Kevin Phan, EO Solutions; David Chaparro, EO Solutions; Zach Gazak, Space Systems Command A&AS
https://doi.org/10.64861/XUQQ5883
Keywords: Space Domain Awareness, Space Situational Awareness, Low Signal-to-Noise Ratio Detection, SNR-Aware Training, Detectability Modeling, Optical Space Object Detection, Imbalanced Training Dynamics, Loss Reweighting
Abstract:
Optical detection of resident space objects (RSOs) is fundamentally constrained by extreme variation in signal-to-noise ratio (SNR) across observed targets. While modern deep object detectors perform well on bright objects, detection reliability degrades sharply for dim RSOs near the noise floor. In operational environments, this leads to inconsistent sensitivity across observing conditions, reduced confidence at detection thresholds, and increased risk of missed detections during marginal imaging periods. We argue that this limitation is not primarily a data volume problem, but a detectability imbalance problem: high-SNR objects dominate training dynamics, biasing models toward brightness-driven cues and suppressing learning from noise-limited observations._x000D_
_x000D_
We present an SNR-aware detection framework that explicitly incorporates physically meaningful detectability estimates into dataset construction, sampling strategy, and loss scaling. Central to this approach is a per-instance SNR estimation pipeline computed directly from image data using Gaussian point spread function (PSF) fitting combined with aperture photometry. For each annotated RSO, a two-dimensional Gaussian model is fit to estimate the object centroid and spatial extent. The object’s signal is measured by integrating pixel intensities within an aperture scaled to the fitted PSF width, ensuring consistent capture of object flux. Local background noise is estimated from a surrounding annular region to reflect representative sky statistics and structured background variability. The SNR is computed as the background-subtracted integrated signal divided by the combined noise contribution from photon statistics and measured background variance within the aperture. This produces a continuous, physically interpretable detectability metric that can be computed retroactively on existing datasets without requiring changes to sensor hardware, labeling procedures, or detector architectures._x000D_
_x000D_
Using these SNR estimates, we reorganize datasets around detectability rather than categorical labels. Objects are stratified into SNR ranges, enabling explicit control over how low-, medium-, and high-SNR regimes are presented during training. We introduce an instance-level SNR-aware sampling strategy that increases exposure to noise-limited RSOs while preserving realistic background structure and contextual variability. Unlike naïve oversampling, this approach avoids excessive duplication of images and maintains operationally representative scene statistics._x000D_
_x000D_
To further mitigate optimization bias, we incorporate SNR-aware loss scaling. Detection losses are reweighted as a bounded function of measured SNR, increasing the relative influence of dim objects while preventing instability from extreme inverse weighting. This directly counteracts gradient dominance from bright objects and ensures that noise-limited RSOs meaningfully influence parameter updates. The method is detector-agnostic and integrates seamlessly with standard one-stage and two-stage architectures._x000D_
_x000D_
We evaluate the framework on a dataset of 49,340 optical images collected from a single operational sensor, spanning a wide range of observing conditions and SNR regimes. Detection performance is assessed across three backbones, Faster R-CNN, RetinaNet, and YOLOX. RetinaNet demonstrates the largest performance gains across low- and mid-SNR bins while maintaining strong performance for bright objects on a held-out partition. For RSOs nearest the noise floor, precision, recall, and F1 all improve substantially over the baseline. In low- to mid-SNR regimes, F1 improves markedly, driven primarily by large recall gains. Even in moderately bright regimes, recall improves while maintaining high precision, and performance on very high-SNR objects remains effectively unchanged, preserving near-perfect detection rates._x000D_
_x000D_
These results demonstrate that explicitly incorporating detectability into training substantially flattens SNR-dependent performance curves and extends reliable detection capability closer to the noise floor. Operationally, this translates to improved sensitivity during suboptimal seeing conditions, shorter exposures, and lower apparent magnitudes without synthetic data generation, cross-sensor normalization, or architectural modification._x000D_
_x000D_
Because SNR is estimated directly from image data using standard photometric principles, the framework can be retroactively applied to legacy datasets and integrated into existing training pipelines with minimal computational overhead. By reframing RSO detection as a detectability-aware optimization problem, this work provides a practical and scalable solution to SNR imbalance and offers a principled pathway for extending low-SNR detection capability in operational space surveillance systems._x000D_
Date of Conference: September 15-18, 2026
Track: AI for SDA Applications