EUVIMEDEUVIMED entdecken
Crossref

Self-Supervised STFT Spectrogram-Based Swin Transformer for Plant Ultrasonic Stress Classification

Eu-Tteum Baek, Eui-Joon Kil

Journal of Korea Multimedia Society · 30.06.2026 · 29 (6) · 855-866

DOI: 10.9717/kmms.2026.29.6.855 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Journal of Korea Multimedia Society, erschienen am 30.06.2026. Verfasst von Eu-Tteum Baek und 1 weiteren Autor:innen.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Recent studies have reported that plants emit sounds under stress conditions, drawing attention to plant ultrasonic emissions as a potential signal for non-invasive plant health monitoring. However, reliably classifying plant stress using only ultrasonic signals acquired in real environments remains challenging. This difficulty stems from the inherent limitations of field data. For example, label accuracy is often not fully guaranteed, class distributions are frequently imbalanced, and the recorded events are extremely short, often on the millisecond scale, making purely supervised learning unstable. To address these limitations and achieve more accurate stress classification, this paper proposes a new method for plant ultrasonic data. First, 500 kHz waveform signals are transformed into STFT-based two-dimensional log-magnitude spectrograms and mapped into image-like tensors. Next, a pretrained Swin Transformer is fine-tuned for the target domain through contrastive self-supervised learning, so that invariance to irrelevant perturbations such as noise is encouraged while discriminative features useful for classification are preserved. The adapted model is then combined with an MLP classification head and further fine-tuned in a supervised manner for plant ultrasonic stress classification. Experimental results show that the proposed method achieves the highest accuracy on the target task compared with baseline models. These findings indicate that combining time-frequency representations, transformer-based transfer learning, and contrastive adaptation is effective in limited-label settings.

Zitieren

APA

Baek, E., & Kil, E. (2026). Self-Supervised STFT Spectrogram-Based Swin Transformer for Plant Ultrasonic Stress Classification. Journal of Korea Multimedia Society, 29(6), 855-866. https://doi.org/10.9717/kmms.2026.29.6.855

Vancouver

Baek E, Kil E. Self-Supervised STFT Spectrogram-Based Swin Transformer for Plant Ultrasonic Stress Classification. Journal of Korea Multimedia Society. 2026;29(6):855-866. doi:10.9717/kmms.2026.29.6.855

BibTeX
@article{Baek2026,
  author = {Baek, Eu-Tteum and Kil, Eui-Joon},
  title = {Self-Supervised STFT Spectrogram-Based Swin Transformer for Plant Ultrasonic Stress Classification},
  journal = {Journal of Korea Multimedia Society},
  year = {2026},
  volume = {29},
  number = {6},
  pages = {855--866},
  doi = {10.9717/kmms.2026.29.6.855}
}
Kostenlos starten →