Scholarlink

<p>Excessive leaching of metamitron into soil and water systems poses significant environmental and public health risks. This study integrates biochar physicochemical characterization, hyperspectral reflectance profiling (400–1000&#xa0;nm), and comparative machine-learning models (PLSR, SVM, RF, and ANN) to predict metamitron attenuation in soil–biochar systems. Three biochar types derived from hazelnut shells, apricot kernel shells, and waste car tires were evaluated at application rates of 5%, 15%, and 25% (w/w). GC–MS analyses revealed statistically significant differences in residual metamitron concentrations among biochar types (<i>p</i> &lt; 0.01), with lignocellulosic biochar showing substantially lower residues than tire-derived biochar. The lowest residual concentration (5.18&#xa0;µg&#xa0;L⁻<sup>1</sup>) was observed for apricot kernel shell biochar at the 25% application rate, while higher overall residues persisted in tire-derived biochar treatments. Hyperspectral analysis identified the 600–690&#xa0;nm region as the most diagnostic spectral window, with reflectance at 670&#xa0;nm exhibiting the strongest correlation with metamitron concentration (r = 0.81–0.82, <i>p</i> &lt; 0.01). Comparative modeling demonstrated marked performance differences among algorithms. Support Vector Regression showed limited predictive capability (R<sup>2</sup> = 0.016, RMSE = 11.43), while Random Forest achieved moderate accuracy (R<sup>2</sup> = 0.753, RMSE = 5.73). The Artificial Neural Network (ANN) model outperformed all alternatives, achieving the highest accuracy (R<sup>2</sup> = 0.965, RMSE = 0.122, MAE = 0.080) and demonstrating consistent generalization across the training, validation, and test datasets. Relative importance analysis (Olden and Jackson method) identified biochar type (+ 29.7%) and application rate (− 33.3%) as the dominant predictors, confirming a nonlinear, threshold-dependent adsorption response. In contrast, BET surface area showed a weak negative contribution (− 1.4%). Overall, these findings demonstrate that integrating hyperspectral reflectance with ANN modeling provides a rapid, non-destructive, and mechanistically informative framework for assessing biochar-based pesticide attenuation under controlled laboratory conditions. External field validation across diverse soils and environmental settings is required to fully assess model transferability.</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

Integrating biochar characterization, hyperspectral signatures, and artificial neural networks for predictive modeling of metamitron leachate attenuation

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年3月9日

查看全文
  • A. Altikat,
  • M. H. Alma,
  • S. Altikat,
  • R. Gürbüz

摘要

Excessive leaching of metamitron into soil and water systems poses significant environmental and public health risks. This study integrates biochar physicochemical characterization, hyperspectral reflectance profiling (400–1000 nm), and comparative machine-learning models (PLSR, SVM, RF, and ANN) to predict metamitron attenuation in soil–biochar systems. Three biochar types derived from hazelnut shells, apricot kernel shells, and waste car tires were evaluated at application rates of 5%, 15%, and 25% (w/w). GC–MS analyses revealed statistically significant differences in residual metamitron concentrations among biochar types (p < 0.01), with lignocellulosic biochar showing substantially lower residues than tire-derived biochar. The lowest residual concentration (5.18 µg L⁻1) was observed for apricot kernel shell biochar at the 25% application rate, while higher overall residues persisted in tire-derived biochar treatments. Hyperspectral analysis identified the 600–690 nm region as the most diagnostic spectral window, with reflectance at 670 nm exhibiting the strongest correlation with metamitron concentration (r = 0.81–0.82, p < 0.01). Comparative modeling demonstrated marked performance differences among algorithms. Support Vector Regression showed limited predictive capability (R2 = 0.016, RMSE = 11.43), while Random Forest achieved moderate accuracy (R2 = 0.753, RMSE = 5.73). The Artificial Neural Network (ANN) model outperformed all alternatives, achieving the highest accuracy (R2 = 0.965, RMSE = 0.122, MAE = 0.080) and demonstrating consistent generalization across the training, validation, and test datasets. Relative importance analysis (Olden and Jackson method) identified biochar type (+ 29.7%) and application rate (− 33.3%) as the dominant predictors, confirming a nonlinear, threshold-dependent adsorption response. In contrast, BET surface area showed a weak negative contribution (− 1.4%). Overall, these findings demonstrate that integrating hyperspectral reflectance with ANN modeling provides a rapid, non-destructive, and mechanistically informative framework for assessing biochar-based pesticide attenuation under controlled laboratory conditions. External field validation across diverse soils and environmental settings is required to fully assess model transferability.