Scholarlink

<p>High-accuracy Sea level prediction is important for understanding marine environments and climate change. In this work, a deep convolutional neural network (CNN) combined with attention mechanism (ADNN) is established for sea level anomaly (SLA) prediction from historical satellite observations. Multi-year (1998–2020) radar altimetry observed SLA pattern samples in the South China Sea are used for model training and testing. Compared with existing deep learning models such as CNN and convolutional long short-term memory (ConvLSTM) network, ADNN demonstrates the highest accuracies of 94.0%, 91.1%, 88.4% and 86.2% for 1-d, 3-d, 5-d and 7-d SLA field predictions, with regional average root mean square errors (RMSE) of 0.27 cm, 0.51 cm, 0.80 cm and 1.09 cm, respectively. The integration of CNN and attention mechanism significantly improves the model performance, especially in estimating short-term sea level changes, with a 74.7% reduction in the RMSE for 1-d predictions compared to the baseline CNN model. Comparative experiments also show that the ADNN model performs well when the input data contains a certain degree of noise. Moreover, a multivariate ADNN (M-ADNN) model is designed to investigate the impacts of environmental variables such as sea surface temperature (SST) and wind on SLA prediction. The model yields a slightly higher accuracy but the results are quite similar to those of the ADNN model. The findings suggest that, although SST or wind can affect sea level changes, the ADNN model demonstrates the ability to identify and learn sufficient information about sea level changes solely from satellite altimetry measurements of SLA, especially for relatively long-term (⩾ 5 d) predictions. This eliminates the need for additional input parameter data, thereby improving the SLA prediction efficiency.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

An attention based deep neural network for predicting sea level anomaly field in the South China Sea

  • 期刊论文

  • 发表日期：

    2025年12月1日

查看全文
  • Qing Xu,
  • Yang Zhang,
  • Xiaobin Yin,
  • Haoyu Liu,
  • Yongcun Cheng

摘要

High-accuracy Sea level prediction is important for understanding marine environments and climate change. In this work, a deep convolutional neural network (CNN) combined with attention mechanism (ADNN) is established for sea level anomaly (SLA) prediction from historical satellite observations. Multi-year (1998–2020) radar altimetry observed SLA pattern samples in the South China Sea are used for model training and testing. Compared with existing deep learning models such as CNN and convolutional long short-term memory (ConvLSTM) network, ADNN demonstrates the highest accuracies of 94.0%, 91.1%, 88.4% and 86.2% for 1-d, 3-d, 5-d and 7-d SLA field predictions, with regional average root mean square errors (RMSE) of 0.27 cm, 0.51 cm, 0.80 cm and 1.09 cm, respectively. The integration of CNN and attention mechanism significantly improves the model performance, especially in estimating short-term sea level changes, with a 74.7% reduction in the RMSE for 1-d predictions compared to the baseline CNN model. Comparative experiments also show that the ADNN model performs well when the input data contains a certain degree of noise. Moreover, a multivariate ADNN (M-ADNN) model is designed to investigate the impacts of environmental variables such as sea surface temperature (SST) and wind on SLA prediction. The model yields a slightly higher accuracy but the results are quite similar to those of the ADNN model. The findings suggest that, although SST or wind can affect sea level changes, the ADNN model demonstrates the ability to identify and learn sufficient information about sea level changes solely from satellite altimetry measurements of SLA, especially for relatively long-term (⩾ 5 d) predictions. This eliminates the need for additional input parameter data, thereby improving the SLA prediction efficiency.