EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

Short-Term Forecasting of Ocean Surface Current Maps Using High-Frequency Radar Observations and LSTM Neural Networks: A Case Study of Southwestern Taiwan

Yi-Chieh Lu, Laurence Zsu-Hsin Chuang, Jian-Wu Lai

Remote Sensing · 09.07.2026 · 18 (14) · 2303

DOI: 10.3390/rs18142303 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Remote Sensing, online erschienen am 09.07.2026. Verfasst von Yi-Chieh Lu und 2 weiteren Autor:innen. Die Arbeit nennt 70 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: National Science and Technology Council, Ocean Affairs Council.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Short-term coastal surface-current forecasts at forecast lead times (τ = 1–12 h) are critical for search and rescue (SAR), pollution response, and vessel routing. From a 2015–2019 archive of hourly CODAR high-frequency radar (HFR) observations off Southwestern Taiwan, we developed grid-point long short-term memory (LSTM) models using historical observations alone, without atmospheric forcing or data assimilation (training 2015–2017, validation 2018, test 2019). Because harmonic tides account for ~19% of the surface-current variance, we tested harmonic detiding under a matched architecture and tuning protocol, comparing a raw-input LSTM with a detided variant (LSTM-HA) that forecasts the detided residual and reconstructs the total current. In the out-of-sample 2019 test year (τ = 12 h), LSTM-HA ranked highest (R = 0.768/0.729 for u/v) and reduced RMSE by ~34% relative to Persistence; both LSTM configurations far exceeded HA-Persistence and tide-free HYCOM, and the LSTM-HA advantage was statistically significant and spatially pervasive. An independent single-drifter Lagrangian proof-of-concept (December 2020) gave 12 h mean separations of 8.52/9.20 km for LSTM/LSTM-HA, comparable to the HFR observations (9.26 km) and below HYCOM. For this tide-influenced focus area, the benefit of LSTM-HA emerges from approximately τ = 3 h and becomes most relevant over τ = 6–12 h. At τ = 1–3 h, the raw-input LSTM performs nearly equivalently while forecasting the total current directly. Broader seasonal validation, including monsoon and typhoon forcing, remains a priority.

Förderung

  • National Science and Technology Council
  • Ocean Affairs Council

Zitieren

APA

Lu, Y., Chuang, L. Z., & Lai, J. (2026). Short-Term Forecasting of Ocean Surface Current Maps Using High-Frequency Radar Observations and LSTM Neural Networks: A Case Study of Southwestern Taiwan. Remote Sensing, 18(14), 2303. https://doi.org/10.3390/rs18142303

Vancouver

Lu Y, Chuang LZ, Lai J. Short-Term Forecasting of Ocean Surface Current Maps Using High-Frequency Radar Observations and LSTM Neural Networks: A Case Study of Southwestern Taiwan. Remote Sensing. 2026;18(14):2303. doi:10.3390/rs18142303

BibTeX
@article{Lu2026,
  author = {Lu, Yi-Chieh and Chuang, Laurence Zsu-Hsin and Lai, Jian-Wu},
  title = {Short-Term Forecasting of Ocean Surface Current Maps Using High-Frequency Radar Observations and LSTM Neural Networks: A Case Study of Southwestern Taiwan},
  journal = {Remote Sensing},
  year = {2026},
  volume = {18},
  number = {14},
  pages = {2303},
  doi = {10.3390/rs18142303}
}
Kostenlos starten →