PaperlayerView exact revision on arXiv

Source-linked AI summary

Raw Imagery Impacting Your AI: Should You Care?

Adrien Dorise, Marjorie Bellizzi, Stéphane May

arXiv:2609.38265v1eess.IVcs.AIcs.CV

TL;DR

Onboard AI may improve mission reactivity and reduce downlink needs, but conventional image-quality criteria may not represent the needs of downstream detection. This study evaluates detectors under image degradation and finds that GSD has the most consistent effect, while severe blur combined with noise marks the most consistent low-quality regime.

  • Problem

    Conventional image-quality criteria may not represent the requirements of downstream AI tasks, while raw optical imagery can have reduced sampling, spatial contrast, and radiometric quality.

    Source: §I · p. 1

  • Method

    The study evaluates ship detection under simulated raw-like imagery degraded through GSD, MTF, and signal-dependent radiometric noise.

    Source: 1) Spatial degradation through the MTF · p. 2, 3) Signal-dependent radiometric noise · p. 2, §V · p. 5

  • Results

    GSD has the most consistent effect, while MTF and SNR responses depend more on resolution and model; severe combined blur and noise gives lower AP50 than corresponding isolated endpoints.

    Source: §IV · p. 4, §V · p. 5

  • Takeaways & Limitations

    Detector performance should be characterized across the expected sensor operating domain to support trade-offs among sensor design, onboard processing, and AI performance.

    Source: §V · p. 6

    Each configuration used a single training run, limiting assessment of performance variability and the significance of small differences.

    Source: §V · p. 6

Abstract

from arXiv · show

Onboard AI is gaining interest for space applications such as vessel, wildfire, and cloud detection, where real-time processing can improve mission reactivity and reduce downlink needs. However, onboard models may operate on raw or minimally processed imagery rather than on restored ground products. This study evaluates how image degradation affects object detection by varying Signal-to-Noise Ratio (SNR), Modulation Transfer Function (MTF) at Nyquist, and Ground Sampling Distance (GSD). Controlled degradations are applied to Very High Resolution Maxar imagery, and three lightweight detectors, YOLOv5s, YOLOX-S, and NanoDet, are evaluated on the resulting operating points. The results show that the impact of image quality depends on the degradation mechanism, and that increasing degradation does not necessarily lead to a proportional decrease in vessel detection performance. GSD produces the most consistent performance shift, while MTF and SNR effects depend more on the model and resolution. Severe combinations of blur and noise produce the largest losses. These results provide task-level information that can support sensor, processing, and AI trade-offs for future onboard systems.

I. INTRODUCTION

Onboard AI may improve mission reactivity and reduce downlink needs, but image-quality criteria for human interpretation may not reflect what downstream detectors require. The study examines task-level sensitivity to simulated raw-like imagery degraded in GSD, MTF, and SNR.

I. INTRODUCTION

  • Onboard inference can improve mission reactivity, reduce downlink volume, and prioritize useful observations for time-sensitive Earth-observation tasks.Examples include vessel, wildfire, and cloud detection.Source: §I · p. 1
  • Conventional image-quality processing may be costly onboard, and visually moderate degradation can remove information important to detection.Sensor requirements based only on conventional image-quality criteria may not represent downstream AI requirements.Source: §I · p. 1
  • The study compares three lightweight detectors across controlled GSD, MTF, and SNR variations to support task-level sensor, processor, and AI co-design.Experiments use GSDs of 100 and 200 cm and separate MTF-only and SNR-only sweeps from a combined sweep.Source: §I · p. 1

A. Image degradation simulation

1) Spatial degradation through the MTF:

The MTF model represents spatial-contrast loss from the optical system and detector. It is converted into a point-spread function for degrading input imagery.

1) Spatial degradation through the MTF:

  • The parametric two-dimensional MTF represents optical-system and detector-induced loss of spatial contrast.Its formulation is defined in the frequency domain.Source: 1) Spatial degradation through the MTF · p. 2
  • The sinc terms approximate detector response along the spatial axes, while the exponential term represents the optical contribution.The parameter γ is determined from a prescribed MTF value at the Nyquist frequency.Source: 1) Spatial degradation through the MTF · p. 2
  • The sampled MTF is converted by inverse Fourier transform into a point-spread function, which is convolved with each input spectral channel.The convolution uses an FFT-based implementation.Source: 1) Spatial degradation through the MTF · p. 2

2) GSD degradation:

The simulation models GSD as a change in spatial sampling, distinct from MTF attenuation of spatial frequencies. It also applies signal-dependent Gaussian noise whose variance depends linearly on signal level.

2) GSD degradation:

  • After MTF degradation, spatial sampling is modified to reproduce the target sensor’s GSD.The image is subsampled using the resampling ratio, distinguishing discrete sampling from MTF attenuation of spatial frequencies.Source: 2) GSD degradation · p. 2

3) Signal-dependent radiometric noise:

  • Radiometric degradation uses a signal-dependent Gaussian noise model with variance assumed to depend linearly on signal level L.The coefficients are determined from two reference operating points.Source: 3) Signal-dependent radiometric noise · p. 2
  • The noise standard deviation is evaluated per pixel, and independent Gaussian noise samples produce signal-varying noise that is clipped to the sensor digital range.The formulation avoids assuming constant variance across the complete image.Source: 3) Signal-dependent radiometric noise · p. 2

B. Dataset

The ship-detection dataset contains 47 high-resolution scenes and more than 24,000 annotated ships across 53 vessel classes.

B. Dataset

  • The dataset contains 47 Maxar-derived scenes at 50 cm GSD, with more than 24,000 annotated ships across 53 vessel classes.The final images are pansharpened from panchromatic and coarser-resolution RGB multispectral bands using weighted Brovey.Source: B. Dataset · p. 2

C. Detection models

The study compares three lightweight object detectors chosen for onboard constraints and reports their detection performance with COCO-style metrics.

C. Detection models

  • YOLOv5s, NanoDet, and YOLOX-S are the three lightweight object-detection architectures evaluated.The study cites onboard latency, memory, and deployment complexity constraints as reasons for selecting lightweight models.Source: C. Detection models · p. 3
  • The models are trained for 100 epochs, averaging approximately 10 hours per model and operating point.Training uses two NVIDIA RTX 4090 GPUs.Source: C. Detection models · p. 3
  • Detection performance is reported using COCO-style AP, AP50, and AP75.Source: C. Detection models · p. 3

D. Experimental design

Controlled experiments at target GSDs of 100 and 200 cm vary MTF and SNR independently and jointly; the original approximately 50 cm imagery is a source condition, not part of the controlled grid.

D. Experimental design

  • The controlled experiments use target GSDs of 100 and 200 cm.The original imagery has a reference GSD of approximately 50 cm but is not included in the controlled MTF and SNR grid.Source: D. Experimental design · p. 3
  • Independent MTF and SNR sweeps isolate spatial blur and radiometric noise, followed by a joint sweep of both degradations.The joint sweep is intended to reproduce combined degradation.Source: D. Experimental design · p. 3
  • The nominal configuration uses MTFNyq = 0.10 and SNR parameters (100, 250), while the sweeps vary these settings.The study deliberately includes severe parameters to make degradation impacts visible.Source: D. Experimental design · p. 3

III. RESULTS

The results focus on AP50 and overall degradation trends. Increasing GSD from 100 to 200 cm consistently reduces performance, though the magnitude varies by detector and operating point.

A. Effect of ground sampling distance

  • GSD increases from 100 to 200 cm reduce detection performance by 3.9 to 7.8 percentage points, depending on configuration.YOLOv5s and YOLOX-S have lower AP50 at 200 cm for all configurations available at both resolutions.Source: A. Effect of ground sampling distance · p. 3
  • NanoDet follows the same overall GSD trend, with one small reversal in the combined MTF and SNR sweep.Source: A. Effect of ground sampling distance · p. 3
  • The GSD effect appears across MTF-only, SNR-only, and combined conditions, but its magnitude depends on detector and operating point.The analysis reports AP, AP50, and AP75 for available configurations, focusing mainly on AP50 and overall trends.Source: A. Effect of ground sampling distance · p. 3, §III · p. 3

B. MTF-only degradation

Severe simultaneous blur and noise produce the largest losses, while MTF and SNR effects vary with resolution and are not consistently proportional to degradation.

B. MTF-only degradation

  • At 100 cm, lower MTF generally reduces performance, while at 200 cm very low MTF tends to hurt but neighboring operating points often reorder.The 100 cm trend is especially clear for YOLOv5s and NanoDet; YOLOX-S trends similarly but not strictly monotonically.Source: B. MTF-only degradation · p. 3

C. SNR-only degradation

  • SNR performance is relatively stable over much of the tested range at 100 cm, but severe noise causes greater losses, especially at 200 cm.At 200 cm, YOLOv5s and NanoDet worsen under severe noise, while YOLOX-S varies at intermediate settings but its lowest-SNR results remain below its best SNR-only points.Source: C. SNR-only degradation · p. 3

D. Combined MTF and SNR degradation

  • The lowest-quality combined MTF and SNR configurations produce substantially lower AP50 than corresponding single-degradation cases across all three detectors.The drop is sharp at 200 cm in the final two combined operating points; performance deteriorates once both degradation mechanisms become strong.Source: D. Combined MTF and SNR degradation · p. 3

IV. DISCUSSION

GSD produces the clearest performance shift, while MTF and SNR effects vary with detector and resolution. Severe combined blur and noise mark the most consistent low-quality regime, informing trade-offs among sensor quality, restoration, and onboard inference.

IV. DISCUSSION

  • Under matched controlled configurations, GSD at 100 cm outperforms 200 cm in mean AP50 by 6.7 and 5.0 points; NanoDet favors 100 cm in 22 of 23 matched cases.Source: §IV · p. 4, Figure 2 · p. 4
  • At fixed GSD, MTF and SNR have no universal ranking: severe MTF affects YOLOv5s and NanoDet more at 100 cm, while YOLOv5s is strongly SNR-sensitive at 200 cm.NanoDet and YOLOX-S show more non-monotonic behavior at 200 cm.Source: §IV · p. 4
  • For every model at both GSDs, the strongest combined blur-and-noise setting lowers AP50 more than the corresponding isolated MTF or SNR endpoint.The mixed sweep contains only eight paired operating points, so its interpolated surfaces are visual aids rather than complete response-surface measurements.Source: §IV · p. 4
  • Moderate degradation does not necessarily cause substantial losses: performance often plateaus at moderate-to-high quality, with neighboring operating points sometimes reversing order.The authors note that these fluctuations may reflect both a performance plateau and stochastic training variability.Source: §IV · p. 4
  • Expected sensor conditions and acceptable task accuracy guide whether to use direct inference, onboard restoration, or degradation-aware detection.Performance thresholds across operating points can also inform sensor-design requirements.Source: §IV · pp. 4–5

V. CONCLUSION

Across the evaluated detectors, GSD has the most consistent effect, while MTF and SNR responses vary with spatial resolution and can be non-monotonic. Small-to-moderate degradation is tolerated within the tested range, but combined severe blur and noise produce the most consistent low-quality regime.

V. CONCLUSION

  • In simulated raw-like optical imagery evaluated with YOLOv5s, NanoDet, and YOLOX-S, GSD produced the most consistent effect: performance generally decreased at 200 cm compared with 100 cm.Source: §V · p. 5
  • Small-to-moderate degradation does not cause a sudden performance breakdown, while severe blur and noise together produce the most consistent failure regime.Under combined severe degradation, AP50 decreases more than under the corresponding isolated degradations.Source: §V · pp. 5–6
  • Because MTF and SNR responses are resolution-dependent and often non-monotonic, sensor requirements should be assessed across the expected operating domain rather than inferred from a monotonic parameter relationship.Such task-oriented evaluation can support trade-offs among sensor design, onboard processing, and AI performance.Source: §V · pp. 5–6
  • Each configuration used a single training run, limiting the ability to separate systematic degradation effects from training variability.The authors recommend repeated runs with multiple random seeds and variability measures or confidence intervals.Source: §V · p. 6

Related papers

Loading 2609.38265v1…