Skip to content

Trajectory-Agnostic Asteroid Detection in TESS with Deep Learning

Home / Papers / Trajectory-Agnostic Asteroid Detection in TESS with Deep Learning

Trajectory-Agnostic Asteroid Detection in TESS with Deep Learning

Authors: Brian P. Powell, Jorge Martínez-Palomera, A. Tuson, Christina Hedges, Jessie Dotson, Jordan Caraballo-Vega Date: 2026-05-12 Paper ID: openalex:2605.12391

Summary

This paper introduces a trajectory-agnostic method for detecting moving objects in TESS image time-series data using a stacked 3D U-Net architecture termed W-Net. Unlike traditional “shift-and-stack” algorithms that require prior assumptions about object velocity and direction, this approach is robust to varying asteroid trajectories. The model incorporates a novel Adaptive Normalization layer to learn optimal data scaling directly from the inputs. The researchers also released the tess-asteroid-ml repository to support community-wide training for future space surveys.

Key Contributions

  • Introduces W-Net, a stacked 3D U-Net architecture that eliminates the need for explicit trajectory assumptions in asteroid detection.
  • Proposes Adaptive Normalization to automatically learn optimal scaling distributions for image time-series.
  • Provides the tess-asteroid-ml codebase to facilitate training and benchmarking in moving object detection.

Open Questions & Future Work

Key Concepts

  • w-net: A deep learning architecture composed of two stacked 3D U-Nets with skip connections for spatiotemporal feature extraction in image time-series.
  • adaptive-normalization: A learned data scaling technique that allows neural networks to determine the optimal range and distribution for input data processing dynamically.

Archivist Review

Approved the W-Net architecture and Adaptive Normalization mechanism as both represent reusable methodologies for time-series image processing. The open question regarding label quality in astronomical datasets was approved as a fundamental bottleneck, whereas the request regarding fast-moving objects was rejected as a routine data augmentation task.

Approved Concepts

  • W-Net: Central architecture for spatiotemporal feature extraction in moving object detection.
  • Adaptive Normalization: Core technique for automating input data scaling learning, replacing static preprocessing.

Approved Open Questions

  • Moving Object Label Quality Bottleneck: This is a foundational performance bottleneck for supervised learning in astronomical surveys.

Rejected Candidates

  • [open_question] Detecting Fast-Moving Asteroids (fast-asteroid-detection-bottleneck) - other: This is a specific training data limitation (synthetic data injection) rather than a fundamental research bottleneck.

Metadata & Links

url
https://arxiv.org/abs/2605.12391
paper_id
2605.12391
paper_source
openalex
domain
computer-vision
tags
computer-vision
architectures
datasets
concept_slugs
w-netadaptive-normalization
dataset_slugs
skill
TimeSeriesSkill
created_at
2026-05-15T06:16:06Z