Scholarlink

<p>The prediction of soil copper (Cu) and zinc (Zn) content by Vis–NIR-SWIR spectroscopy can be performed indirectly due to the correlation of these metals with spectrally active soil organic-mineral compounds (clay, iron oxides and organic matter). However, it is not yet known which is the best combination of machine learning and spectral processing techniques that results in increased accuracy of soil Cu and Zn content predictions. The purposes of this study are (i) to identify absorption profiles mostly correlated to soil available Cu and Zn content and (ii) to assess the effect of different spectral pre-processing techniques and machine learning methods on the accuracy of models adopted to predict soil available Cu and Zn contents. A broad database on vineyards’ subtropical soils (n = 1,482) in Southern Brazil was developed. Cu and Zn contents were extracted by Mehlich-1 procedure. Prediction models were calibrated with spectral data without and after seven spectral pre-processing techniques. Four machine learning methods were tested: Partial Least Square regression (PLSR), Cubist, Support Vector Machine (SVM) and Random Forest (RF). The best spectral pre-processing/machine learning method combination was selected based on models’ validation results, by taking into account the coefficient of determination (R<sup>2</sup>), root mean square error (RMSE), bias, mean absolute error (MAE) and ratios of performance to interquartile range (RPIQ). The combination of Vis–NIR-SWIR spectral data to the RF machine learning method, combined to 1st derivative of Savitzky-Golay spectra (RF + SGD-1d), resulted in the best model to make high accuracy predictions of soil available Cu (R<sup>2</sup> = 0.91; RMSE = 32.46&#xa0;mg&#xa0;kg<sup>−1</sup>; bias = 0.46&#xa0;mg&#xa0;kg<sup>−1</sup>; MAE = 17.19&#xa0;mg&#xa0;kg<sup>−1</sup>; RPIQ = 2.51) and Zn (R<sup>2</sup> = 0.92; RMSE = 3.15&#xa0;mg&#xa0;kg<sup>−1</sup>; bias = -0.71&#xa0;mg&#xa0;kg<sup>−1</sup>; MAE = 2.28&#xa0;mg&#xa0;kg<sup>−1</sup>; RPIQ = 3.71) content at the validation stage. Based on the recorded data, the calibration of prediction models applied to predict the available Cu and Zn content is influenced by the combination of the machine learning method to pre-processing techniques.</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

Combining Vis-NIR-SWIR Spectroscopy and Machine Learning to Predict Soil Available Copper and Zinc in Southern Brazil Vineyards

  • 期刊论文

  • 发表日期：

    2025年4月3日

查看全文
  • Daniely Vaz da Silva Sangoi,
  • Ricardo Simão Diniz Dalmolin,
  • Jean Michel Moura-Bueno,
  • Fabrício de Araújo Pedron,
  • Gustavo Brunetto,
  • Jacson Hindersmann,
  • Douglas Luiz Grando,
  • Agnes Estela Fontana

摘要

The prediction of soil copper (Cu) and zinc (Zn) content by Vis–NIR-SWIR spectroscopy can be performed indirectly due to the correlation of these metals with spectrally active soil organic-mineral compounds (clay, iron oxides and organic matter). However, it is not yet known which is the best combination of machine learning and spectral processing techniques that results in increased accuracy of soil Cu and Zn content predictions. The purposes of this study are (i) to identify absorption profiles mostly correlated to soil available Cu and Zn content and (ii) to assess the effect of different spectral pre-processing techniques and machine learning methods on the accuracy of models adopted to predict soil available Cu and Zn contents. A broad database on vineyards’ subtropical soils (n = 1,482) in Southern Brazil was developed. Cu and Zn contents were extracted by Mehlich-1 procedure. Prediction models were calibrated with spectral data without and after seven spectral pre-processing techniques. Four machine learning methods were tested: Partial Least Square regression (PLSR), Cubist, Support Vector Machine (SVM) and Random Forest (RF). The best spectral pre-processing/machine learning method combination was selected based on models’ validation results, by taking into account the coefficient of determination (R2), root mean square error (RMSE), bias, mean absolute error (MAE) and ratios of performance to interquartile range (RPIQ). The combination of Vis–NIR-SWIR spectral data to the RF machine learning method, combined to 1st derivative of Savitzky-Golay spectra (RF + SGD-1d), resulted in the best model to make high accuracy predictions of soil available Cu (R2 = 0.91; RMSE = 32.46 mg kg−1; bias = 0.46 mg kg−1; MAE = 17.19 mg kg−1; RPIQ = 2.51) and Zn (R2 = 0.92; RMSE = 3.15 mg kg−1; bias = -0.71 mg kg−1; MAE = 2.28 mg kg−1; RPIQ = 3.71) content at the validation stage. Based on the recorded data, the calibration of prediction models applied to predict the available Cu and Zn content is influenced by the combination of the machine learning method to pre-processing techniques.