pith. sign in

arxiv: 2511.10834 · v3 · submitted 2025-11-13 · 💻 cs.LG · cs.DC

EarthSight: A Distributed Framework for Low-Latency Satellite Intelligence

Ansel Kaplan Erol , Seungjun Lee , Divya Mahajan This is my paper

Pith reviewed 2026-05-17 21:47 UTC · model grok-4.3

classification 💻 cs.LG cs.DC
keywords satellite imageryonboard machine learningdistributed computinglow-latency systemsmulti-task inferencepriority schedulingdynamic filtering
0
0 comments X

The pith

By distributing satellite image analysis between orbit and ground with shared model backbones and priority scheduling, EarthSight reduces compute time by 1.9 times and latency from 51 to 21 minutes.

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

EarthSight addresses delays in satellite imagery delivery for critical applications by shifting from isolated satellite processing to a coordinated system between orbit and ground. It introduces multi-task inference using shared backbones to handle multiple vision tasks efficiently on satellites, a ground-station scheduler that aggregates requests and assigns compute resources, and dynamic filter ordering to discard low-value images early. These elements leverage global ground context and onboard adaptability to respect strict power and bandwidth limits. Simulator evaluations show concrete improvements in speed and responsiveness compared to existing approaches.

Core claim

EarthSight redefines satellite image intelligence as a distributed decision problem. Its three innovations—multi-task inference with shared backbones, ground-station query scheduler with priority prediction, and dynamic filter ordering based on selectivity, accuracy, and cost—enable scalable low-latency analysis. Simulator tests confirm a 1.9x reduction in average compute time per image and a drop in 90th percentile latency from 51 to 21 minutes.

What carries the argument

Distributed runtime framework that integrates onboard multi-task inference using shared backbones with ground-station query scheduling and dynamic filter ordering to manage resources across the constellation.

If this is right

  • Average compute time per image drops by a factor of 1.9.
  • 90th percentile end-to-end latency falls from 51 minutes to 21 minutes.
  • Resource use decreases, allowing more images to be analyzed within mission constraints.
  • High-priority images reach users faster for applications like disaster response.

Where Pith is reading between the lines

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

  • The framework could inform designs for other edge computing in space or remote sensor networks.
  • Combining it with advanced compression techniques might yield even lower latencies in practice.
  • Testing on varied satellite constellations would validate scalability under different orbital conditions.
  • It suggests rethinking traditional centralized ground processing for orbital data streams.

Load-bearing premise

The satellite simulator accurately represents real power, compute, and communication limits, and shared backbones maintain accuracy without degradation across tasks.

What would settle it

Running EarthSight on actual satellites and measuring if the compute time and latency reductions match the simulator results under real conditions.

Figures

Figures reproduced from arXiv: 2511.10834 by Ansel Kaplan Erol, Divya Mahajan, Seungjun Lee.

Figure 1
Figure 1. Figure 1: LEO Earth observation systems perform on-board [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the EARTHSIGHT system. EARTH￾SIGHT integrates onboard multi-task models with predictive scheduling to enable query-driven, low-latency analysis of satellite images. by sharing feature extraction across tasks, reducing memory overhead, and supporting multiple mission goals without de￾ploying multiple large models [27]. These models use a shared backbone to extract latent features, followed by li… view at source ↗
Figure 3
Figure 3. Figure 3: EARTHSIGHT’s multi-task model architecture. A shared backbone processes the input image to produce a latent representation, which is used by lightweight, task-specific heads to perform heterogeneous tasks such as classification. the transmission order to maximize responsiveness. Together, these components form a flexible, resource-aware system that adapts to changing mission conditions while maximizing the… view at source ↗
Figure 4
Figure 4. Figure 4: Timing diagrams illustrating filter execution strategies on a CPU-xPU system: (a) Serval’s (baseline) sequential [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Comparison of model size and performance across [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: 90th percentile tail latency, measured from first [PITH_FULL_IMAGE:figures/full_fig_p009_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Distribution of avoidable image latencies for high priority ( [PITH_FULL_IMAGE:figures/full_fig_p010_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Ablation study results show the impact of individual [PITH_FULL_IMAGE:figures/full_fig_p010_8.png] view at source ↗
Figure 10
Figure 10. Figure 10: High-priority intelligence images. (a) shows a [PITH_FULL_IMAGE:figures/full_fig_p015_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: A high-priority image showing an anomalous vehi [PITH_FULL_IMAGE:figures/full_fig_p016_11.png] view at source ↗
read the original abstract

Low-latency delivery of satellite imagery is essential for time-critical applications such as disaster response, intelligence, and infrastructure monitoring. However, traditional pipelines rely on downlinking all captured images before analysis, introducing delays of hours to days due to restricted communication bandwidth. To address these bottlenecks, emerging systems perform onboard machine learning to prioritize which images to transmit. However, these solutions typically treat each satellite as an isolated compute node, limiting scalability and efficiency. Redundant inference across satellites and tasks further strains onboard power and compute costs, constraining mission scope and responsiveness. We present EarthSight, a distributed runtime framework that redefines satellite image intelligence as a distributed decision problem between orbit and ground. EarthSight introduces three core innovations: (1) multi-task inference on satellites using shared backbones to amortize computation across multiple vision tasks; (2) a ground-station query scheduler that aggregates user requests, predicts priorities, and assigns compute budgets to incoming imagery; and (3) dynamic filter ordering, which integrates model selectivity, accuracy, and execution cost to reject low-value images early and conserve resources. EarthSight leverages global context from ground stations and resource-aware adaptive decisions in orbit to enable constellations to perform scalable, low-latency image analysis within strict downlink bandwidth and onboard power budgets. Evaluations using a prior established satellite simulator show that EarthSight reduces average compute time per image by 1.9x and lowers 90th percentile end-to-end latency from first contact to delivery from 51 to 21 minutes compared to the state-of-the-art baseline.

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 / 2 minor

Summary. The manuscript introduces EarthSight, a distributed runtime framework for low-latency satellite image intelligence. It redefines the problem as a joint orbit-ground decision process and contributes three techniques: multi-task inference via shared backbones to amortize computation across vision tasks, a ground-station query scheduler that aggregates requests and assigns compute budgets, and dynamic filter ordering that incorporates model selectivity, accuracy, and cost to drop low-value images early. Evaluations on a prior established satellite simulator report a 1.9× reduction in average compute time per image and a drop in 90th-percentile end-to-end latency from 51 to 21 minutes relative to a stated state-of-the-art baseline.

Significance. If the simulator faithfully reproduces real power, bandwidth, and orbital constraints, the framework could meaningfully advance time-critical satellite applications such as disaster response by reducing downlink volume and enabling scalable constellation-level inference. The integration of shared-backbone multi-task learning with ground-orchestrated scheduling and early rejection is a coherent systems contribution. The use of an established simulator supports reproducibility of the reported numbers, though the overall significance remains conditional on validation of the modeling assumptions.

major comments (2)
  1. [Evaluation] Evaluation section: the headline claims of 1.9× compute-time reduction and 90th-percentile latency improvement from 51 min to 21 min rest exclusively on runs inside the cited satellite simulator. No sensitivity analysis is provided for variations in power draw, downlink bandwidth variability, compute heterogeneity, or orbital contact schedules, nor is any hardware cross-validation reported. Because these metrics constitute the primary empirical support for the framework’s value, the absence of such checks is load-bearing for the central performance claims.
  2. [Evaluation] Evaluation section: the manuscript reports no ablation studies isolating the individual contributions of shared-backbone multi-task inference, the ground-station scheduler, and dynamic filter ordering. It also omits any quantification of accuracy trade-offs or task-specific degradation that may arise from the shared backbone. These omissions prevent assessment of whether the observed speedups are robust or come at an unacceptable cost to the vision-task quality that the system is intended to deliver.
minor comments (2)
  1. [Abstract and Evaluation] The abstract and evaluation results should include error bars, confidence intervals, or statistical tests for the reported speedups and latency figures.
  2. [Related Work / Evaluation] Clarify the precise definition and implementation details of the state-of-the-art baseline used for comparison, including any differences in assumptions about onboard resources or task sets.

Simulated Author's Rebuttal

2 responses · 1 unresolved

We thank the referee for the constructive feedback. We address the major comments below and describe the revisions we will make to strengthen the empirical support for our claims.

read point-by-point responses

  1. Referee: [Evaluation] Evaluation section: the headline claims of 1.9× compute-time reduction and 90th-percentile latency improvement from 51 min to 21 min rest exclusively on runs inside the cited satellite simulator. No sensitivity analysis is provided for variations in power draw, downlink bandwidth variability, compute heterogeneity, or orbital contact schedules, nor is any hardware cross-validation reported. Because these metrics constitute the primary empirical support for the framework’s value, the absence of such checks is load-bearing for the central performance claims.

    Authors: We agree that additional sensitivity analysis would improve confidence in the results. In the revised manuscript we will add experiments that vary power draw, downlink bandwidth, compute heterogeneity, and orbital contact schedules within the established simulator and report the resulting ranges for compute time and latency. Hardware cross-validation lies outside the scope of this simulation study; we will explicitly discuss this limitation and the fidelity of the simulator's power, bandwidth, and orbital models. revision: partial

  2. Referee: [Evaluation] Evaluation section: the manuscript reports no ablation studies isolating the individual contributions of shared-backbone multi-task inference, the ground-station scheduler, and dynamic filter ordering. It also omits any quantification of accuracy trade-offs or task-specific degradation that may arise from the shared backbone. These omissions prevent assessment of whether the observed speedups are robust or come at an unacceptable cost to the vision-task quality that the system is intended to deliver.

    Authors: We will add a dedicated ablation subsection that disables each component in turn (shared backbone, ground scheduler, dynamic filter ordering) and quantifies the incremental effect on average compute time and 90th-percentile latency. We will also report task-specific accuracy metrics (precision, recall, F1) for each vision task when using the shared backbone versus independent models, thereby quantifying any accuracy trade-offs. revision: yes

standing simulated objections not resolved
  • Hardware cross-validation on physical satellite platforms

Circularity Check

0 steps flagged

No circularity: empirical simulator results are independent of any derivation chain

full rationale

The paper describes a distributed runtime framework with three innovations (shared-backbone multi-task inference, ground-station query scheduler, dynamic filter ordering) and reports performance numbers obtained by running the system inside a prior established satellite simulator. These numbers are direct empirical measurements of compute time and end-to-end latency; they are not obtained by fitting parameters to a subset of data and then re-using the fit as a prediction, nor by any equation that reduces to its own inputs by construction. No self-citation is invoked as a uniqueness theorem or load-bearing premise for the central claims. The evaluation therefore stands as an external benchmark rather than a self-referential derivation.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

The framework description relies on standard assumptions about satellite compute budgets and communication links but introduces no explicit free parameters, axioms, or invented entities beyond the named scheduling and filtering mechanisms.

pith-pipeline@v0.9.0 · 5579 in / 1166 out tokens · 84905 ms · 2026-05-17T21:47:11.443949+00:00 · methodology

discussion (0)

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

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Equinox: Decentralized Scheduling for Hardware-Aware Orbital Intelligence

    cs.DC 2026-04 unverdicted novelty 5.0

    Equinox uses a barrier-function-derived marginal cost to enable value-based adaptive scheduling and neighbor offloading in energy-constrained satellite constellations, yielding 20-31% throughput gains and higher batte...

Reference graph

Works this paper leans on

53 extracted references · 53 canonical work pages · cited by 1 Pith paper · 2 internal anchors

  1. [1]

    McDowell

    Jonathan C. McDowell. The Low Earth Orbit Satellite Population and Impacts of the SpaceX Starlink Constel- 11 lation.The Astrophysical Journal Letters, 892(2):L36,

    work page

  2. [2]

    https://dx.doi.org/10.3847/2041-8213/ ab8016

    work page doi:10.3847/2041-8213/ 2041

  3. [3]

    Large Constellations of Low-Altitude Satel- lites: A Primer, 2023

    CBO. Large Constellations of Low-Altitude Satel- lites: A Primer, 2023. https://www.cbo.gov/ publication/59175

    work page 2023

  4. [4]

    Our constellations

    Planet Labs. Our constellations. https://planet. com/our-constellations

    work page

  5. [5]

    LEMUR space plat- form

    Spire Global Inc. LEMUR space plat- form. https://spire.com/space-services/ lemur-space-platform/

    work page

  6. [6]

    Au- tonomous Satellite Wildfire Detection Using Hyperspec- tral Imagery and Neural Networks: A Case Study on Australian Wildfire.Remote Sensing, 15(3):720, 2023

    Kathiravan Thangavel, Dario Spiller, Roberto Saba- tini, Stefania Amici, Sarathchandrakumar Thottuchirayil Sasidharan, Haytham Fayek, and Pier Marzocca. Au- tonomous Satellite Wildfire Detection Using Hyperspec- tral Imagery and Neural Networks: A Case Study on Australian Wildfire.Remote Sensing, 15(3):720, 2023. https://www.mdpi.com/2072-4292/15/3/720

    work page 2023

  7. [7]

    Kouzani, and M

    Arsalan Tahir, Hafiz Suliman Munawar, Junaid Akram, Muhammad Adil, Shehryar Ali, Abbas Z. Kouzani, and M. A. Parvez Mahmud. Automatic Target Detec- tion from Satellite Imagery Using Machine Learning. Sensors, 22(3):1147, 2022. https://www.mdpi.com/ 1424-8220/22/3/1147

    work page 2022

  8. [8]

    Detection of Military Targets from Satellite Images using Deep Convolutional Neural Networks

    Harika Bandarupally, Harshitha Reddy Talusani, and T Sridevi. Detection of Military Targets from Satellite Images using Deep Convolutional Neural Networks. In 2020 IEEE 5th International Conference on Comput- ing Communication and Automation (ICCCA), pages 531–535, 2020. https://ieeexplore.ieee.org/ abstract/document/9250864

    work page arXiv 2020

  9. [9]

    Monitoring agriculture areas with satellite images and deep learning.Applied Soft Computing, 95:106565,

    Thanh Tam Nguyen, Thanh Dat Hoang, Minh Tam Pham, Tuyet Trinh Vu, Thanh Hung Nguyen, Quyet-Thang Huynh, and Jun Jo. Monitoring agriculture areas with satellite images and deep learning.Applied Soft Computing, 95:106565,

    work page

  10. [10]

    https://www.sciencedirect.com/science/ article/pii/S1568494620305032

    work page

  11. [11]

    Satellite data analysis and geographic informa- tion system for urban planning: A systematic review

    Hafsa Ouchra, Abdessamad Belangour, and Allae Er- raissi. Satellite data analysis and geographic informa- tion system for urban planning: A systematic review. In2022 International Conference on Data Analytics for Business and Industry (ICDABI), pages 558–564,

    work page

  12. [12]

    https://ieeexplore.ieee.org/abstract/ document/10041487

    work page arXiv

  13. [13]

    Our constellations: Dove, rapideye, and skysat fleets, 2025

    Planet Labs Inc. Our constellations: Dove, rapideye, and skysat fleets, 2025. Accessed: 2025-04-13

    work page 2025

  14. [14]

    Ushering in the low earth orbit compute cloud era.Com- puter, 58(10):78–86, 2025

    Ketan Bhardwaj, Vaibhav Bhosale, and Ada Gavrilovska. Ushering in the low earth orbit compute cloud era.Com- puter, 58(10):78–86, 2025

    work page 2025

  15. [15]

    A Review on Early Forest Fire Detection Systems Using Optical Re- mote Sensing.Sensors, 20(22):6442, 2020

    Panagiotis Barmpoutis, Periklis Papaioannou, Kosmas Dimitropoulos, and Nikos Grammalidis. A Review on Early Forest Fire Detection Systems Using Optical Re- mote Sensing.Sensors, 20(22):6442, 2020. https: //www.mdpi.com/1424-8220/20/22/6442

    work page 2020

  16. [16]

    van Rijn, Holger H

    Devis Tuia, Konrad Schindler, Begüm Demir, Xiao Xi- ang Zhu, Mrinalini Kochupillai, Sašo Džeroski, Jan N. van Rijn, Holger H. Hoos, Fabio Del Frate, Mihai Datcu, V olker Markl, Bertrand Le Saux, Rochelle Schnei- der, and Gustau Camps-Valls. Artificial Intelligence to Advance Earth Observation: A review of models, recent trends, and pathways forward.IEEE G...

    work page

  17. [17]

    https://ieeexplore.ieee.org/abstract/ document/10669817

    work page arXiv

  18. [18]

    A Lightweight and Adaptive Image Inference Strategy for Earth Observation on LEO Satellites.Re- mote Sensing, 17(7):1175, 2025

    Bo Wang, Yuhang Fang, Dongyan Huang, Zelin Lu, and Jiaqi Lv. A Lightweight and Adaptive Image Inference Strategy for Earth Observation on LEO Satellites.Re- mote Sensing, 17(7):1175, 2025. https://www.mdpi. com/2072-4292/17/7/1175

    work page 2025

  19. [19]

    Known Knowns and Unknowns: Near- realtime Earth Observation Via Query Bifurcation in Serval

    Bill Tao, Om Chabra, Ishani Janveja, Indranil Gupta, and Deepak Vasisht. Known Knowns and Unknowns: Near- realtime Earth Observation Via Query Bifurcation in Serval. In21st USENIX Symposium on Networked Sys- tems Design and Implementation (NSDI 24), pages 809– 824, 2024. https://www.usenix.org/conference/ nsdi24/presentation/tao

    work page 2024

  20. [20]

    L2D2: Low latency distributed downlink for LEO satel- lites

    Deepak Vasisht, Jayanth Shenoy, and Ranveer Chandra. L2D2: Low latency distributed downlink for LEO satel- lites. InProceedings of the 2021 ACM SIGCOMM 2021 Conference, pages 151–164, 2021. https://dl.acm. org/doi/10.1145/3452296.3472932

    work page doi:10.1145/3452296.3472932 2021

  21. [21]

    Planet analytic feeds: Performance metrics, April 2021

    Planet Labs Inc. Planet analytic feeds: Performance metrics, April 2021. Accessed: 2025-04-13

    work page 2021

  22. [22]

    Kodan: Addressing the computational bottleneck in space

    Bradley Denby, Krishna Chintalapudi, Ranveer Chandra, Brandon Lucia, and Shadi Noghabi. Kodan: Addressing the computational bottleneck in space. InProceedings of the 28th ACM International Conference on Architec- tural Support for Programming Languages and Operat- ing Systems, Volume 3, ASPLOS 2023, page 392–403, New York, NY , USA, 2023. Association for ...

    work page 2023

  23. [23]

    Ardagna, Shangguang Wang, and Mengwei Xu

    Chen Yang, Qibo Sun, Qiyang Zhang, Hao Lu, Clau- dio A. Ardagna, Shangguang Wang, and Mengwei Xu. Toward efficient satellite computing through adaptive compression.IEEE Transactions on Services Comput- ing, 17(6):4411–4422, 2024. 12

    work page 2024

  24. [24]

    Orbital edge com- puting: Nanosatellite constellations as a new class of computer system

    Bradley Denby and Brandon Lucia. Orbital edge com- puting: Nanosatellite constellations as a new class of computer system. InProceedings of the Twenty-Fifth International Conference on Architectural Support for Programming Languages and Operating Systems, AS- PLOS ’20, page 939–954, New York, NY , USA, 2020. Association for Computing Machinery

    work page 2020

  25. [25]

    Sentinel Satellite Data, 2025

    European Space Agency. Sentinel Satellite Data, 2025

    work page 2025

  26. [26]

    CER SSF Aqua-FM3-MODIS Edition4A, 2025

    NASA Atmospheric Science Data Center. CER SSF Aqua-FM3-MODIS Edition4A, 2025

    work page 2025

  27. [27]

    Spirepedia: Spire’s satellite constella- tion, 2025

    Spire Global Inc. Spirepedia: Spire’s satellite constella- tion, 2025. Accessed: 2025-04-13

    work page 2025

  28. [28]

    Canadian space agency assigns can$72 million contract to spire global canada to design wild- firesat mission, February 2025

    Spire Global Inc. Canadian space agency assigns can$72 million contract to spire global canada to design wild- firesat mission, February 2025. Accessed: 2025-04-13

    work page 2025

  29. [29]

    Acero, Jonathan Diaz, Ronald Hurtado- Velasco, Sergio Ramiro Gonzalez Bautista, Sonia Rincón, Francisco L

    Ignacio F. Acero, Jonathan Diaz, Ronald Hurtado- Velasco, Sergio Ramiro Gonzalez Bautista, Sonia Rincón, Francisco L. Hernández, Julian Rodriguez- Ferreira, and Jesus Gonzalez-Llorente. A method for validating cubesat satellite eps through power budget analysis aligned with mission requirements.IEEE Ac- cess, 11:43316–43332, 2023

    work page 2023

  30. [30]

    Fedspace: An efficient federated learning framework at satellites and ground stations.arXiv preprint arXiv:2202.01267, 2022

    Jinhyun So, Kevin Hsieh, Behnaz Arzani, Shadi Noghabi, Salman Avestimehr, and Ranveer Chandra. Fedspace: An efficient federated learning framework at satellites and ground stations.arXiv preprint arXiv:2202.01267, 2022

    work page arXiv 2022

  31. [31]

    Cluster Computing 6(3), 215–226 (Jul 2003), https://doi.org/10.1023/A: 1023588520138

    Rich Caruana. Multitask Learning.Machine Learning, 28(1):41–75, 1997. https://doi.org/10.1023/A: 1007379606734

    work page doi:10.1023/a: 1997

  32. [32]

    Unleashing the power of multi- task learning: A comprehensive survey spanning traditional, deep, and pretrained foundation model eras.arXiv preprint arXiv:2404.18961, 2024

    Jun Yu, Yutong Dai, Xiaokang Liu, Jin Huang, Yishan Shen, Ke Zhang, Rong Zhou, Eashan Adhikarla, Wenx- uan Ye, Yixin Liu, Zhaoming Kong, Kai Zhang, Yilong Yin, Vinod Namboodiri, Brian D. Davison, Jason H. Moore, and Yong Chen. Unleashing the Power of Multi- Task Learning: A Comprehensive Survey Spanning Tra- ditional, Deep, and Pretrained Foundation Model...

    work page arXiv 2024

  33. [33]

    Gleason, and Subhransu Maji

    Rangel Daroya, Luisa Vieira Lucchese, Travis Simmons, Punwath Prum, Tamlin Pavelsky, John Gardner, Colin J. Gleason, and Subhransu Maji. Improving Satellite Im- agery Masking using Multi-task and Transfer Learning, 2024.http://arxiv.org/abs/2412.08545

    work page arXiv 2024

  34. [34]

    An Overview of Multi-Task Learning in Deep Neural Networks

    Sebastian Ruder. An Overview of Multi-Task Learning in Deep Neural Networks, 2017. http://arxiv.org/ abs/1706.05098

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  35. [35]

    Cross-stitch Networks for Multi- task Learning, 2016

    Ishan Misra, Abhinav Shrivastava, Abhinav Gupta, and Martial Hebert. Cross-stitch Networks for Multi- task Learning, 2016. http://arxiv.org/abs/1604. 03539

    work page 2016

  36. [36]

    Accelerated Training for Massive Classification via Dynamic Class Selection

    Xingcheng Zhang, Lei Yang, Junjie Yan, and Dahua Lin. Accelerated Training for Massive Classification via Dynamic Class Selection, 2018. http://arxiv.org/ abs/1801.01687

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  37. [37]

    Planet Tasking Product, 2025

    Planet Labs Inc. Planet Tasking Product, 2025

    work page 2025

  38. [38]

    USB Accelerator datasheet, 2019.https: //coral.ai/docs/accelerator/datasheet/

    Google Coral. USB Accelerator datasheet, 2019.https: //coral.ai/docs/accelerator/datasheet/

    work page 2019

  39. [39]

    Nvidia jetson agx orin series

    Leela S Karumbunathan. Nvidia jetson agx orin series. 2022. https://www.nvidia.com/content/ dam/en-zz/Solutions/gtcf21/jetson-orin/ nvidia-jetson-agx-orin-technical-brief.pdf

    work page 2022

  40. [40]

    LandCover.ai: Dataset for Automatic Mapping of Buildings, Woodlands, Water and Roads from Aerial Imagery, 2022

    Adrian Boguszewski, Dominik Batorski, Na- talia Ziemba-Jankowska, Tomasz Dziedzic, and Anna Zambrzycka. LandCover.ai: Dataset for Automatic Mapping of Buildings, Woodlands, Water and Roads from Aerial Imagery, 2022. http://arxiv.org/abs/2005.02264

    work page arXiv 2022

  41. [41]

    Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. EuroSAT: A Novel Dataset and Deep Learning Benchmark for Land Use and Land Cover Classification.IEEE Journal of Selected Top- ics in Applied Earth Observations and Remote Sensing, 12(7):2217–2226, 2019. https://ieeexplore.ieee. org/document/8736785

    work page arXiv 2019

  42. [42]

    PatternNet: A benchmark dataset for per- formance evaluation of remote sensing image retrieval

    Weixun Zhou, Shawn Newsam, Congmin Li, and Zhen- feng Shao. PatternNet: A benchmark dataset for per- formance evaluation of remote sensing image retrieval. ISPRS Journal of Photogrammetry and Remote Sensing, 145:197–209, 2018. https://www.sciencedirect. com/science/article/pii/S0924271618300042

    work page 2018

  43. [43]

    Efficient methods for agile earth observation satellite scheduling

    Xiaoyu Zhao, Zhaohui Wang, Jimin Lv, and Yingwu Chen. Efficient methods for agile earth observation satellite scheduling. In2019 IEEE Symposium Series on Computational Intelligence (SSCI), pages 3158–3164, 2019

    work page 2019

  44. [44]

    Mota Macambira, Celso Barbosa Car- valho, and José Ferreira de Rezende

    Renata do N. Mota Macambira, Celso Barbosa Car- valho, and José Ferreira de Rezende. Energy-efficient routing in LEO satellite networks for extending satel- lites lifetime.Computer Communications, 195:463–475,

    work page

  45. [45]

    https://www.sciencedirect.com/science/ article/pii/S0140366422003486

    work page

  46. [46]

    Yu, Stephen J

    Jianping Gou, B. Yu, Stephen J. Maybank, and Dacheng Tao. Knowledge distillation: A survey.International Journal of Computer Vision, 129:1789 – 1819, 2020. 13

    work page 2020

  47. [47]

    Advancing model pruning via bi-level optimization

    Yihua Zhang, Yuguang Yao, Parikshit Ram, Pu Zhao, Tianlong Chen, Min-Fong Hong, Yanzhi Wang, and Sijia Liu. Advancing model pruning via bi-level optimization. ArXiv, abs/2210.04092, 2022

    work page arXiv 2022

  48. [48]

    Nair, and Divya Mahajan

    Irene Wang, Prashant J. Nair, and Divya Mahajan. Fluid: Mitigating stragglers in federated learning using invari- ant dropout, 2023

    work page 2023

  49. [49]

    Once-for-all: Train one network and specialize it for efficient deployment, 2020

    Han Cai, Chuang Gan, Tianzhe Wang, Zhekai Zhang, and Song Han. Once-for-all: Train one network and specialize it for efficient deployment, 2020

    work page 2020

  50. [50]

    Effec- tive video scene analysis for a nanosatellite based on an onboard deep learning method.Remote Sensing, 15(8), 2023

    Natnael Alemayehu Tamire and Hae-Dong Kim. Effec- tive video scene analysis for a nanosatellite based on an onboard deep learning method.Remote Sensing, 15(8), 2023

    work page 2023

  51. [51]

    Real-time aerial fire detection on resource-constrained devices using knowl- edge distillation.arXiv preprint arXiv:2502.20979, 2025

    Sabina Jangirova, Branislava Jankovic, Waseem Ullah, Latif U Khan, and Mohsen Guizani. Real-time aerial fire detection on resource-constrained devices using knowl- edge distillation.arXiv preprint arXiv:2502.20979, 2025

    work page arXiv 2025

  52. [52]

    In-orbit process- ing or not? sunlight-aware task scheduling for energy- efficient space edge computing networks, 2024

    Weisen Liu, Zeqi Lai, Qian Wu, Hewu Li, Qi Zhang, Zonglun Li, Yuanjie Li, and Jun Liu. In-orbit process- ing or not? sunlight-aware task scheduling for energy- efficient space edge computing networks, 2024

    work page 2024

  53. [53]

    Approximation algorithms for stochastic boolean function evaluation and stochastic submodu- lar set cover,

    Amol Deshpande, Lisa Hellerstein, and Devorah Kletenik, “Approximation algorithms for stochastic boolean function evaluation and stochastic submodu- lar set cover,” inProceedings of the Twenty-Fifth Annual ACM-SIAM Symposium on Discrete Algorithms (SODA), 2014, pp. 1453–1472. 14 APPENDIX A Evaluation Scenarios Details In greater detail, the three scenario...

    work page 2014