Scholarlink

<p>Coastal water quality (CWQ) modelling is essential for sustainable management under increasing anthropogenic and climatic pressures. This study proposed an attention-based deep learning model, TabNet, integrated with the Quokka Swarm Optimisation (QSO) algorithm to enhance predictive accuracy, stability, and interpretability. Long-term in-situ CWQ data from Deep Bay, Hong Kong, were synchronised with AlphaEarth satellite-derived features to capture spatial-temporal variability of ammonia nitrogen (NH<sub>3</sub>-N). Data preprocessing involved temporal alignment, interpolation, and chronological splitting to maintain the integrity of the time series. The baseline model achieved a coefficient of determination (R<sup>2</sup> = 0.83), whereas the QSO-optimised model improved predictive accuracy to (R<sup>2</sup> = 0.91). Explainable Artificial Intelligence (XAI) approaches, including Local Interpretable Model Agnostic Explanations and Partial Dependence Plots, identified signified AlphaEarth features (e.g., A42, A53, and A61) as strong positive contributors to NH<sub>3</sub>-N variability, while features such as A25 and A34 showed negative influence, highlighting the nonlinear role of satellite-derived predictors in coastal nutrient dynamics. Geographic Information System (GIS) based spatial validation confirmed the consistency of model outputs, demonstrating the reliability and scalability of the proposed AI-driven model for sustainable CWQ assessment and 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

Satellite-Informed Transformer Model for Enhanced Coastal Water Quality Modelling

  • 期刊论文

  • 发表日期：

    2026年4月23日

查看全文
  • Mehvish Bilal,
  • Zaher Mundher Yaseen,
  • Syed Masiur Rahman

摘要

Coastal water quality (CWQ) modelling is essential for sustainable management under increasing anthropogenic and climatic pressures. This study proposed an attention-based deep learning model, TabNet, integrated with the Quokka Swarm Optimisation (QSO) algorithm to enhance predictive accuracy, stability, and interpretability. Long-term in-situ CWQ data from Deep Bay, Hong Kong, were synchronised with AlphaEarth satellite-derived features to capture spatial-temporal variability of ammonia nitrogen (NH3-N). Data preprocessing involved temporal alignment, interpolation, and chronological splitting to maintain the integrity of the time series. The baseline model achieved a coefficient of determination (R2 = 0.83), whereas the QSO-optimised model improved predictive accuracy to (R2 = 0.91). Explainable Artificial Intelligence (XAI) approaches, including Local Interpretable Model Agnostic Explanations and Partial Dependence Plots, identified signified AlphaEarth features (e.g., A42, A53, and A61) as strong positive contributors to NH3-N variability, while features such as A25 and A34 showed negative influence, highlighting the nonlinear role of satellite-derived predictors in coastal nutrient dynamics. Geographic Information System (GIS) based spatial validation confirmed the consistency of model outputs, demonstrating the reliability and scalability of the proposed AI-driven model for sustainable CWQ assessment and management.