira, J. Dias, Benchmarking edge computing devices for grape bunches and trunks detection using accelerated object detection single shot multibox deep learning models, Engineering Applications of Artificial Intelligence 1...

  8. [16]

    I. J. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S.Ozair, A.Courville, Y.Bengio, Generativeadversarialnetworks, 2014. URL:https://arxiv.org/abs/1406.2661.arXiv:1406.2661

  9. [17]

    Shorten, T

    C. Shorten, T. M. Khoshgoftaar, A survey on image data augmentation for deep learning, Journal of big data 6 (2019) 1–48

  10. [18]

    Beery, G

    S. Beery, G. Van Horn, P. Perona, Recognition in terra incognita, in: Proceedings of the European conference on computer vision (ECCV), 2018, pp. 456–473

  11. [19]

    Rombach, A

    R. Rombach, A. Blattmann, D. Lorenz, P. Esser, B. Ommer, High- resolution image synthesis with latent diffusion models, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recogni- tion, 2022, pp. 10684–10695

  12. [20]

    P. N. Mueller, Attention-enhanced conditional-diffusion-based data synthesis for data augmentation in machine fault diagnosis, En- gineering Applications of Artificial Intelligence 131 (2024) 107696. URL:https://doi.org/10.1016/j.engappai.2023.107696. doi:10. 1016/j.engappai.2...

  13. [21]

    Dalal, B

    N. Dalal, B. Triggs, Histograms of oriented gradients for human detec- tion, in: 2005 IEEE computer society conference on computer vision and pattern recognition (CVPR’05), volume 1, Ieee, 2005, pp. 886–893. 39

  14. [22]

    Cortes, V

    C. Cortes, V. Vapnik, Support-vector networks, Machine learning 20 (1995) 273–297

  15. [23]

    S. Ren, K. He, R. Girshick, J. Sun, Faster r-cnn: Towards real-time object detection with region proposal networks, 2016. URL:https:// arxiv.org/abs/1506.01497.arXiv:1506.01497

  16. [24]

    G.-S. Xia, X. Bai, J. Ding, Z. Zhu, S. Belongie, J. Luo, M. Datcu, M. Pelillo, L. Zhang, Dota: A large-scale dataset for object detection in aerial images, in: Proceedings of the IEEE conference on computer vision and pattern recognition, 2018, pp. 3974–3983

  17. [25]

    Redmon, S

    J. Redmon, S. Divvala, R. Girshick, A. Farhadi, You only look once: Unified, real-time object detection, in: Proceedings of the IEEE confer- ence on computer vision and pattern recognition, 2016, pp. 779–788

  18. [26]

    T.-Y. Lin, P. Goyal, R. Girshick, K. He, P. Dollár, Focal loss for dense object detection, in: Proceedings of the IEEE international conference on computer vision, 2017, pp. 2980–2988

  19. [27]

    Q. Liu, X. Xiang, Y. Wang, Z. Luo, F. Fang, Aircraft detection in remote sensing image based on corner clustering and deep learning, Engineering Applications of Artificial Intelligence 87 (2020) 103333. URL:https://doi.org/10.1016/j.engappai.2019.103333. doi:10. 1016/j.engappa...

  20. [28]

    İlmak, T

    D. İlmak, T. Bakirman, E. Sertel, Exploring you only look once v8 and v9 for efficient airplane detection in very high resolution remote sensing imagery, Engineering Applications of Artificial Intelligence 160 (2025) 111854. URL:https://doi.org/10.1016/j.engappai.2025.111854. ...

  21. [29]

    S. Liu, F. Shao, J. Xue, Recent advances of yolo-based object de- tection in remote sensing imagery: A comprehensive review, En- gineering Applications of Artificial Intelligence 179 (2026) 115171. URL:https://doi.org/10.1016/j.engappai.2026.115171. doi:10. 1016/j.engappai.2026.115171

  22. [30]

    Zhang, T

    X. Zhang, T. Lu, J. Wang, S. Fu, F. Gao, Small object detection by edge-aware neural network, Engineering Applications of Artificial 40 Intelligence 138 (2024) 109406. URL:https://doi.org/10.1016/j. engappai.2024.109406. doi:10.1016/j.engappai.2024.109406

  23. [31]

    J. R. Wertz, D. F. Everett, J. J. Puschell, Space mission engineering: the new smad, (No Title) (2011)

  24. [32]

    B. Yost, S. Weston, State-of-the-art small spacecraft technology, Tech- nical Report, 2024

  25. [33]

    Puig-Suari, C

    J. Puig-Suari, C. Turner, W. Ahlgren, Development of the standard cubesat deployer and a cubesat class picosatellite, in: 2001 IEEE aerospaceconferenceproceedings(Cat.No.01TH8542), volume1, IEEE, 2001, pp. 1–347

  26. [34]

    V. P. Katuntsev, Approaches to decompression safety support of eva for orbital and interplanetary missions, Acta Astronautica 66 (2010) 96–101

  27. [35]

    T. Ma, M. Zhang, K. Chen, Z. Song, Y. Wu, X. Qian, Asymnvm: An efficient framework for implementing persistent data structures on asym- metric nvm architecture, in: Proceedings of the Twenty-Fifth Interna- tional Conference on Architectural Support for Programming Languages an...

  28. [36]

    S. S. Gao, Q. Luo, F. Zhu, Circularly polarized antennas, John Wiley & Sons, 2014

  29. [37]

    Costantine, Y

    J. Costantine, Y. Tawk, I. Maqueda, M. Sakovsky, G. Olson, S. Pelle- grino, C.G.Christodoulou, Uhfdeployablehelicalantennasforcubesats, IEEE Transactions on Antennas and Propagation 64 (2016) 3752–3759

  30. [38]

    URL:https://www.nvidia

    NVIDIA Corporation, NVIDIA T4 Tensor Core GPU Architec- ture, Whitepaper, NVIDIA, 2018. URL:https://www.nvidia. com/content/dam/en-zz/Solutions/Data-Center/tesla-t4/ t4-tensor-core-datasheet-951643.pdf

  31. [39]

    Rapuano, G

    E. Rapuano, G. Meoni, T. Pacini, G. Dinelli, G. Furano, G. Giuffrida, L. Fanucci, An fpga-based hardware accelerator for cnns inference on board satellites: Benchmarking with myriad 2-based solution for the cloudscout case study, Remote Sensing 13 (2021) 1518. 41

  32. [40]

    ai/docs/edgetpu/models-intro, 2020

    Google LLC, Tensorflow models on the edge tpu,https://www.coral. ai/docs/edgetpu/models-intro, 2020. Último acceso: 07 de marzo de 2026

  33. [41]

    URL: https://docs.ultralytics.com/modes/export/, Último acceso: 07 de marzo de 2026

    Ultralytics, Ultralytics yolo export formats: Edge tpu, 2024. URL: https://docs.ultralytics.com/modes/export/, Último acceso: 07 de marzo de 2026

  34. [42]

    tensorflow.org/lite/performance/post_training_integer_ quant, 2023

    TensorFlow, Post-training integer quantization,https://www. tensorflow.org/lite/performance/post_training_integer_ quant, 2023. TensorFlow Lite Documentation. Último acceso: 07 de marzo de 2026

  35. [43]

    URL:https://simera-sense

    Simera Sense, Multiscape200, 2025. URL:https://simera-sense. com/products/multiscape200/, Último acceso: 18 de marzo de 2026

  36. [44]

    URL:https: //dragonflyaerospace.com/products/mantis/, Último acceso: 18 de marzo de 2026

    Dragonfly Aerospace, Mantis imager, 2024. URL:https: //dragonflyaerospace.com/products/mantis/, Último acceso: 18 de marzo de 2026

  37. [45]

    URL:https://tensortech.co/product/detail/ adcs_for_over_30_kg, Último acceso: 18 de marzo de 2026

    Tensor Tech, Attitude determination and control system for small satel- lites (>30 kg), 2025. URL:https://tensortech.co/product/detail/ adcs_for_over_30_kg, Último acceso: 18 de marzo de 2026

  38. [46]

    Krenz, P

    R. Krenz, P. Sroka, M. Sybis, I. Zainutdinov, K. Wesołowski, A low-cost high-throughput c-band communication system for cubesats, Electron- ics 12 (2023). URL:https://www.mdpi.com/2079-9292/12/20/4374. doi:10.3390/electronics12204374

  39. [47]

    URL:https://www.ericsson.com/en/ reports-and-papers/white-papers/digital-airspace, white pa- per, PDF

    Ericsson, Digital Airspace: The Sky - The Digital Frontier, Tech- nical Report, Ericsson, 2020. URL:https://www.ericsson.com/en/ reports-and-papers/white-papers/digital-airspace, white pa- per, PDF. Último acceso: 18 de marzo de 2026

  40. [48]

    Kuwahara, Introduction to cubesat command and data handling sys- tem, 2021

    T. Kuwahara, Introduction to cubesat command and data handling sys- tem, 2021. URL:https://www.unoosa.org/oosa/en/ourwork/psa/ hsti/kibocube.html, kiboCUBE Academy Lecture 10, PDF. Último acceso: 18 de marzo de 2026. 42

  41. [49]

    Buslaev, V

    A. Buslaev, V. I. Iglovikov, E. Khvedchenya, A. Parinov, M. Druzhinin, A. A. Kalinin, Albumentations: Fast and flexible image augmentation, Information 11 (2020) 125. doi:10.3390/info11020125

  42. [50]

    W. Liu, D. Anguelov, D. Erhan, C. Szegedy, S. Reed, C.-Y. Fu, A. C. Berg, Ssd: Single shot multibox detector, in: European conference on computer vision, Springer, 2016, pp. 21–37

  43. [51]

    G.Jocher, J.Qiu, Ultralyticsyolo11, 2024.URL:https://github.com/ ultralytics/ultralytics

  44. [52]

    Y. Zhao, W. Lv, S. Xu, J. Wei, G. Wang, Q. Dang, Y. Liu, J. Chen, Detrs beat yolos on real-time object detection, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2024, pp. 16965–16974

  45. [53]

    N.Carion, F.Massa, G.Synnaeve, N.Usunier, A.Kirillov, S.Zagoruyko, End-to-end object detection with transformers, in: European conference on computer vision, Springer, 2020, pp. 213–229

  46. [54]

    F. C. Akyon, S. O. Altinuc, A. Temizel, Slicing aided hyper inference and fine-tuning for small object detection, in: 2022 IEEE international conference on image processing (ICIP), IEEE, 2022, pp. 966–970. 43

Pith tools

Reviewed July 31, 2026 · model on record in the stance chip above.