Support-Gated Spatial Evidence Rejection for Open-Set Event Recognition in Distributed Acoustic Sensing
Distributed acoustic sensing can distinguish activities near a water perimeter, but confident predictions may still be unreliable. We propose support-gated spatial evidence rejection (SGER), a post-training decision rule for a Transformer initialized by masked autoencoder pretraining. Four views each remove a different spatial token band and score the candidate selected from the full image. The lower-quartile view margin is combined with the full-image margin, and an extreme-value class-support gate excludes unsupported candidates. The acceptance threshold is fitted using known-event calibration images. On a retrospective benchmark of 10,000 images, SGER achieves 90.96% overall open-set accuracy and 91.01% macro F1. It retains 92.51% of known events with 95.88% accuracy among accepted known events and rejects all 2000 background images. Overall accuracy exceeds an OpenMax implementation using the same backbone by 2.46 percentage points and a full-image-margin rule using the same support gate by 1.28 percentage points. Gains of 2.47–2.59 percentage points are retained under the tested image-noise and missing-row perturbations. The evaluation focuses on background rejection; the unlabeled MAE pretraining set also includes background images.
Authors
- Ziwen Pan (ORCID: https://orcid.org/0009-0007-6703-0529)
- Hongyu Tao (ORCID: https://orcid.org/0009-0003-4528-0886)
- Zhao Zhang (ORCID: https://orcid.org/0000-0002-7323-5500)
- Na Sun (ORCID: https://orcid.org/0009-0001-4926-5371)
- Mingyang Sun
- Miao Yu
Institutions
- Chinese Academy of Sciences (CN)
- Shenzhen Institutes of Advanced Technology (CN)
- Changchun University (CN)
Publication Details
- Journal
- Sensors
- Published
- 2026-11-19
- DOI
- https://doi.org/10.3390/s26227313
- Primary Topic
- Advanced Neural Network Applications
- Type
- article
- Field-Weighted Citation Impact
- 0.00