Scholarlink

<p>Accurate monitoring of coastal eutrophication is critical for maintaining marine ecosystem health and supporting environmental management. However, retrieving eutrophication indicators through remote sensing remains a persistent challenge due to the complex nonlinear relationships between satellite signals and water chemistry. In this study, we developed a hybrid transformer-support vector regression (SVR) model to overcome these limitations. This novel architecture bridges the gap by synergizing deep spatiotemporal feature extraction with robust small-sample regression, providing a scalable framework for coastal eutrophication monitoring. Applied to Qingdao coastal waters (2000–2022) using the Moderate-resolution Imaging Spectroradiometer (MODIS) data, the proposed model achieved state-of-the-art accuracy for chemical oxygen demand (COD) (<i>R</i><sup>2</sup>=0.703) and soluble reactive phosphorus (SRP) (<i>R</i><sup>2</sup>= 0.651), surpassing classical Random Forest (RF), 1D-CNN, and hybrid baselines (CNN/LSTM-SVR). While dissolved inorganic nitrogen (DIN) retrieval remains challenging (<i>R</i><sup>2</sup>=0.362) due to its non-optical nature and data resolution constraints, the model exhibited notable error stability. Spatiotemporal analysis revealed that water depth, latitude, and tidal flat proximity regulate spatial heterogeneity, while land-based nutrient fluxes and sediment-water interface processes drive summer-autumn eutrophication peaks. This research offers a robust machine learning approach for long-term coastal monitoring and valuable geospatial insights for coastal environmental management.</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

Hybrid transformer-SVR framework for coastal eutrophication assessment through satellite-based retrieval of non-optically active water quality parameters

  • 期刊论文

  • 发表日期：

    2026年3月7日

查看全文
  • Chengyang Guan,
  • Longkun Zhang,
  • Qingchun Guan,
  • Xiaoxue Tang,
  • Junwen Chen

摘要

Accurate monitoring of coastal eutrophication is critical for maintaining marine ecosystem health and supporting environmental management. However, retrieving eutrophication indicators through remote sensing remains a persistent challenge due to the complex nonlinear relationships between satellite signals and water chemistry. In this study, we developed a hybrid transformer-support vector regression (SVR) model to overcome these limitations. This novel architecture bridges the gap by synergizing deep spatiotemporal feature extraction with robust small-sample regression, providing a scalable framework for coastal eutrophication monitoring. Applied to Qingdao coastal waters (2000–2022) using the Moderate-resolution Imaging Spectroradiometer (MODIS) data, the proposed model achieved state-of-the-art accuracy for chemical oxygen demand (COD) (R2=0.703) and soluble reactive phosphorus (SRP) (R2= 0.651), surpassing classical Random Forest (RF), 1D-CNN, and hybrid baselines (CNN/LSTM-SVR). While dissolved inorganic nitrogen (DIN) retrieval remains challenging (R2=0.362) due to its non-optical nature and data resolution constraints, the model exhibited notable error stability. Spatiotemporal analysis revealed that water depth, latitude, and tidal flat proximity regulate spatial heterogeneity, while land-based nutrient fluxes and sediment-water interface processes drive summer-autumn eutrophication peaks. This research offers a robust machine learning approach for long-term coastal monitoring and valuable geospatial insights for coastal environmental management.