Scholarlink

<p>This study presents a multi-scale assessment framework that integrates multi-source remote sensing and machine learning techniques to evaluate soil physical and chemical characteristics, focusing on estimating topsoil organic carbon content and soil material strength in the Del Centro region, Mexico. A comprehensive suite of 389 predictive variables, including spectral indices, elevation models at varying resolutions, terrain derivatives, lithological units, and categorical land surface information, was assembled to test nine model scenarios. These scenarios combined variables at different spatial scales to evaluate their influence on model performance. Among the scenarios, the model using terrain derivatives and categorical inputs achieved the highest performance (<i>R</i><sup>2</sup>&#xa0;=&#xa0;0.84, adj <i>R</i><sup>2</sup>&#xa0;=&#xa0;0.81), with a low MAE of 0.16 and GTW-MAE of 0.98. Sentinel spectral data alone also performed robustly (<i>R</i><sup>2</sup>&#xa0;=&#xa0;0.83), while models based solely on lithological units showed limited predictive capability (<i>R</i><sup>2</sup>&#xa0;=&#xa0;0.53, adj <i>R</i><sup>2</sup>&#xa0;=&#xa0;0.51). Models incorporating integrated predictors (all variables) performed well overall (<i>R</i><sup>2</sup>&#xa0;=&#xa0;0.81), indicating the benefit of fusing diverse data sources. Cross-validation results supported the consistency of the models, with CV-<i>R</i><sup>2</sup> ranging from 0.39 to 0.56 across different configurations. Top-performing models demonstrated that combining terrain attributes and spectral information significantly improves the estimation of soil properties, including material strength. In contrast, reliance on coarse-resolution elevation data or single-category variables resulted in diminished predictive accuracy. The findings also underscore that integrating variables at multiple analysis scales leads to considerable improvements in model robustness, with adjusted <i>R</i><sup>2</sup> gains up to 73% compared to the least effective model. This research highlights the potential of combining advanced remote sensing products and multi-scale data analytics to support more precise digital soil mapping.</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

Multi-scale strength and characteristic assessment of soil materials using multi-source remote sensing and machine learning

  • 期刊论文

  • 发表日期：

    2025年10月23日

查看全文
  • S Thenmozhi,
  • Amal Alshardan,
  • Abdulaziz G Alghamdi,
  • T Pradeep

摘要

This study presents a multi-scale assessment framework that integrates multi-source remote sensing and machine learning techniques to evaluate soil physical and chemical characteristics, focusing on estimating topsoil organic carbon content and soil material strength in the Del Centro region, Mexico. A comprehensive suite of 389 predictive variables, including spectral indices, elevation models at varying resolutions, terrain derivatives, lithological units, and categorical land surface information, was assembled to test nine model scenarios. These scenarios combined variables at different spatial scales to evaluate their influence on model performance. Among the scenarios, the model using terrain derivatives and categorical inputs achieved the highest performance (R2 = 0.84, adj R2 = 0.81), with a low MAE of 0.16 and GTW-MAE of 0.98. Sentinel spectral data alone also performed robustly (R2 = 0.83), while models based solely on lithological units showed limited predictive capability (R2 = 0.53, adj R2 = 0.51). Models incorporating integrated predictors (all variables) performed well overall (R2 = 0.81), indicating the benefit of fusing diverse data sources. Cross-validation results supported the consistency of the models, with CV-R2 ranging from 0.39 to 0.56 across different configurations. Top-performing models demonstrated that combining terrain attributes and spectral information significantly improves the estimation of soil properties, including material strength. In contrast, reliance on coarse-resolution elevation data or single-category variables resulted in diminished predictive accuracy. The findings also underscore that integrating variables at multiple analysis scales leads to considerable improvements in model robustness, with adjusted R2 gains up to 73% compared to the least effective model. This research highlights the potential of combining advanced remote sensing products and multi-scale data analytics to support more precise digital soil mapping.