Aidan Lorenz, ARKA; Shawn Abernethy, ARKA; Mike Fischer, ARKA; Jacob Griesbach, ARKA; W. Jody Mandeville, InTrack Radar Technologies; Sid Arora, InTrack Radar Technologies; Tim McLaughlin, InTrack Radar Technologies; Harshitha Challa, West Virginia University; Piyush Mehta, West Virginia University
Keywords: machine learning, object detection, convolutional neural networks, CNNs, LNT, debris detection
Abstract:
With the increased focus on discovering and tracking orbital debris smaller than 10 centimeters, efforts remain underway to push detection capabilities closer to the noise floor in ground-based electro-optical (EO) sensor data. Traditionally, rate tracking with extended integration times is used to enhance the signal-to-noise ratio of dim targets, enabling detection against background noise. This technique does not work for discovering dim objects that are not already cataloged since the motion of such targets is not known. Other methods such as multi-frame velocity matched filtering are computationally intensive and tedious. As part of the IARPA SINTRA program, we explore training a machine learning detection model on a large simulated multi-frame dataset of sidereal collects focused on low Earth orbit (LEO) resident space objects (RSOs) of known visual magnitudes and benchmark this model on a comparable holdout test dataset consisting of real data.
To supplement real-world data, SatSim, an open-source tool, was leveraged to generate a large-scale synthetic dataset resembling the data from a specific ground-based sensor. The dataset includes RSOs with varying visual magnitudes extending down to the noise floor while maintaining pixel-perfect bounding boxes. The simulated data was used to create a dataset sufficient to train and validate machine learning models and further benchmark them as a function of visual magnitude. Real data was collected from Pine Park Observatory (PPO) in Colorado Springs using their narrow field of view (NFOV) EO sensor. This sensor was the one modeled in the SatSim simulated dataset. A traditional single-frame detection scheme was applied to label the centroids of all conventionally detected RSOs within the field of view (FOV) of the sensor. We then developed a primitive blob detection algorithm to infer approximate RSO bounding boxes from the centroid information. The real data was used to create a holdout dataset on which to run inference using the object detection model trained on simulated data in an effort to find RSOs not identified by the traditional detector used for labeling the real data.
With the synthetic tagged EO dataset, we trained various deep learning models, all using the EfficientDet convolutional neural network architecture, for RSO detection exploring performance differences when modifying several factors in the data including single-frame vs multi-frame datapoints and 8-bit vs 16-bit imagery. With each model, our initial benchmark evaluates detection performance as a function of visual magnitude, using the magnitude metadata associated to RSOs in the simulated dataset. A key aspect of this study is the integration of temporal information. Formatting model inputs as 3-channel temporal images – as opposed to single channel frames – provides the model with temporal information theoretically allowing the model to implicitly learn filtering techniques to pull dim RSOs out of the noise.
We then assess model generalization to real data, using models trained exclusively on simulated data to run inference on real data (in addition to some experimentation with finetuning on real data). When performing inference on the real data, we flag false positives for further study to determine if these are potentially RSOs that the traditional detector missed. These results serve as benchmarks for comparing RSO detection between the traditional physics-based detector used to label the real data, and the deep learning model.
Small RSOs remain difficult to detect using traditional techniques, and as part of the IARPA SINTRA program, the team’s efforts aim to characterize the performance of deep learning models commonly used for object detection against this difficult class of targets in ground-based EO imagery. The goal of the IARPA SINTRA program is to improve the current state-of-the-art to detect, track, and characterize Lethal Non-Trackable (LNT) space debris.
Date of Conference: September 16-19, 2025
Track: Machine Learning for SDA Applications