Scholarlink

<p>Due to significant nitrogen inputs, nitrate concentrations in groundwater in Germany often exceed drinking water thresholds, necessitating better prediction models and management strategies. This study uses Random Forest models with Explainable AI (SHAP values) to predict nitrate concentrations in groundwater across Baden-Württemberg and Lower Saxony and evaluate the role of denitrification potential. Results show accurate predictions in Baden-Württemberg but notable errors in Lower Saxony. These discrepancies are attributed to unaccounted spatial variability of denitrification potential. SHAP values quantify and visualize its influence. Incorporating chemical parameters such as iron and potassium, closely associated with denitrification, significantly enhances model performance in Lower Saxony (R<sup>2</sup>&#xa0;rising from 0.06 to 0.72). This highlights the role of anoxic conditions and electron acceptors in nitrate reduction. Without chemical parameters, SHAP reveals how spatial predictors in Lower Saxony reflect random patterns rather than causal relationships. Combining Random Forest models and SHAP offers valuable insights into nitrate dynamics and denitrification.</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

Abschätzung des Denitrifikationspotenzials mit Random-Forest-Modellen unter Verwendung hydrochemischer Kovariablen

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年8月27日

查看全文
  • Marc Ohmer,
  • Tanja Liesch,
  • Julian Xanke

摘要

Due to significant nitrogen inputs, nitrate concentrations in groundwater in Germany often exceed drinking water thresholds, necessitating better prediction models and management strategies. This study uses Random Forest models with Explainable AI (SHAP values) to predict nitrate concentrations in groundwater across Baden-Württemberg and Lower Saxony and evaluate the role of denitrification potential. Results show accurate predictions in Baden-Württemberg but notable errors in Lower Saxony. These discrepancies are attributed to unaccounted spatial variability of denitrification potential. SHAP values quantify and visualize its influence. Incorporating chemical parameters such as iron and potassium, closely associated with denitrification, significantly enhances model performance in Lower Saxony (R2 rising from 0.06 to 0.72). This highlights the role of anoxic conditions and electron acceptors in nitrate reduction. Without chemical parameters, SHAP reveals how spatial predictors in Lower Saxony reflect random patterns rather than causal relationships. Combining Random Forest models and SHAP offers valuable insights into nitrate dynamics and denitrification.