Scholarlink

<p>Accurate, high‑resolution mapping of water turbidity is essential for inland ecosystem management, yet conventional field sampling and simple remote‑sensing techniques remain labour‑intensive and scale‑limited. Airborne hyperspectral imagery offers rich spectral detail, but its operational use for quantitative turbidity retrieval has been hampered by sparse ground truth and the lack of scalable, interpretable workflows. To address these gaps, we present an end‑to‑end machine‑learning framework that couples AVIRIS‑NG reflectance (400–800&#xa0;nm, 8&#xa0;m resolution) with different machine learning based ensemble regressors: Random Forest (RF), Gradient Boosting (GB), Extreme Gradient Boosting (XGBoost), and a Voting ensemble (VR), trained on just 33 in situ turbidity samples. We masked water pixels using NDWI &gt; 0, subsampled ~ 100,000 spectra, and tuned models with five-fold cross-validation. Performance was reported on a 30% hold-out test set. XGBoost achieved the highest predictive accuracy (R² = 0.87; RMSE = 4.68 NTU), followed by VR (R² = 0.84; RMSE = 5.33 NTU) and RF (R² = 0.81; RMSE = 5.73 NTU); GB was more conservative (R² = 0.74; RMSE = 6.85 NTU). Notably, despite its lower overall accuracy, RF captured the widest turbidity range, up to 117 NTU, highlighting its strength in extreme-value retrieval. Pearson correlation at diagnostic bands in the mid-visible (619–642&#xa0;nm) and red‑edge (772–787&#xa0;nm) regions yielded <i>r</i> = 0.68–0.90, confirming strong linear links between spectral reflectance and turbidity. Zone‑wise SHAP analyses revealed that mid‑visible bands dominate under low turbidity (&lt; 15 NTU), whereas green and red‑edge wavelengths drive predictions in moderate (15–45 NTU) and high (&gt; 45 NTU) regimes. Our workflow matches or surpasses spectral‑library methods without extensive field campaigns, offering a scalable, interpretable solution for hyperspectral turbidity monitoring that can inform real‑time water‑quality 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

Machine learning-driven spatial-spectral-contextual analysis of airborne hyperspectral data for water quality assessment

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年11月21日

查看全文
  • Anant Dikshit,
  • Vaibhav Garg,
  • Prasun Kumar Gupta

摘要

Accurate, high‑resolution mapping of water turbidity is essential for inland ecosystem management, yet conventional field sampling and simple remote‑sensing techniques remain labour‑intensive and scale‑limited. Airborne hyperspectral imagery offers rich spectral detail, but its operational use for quantitative turbidity retrieval has been hampered by sparse ground truth and the lack of scalable, interpretable workflows. To address these gaps, we present an end‑to‑end machine‑learning framework that couples AVIRIS‑NG reflectance (400–800 nm, 8 m resolution) with different machine learning based ensemble regressors: Random Forest (RF), Gradient Boosting (GB), Extreme Gradient Boosting (XGBoost), and a Voting ensemble (VR), trained on just 33 in situ turbidity samples. We masked water pixels using NDWI > 0, subsampled ~ 100,000 spectra, and tuned models with five-fold cross-validation. Performance was reported on a 30% hold-out test set. XGBoost achieved the highest predictive accuracy (R² = 0.87; RMSE = 4.68 NTU), followed by VR (R² = 0.84; RMSE = 5.33 NTU) and RF (R² = 0.81; RMSE = 5.73 NTU); GB was more conservative (R² = 0.74; RMSE = 6.85 NTU). Notably, despite its lower overall accuracy, RF captured the widest turbidity range, up to 117 NTU, highlighting its strength in extreme-value retrieval. Pearson correlation at diagnostic bands in the mid-visible (619–642 nm) and red‑edge (772–787 nm) regions yielded r = 0.68–0.90, confirming strong linear links between spectral reflectance and turbidity. Zone‑wise SHAP analyses revealed that mid‑visible bands dominate under low turbidity (< 15 NTU), whereas green and red‑edge wavelengths drive predictions in moderate (15–45 NTU) and high (> 45 NTU) regimes. Our workflow matches or surpasses spectral‑library methods without extensive field campaigns, offering a scalable, interpretable solution for hyperspectral turbidity monitoring that can inform real‑time water‑quality management.