Scholarlink

Purpose <p>Iron oxides are essential mineral components in soil, serving as key indicators of soil maturity, redox conditions, and the degree of organic matter accumulation. Therefore, accurate and timely monitoring of soil iron oxide content is of significant practical importance for land management, soil health assessment, and ecological restoration.</p> Materials and methods <p>This study collected soil samples from a typical karst region in Guangxi and used five smartphones to capture images of the soil samples. Visible spectral information (RGB color parameters) was extracted from the images and transformed into different color space models. Simultaneously, high-spectral data of the soil were obtained using an Analytical Spectral Devices (ASD) FieldSpec 3 spectrometer. A comparative analysis was conducted to explore the feasibility of using smartphone images to estimate the content of iron oxides. The study also employed three algorithms, including Backpropagation Neural Network (BPNN), Support Vector Machine (SVM), and Random Forest (RF), to predict the content of free iron oxides (Fe<sub>d</sub>), total iron content (Fe<sub>t</sub>), and iron activity ratio (Fe<sub>d</sub>/Fe<sub>t</sub>*100%).</p> Results and conclusions <p>The results indicate that most inversion models based on color parameters captured by smartphones achieved suitable accuracy, meeting the requirements for modeling. The spectral acquisition capabilities of smartphones were found to be comparable to, or slightly lower than, those of the spectrometer. This suggests that smartphones are capable of effectively acquiring soil spectral data and facilitating quantitative inversion studies of soil properties. Predictions of Fe<sub>d</sub>, Fe<sub>t</sub>, and Fe<sub>d</sub>/Fe<sub>t</sub>*100% were made using the three algorithms, and all models exhibited reasonable performance. A comprehensive analysis of prediction accuracy, absolute error maps, and scatter plots revealed that the spectral response characteristics of different smartphones significantly influenced inversion accuracy. The prediction performance, ranked from highest to lowest, was as follows: Phone 5 &gt; Phone 4 &gt; Phone 3 &gt; Phone 2 &gt; Phone 1, with Phone 5 showing the best prediction accuracy. Among the inversion models, the RF outperformed both BPNN and SVM, exhibiting superior accuracy and stability.</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 machine learning-based model for soil iron oxides estimation using fused hyperspectral and multispectral data

  • 期刊论文

  • 发表日期：

    2026年4月16日

查看全文
  • Yongshi Tan,
  • Zhenxi Wei,
  • Yan Xiao,
  • Yulin Huang,
  • Zongxin Li,
  • Shuting Yang,
  • Lin Zou,
  • Lanhui Yang,
  • Yusong Deng

摘要

Purpose

Iron oxides are essential mineral components in soil, serving as key indicators of soil maturity, redox conditions, and the degree of organic matter accumulation. Therefore, accurate and timely monitoring of soil iron oxide content is of significant practical importance for land management, soil health assessment, and ecological restoration.

Materials and methods

This study collected soil samples from a typical karst region in Guangxi and used five smartphones to capture images of the soil samples. Visible spectral information (RGB color parameters) was extracted from the images and transformed into different color space models. Simultaneously, high-spectral data of the soil were obtained using an Analytical Spectral Devices (ASD) FieldSpec 3 spectrometer. A comparative analysis was conducted to explore the feasibility of using smartphone images to estimate the content of iron oxides. The study also employed three algorithms, including Backpropagation Neural Network (BPNN), Support Vector Machine (SVM), and Random Forest (RF), to predict the content of free iron oxides (Fed), total iron content (Fet), and iron activity ratio (Fed/Fet*100%).

Results and conclusions

The results indicate that most inversion models based on color parameters captured by smartphones achieved suitable accuracy, meeting the requirements for modeling. The spectral acquisition capabilities of smartphones were found to be comparable to, or slightly lower than, those of the spectrometer. This suggests that smartphones are capable of effectively acquiring soil spectral data and facilitating quantitative inversion studies of soil properties. Predictions of Fed, Fet, and Fed/Fet*100% were made using the three algorithms, and all models exhibited reasonable performance. A comprehensive analysis of prediction accuracy, absolute error maps, and scatter plots revealed that the spectral response characteristics of different smartphones significantly influenced inversion accuracy. The prediction performance, ranked from highest to lowest, was as follows: Phone 5 > Phone 4 > Phone 3 > Phone 2 > Phone 1, with Phone 5 showing the best prediction accuracy. Among the inversion models, the RF outperformed both BPNN and SVM, exhibiting superior accuracy and stability.