Scholarlink

<p>Hyperspectral remote sensing technology addressed the limitations of traditional monitoring methods in rapidly acquiring large-scale pollution information, enabling accurate inversion of regional soil heavy metal contents. However, spectral variations caused by complex environmental factors and soil multicomponent coupling effects create prediction accuracy bottlenecks for conventional linear/nonlinear machine learning models using satellite spectra. This study proposes a stacking ensemble framework based on spectral transformation-derived two-dimensional spectral indices (2D-SIs). Piecewise Direct Standardization (PDS) algorithm was implemented to calibrate GF-5 satellite hyperspectral imagery through laboratory spectral band fusion optimized via mutual information maximization. Subsequent processes involved systematic construction of lead (Pb) 2D-SIs and feature selection via Genetic Algorithm optimization and Spearman rank correlation analysis. The framework integrates Partial Least Squares Regression, Support Vector Machines, and eXtreme Gradient Boosting as base learners, with Random Forest meta-model training utilizing meta-features generated via fivefold cross-validation. Experimental results demonstrate that PDS-calibrated 2D-SIs significantly enhance the Pb predictive capability over traditional one-dimensional bands, with the validation set Coefficient of Determination (R<sup>2</sup>) improvements ranging from 0.185 to 0.544 across indices except for the Normalized Product-Difference Index. The stacking models of all indices consistently outperform individual algorithms, with the Normalized Difference Index based model achieving optimal performance (validation set: R<sup>2</sup> = 0.929, RMSE = 325.658&#xa0;mg/kg, RPD = 2.535). This study demonstrates three pivotal advancements in remote sensing for heavy metal inversion: the essential role of spectral calibration in reducing environmental interference, the improved feature extraction enabled by 2D-SIs, and the superior performance of stacking models in modeling complex soil systems. These innovations collectively enhance methodology for regional monitoring applications.</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

A stacking ensemble modeling to map soil lead content: driven by spectral calibration and two-dimensional spectral index with GF-5 hyperspectral data

  • 期刊论文

  • 发表日期：

    2025年8月23日

查看全文
  • Yulan Tang,
  • Zhao Wang,
  • Xiaohan Zhang,
  • Diannan Huang,
  • Yue Feng,
  • Jingli Wang

摘要

Hyperspectral remote sensing technology addressed the limitations of traditional monitoring methods in rapidly acquiring large-scale pollution information, enabling accurate inversion of regional soil heavy metal contents. However, spectral variations caused by complex environmental factors and soil multicomponent coupling effects create prediction accuracy bottlenecks for conventional linear/nonlinear machine learning models using satellite spectra. This study proposes a stacking ensemble framework based on spectral transformation-derived two-dimensional spectral indices (2D-SIs). Piecewise Direct Standardization (PDS) algorithm was implemented to calibrate GF-5 satellite hyperspectral imagery through laboratory spectral band fusion optimized via mutual information maximization. Subsequent processes involved systematic construction of lead (Pb) 2D-SIs and feature selection via Genetic Algorithm optimization and Spearman rank correlation analysis. The framework integrates Partial Least Squares Regression, Support Vector Machines, and eXtreme Gradient Boosting as base learners, with Random Forest meta-model training utilizing meta-features generated via fivefold cross-validation. Experimental results demonstrate that PDS-calibrated 2D-SIs significantly enhance the Pb predictive capability over traditional one-dimensional bands, with the validation set Coefficient of Determination (R2) improvements ranging from 0.185 to 0.544 across indices except for the Normalized Product-Difference Index. The stacking models of all indices consistently outperform individual algorithms, with the Normalized Difference Index based model achieving optimal performance (validation set: R2 = 0.929, RMSE = 325.658 mg/kg, RPD = 2.535). This study demonstrates three pivotal advancements in remote sensing for heavy metal inversion: the essential role of spectral calibration in reducing environmental interference, the improved feature extraction enabled by 2D-SIs, and the superior performance of stacking models in modeling complex soil systems. These innovations collectively enhance methodology for regional monitoring applications.