Scholarlink

<p>Soil organic carbon (SOC) is essential for evaluating soil health and ecosystem functionality. Despite the increasing application of machine learning (ML) algorithms to predict the SOC, their accuracy in high-resolution mapping remains underexplored. Accurate spatial SOC prediction is crucial for advising soil/land management and carbon sequestration strategies. The present study was carried out to assess the prediction performance of different ML algorithms for SOC prediction in surface soil (0–30&#xa0;cm) of South Gujarat region of India at a regional scale. A total of 507 soil samples and 58 environmental covariates (Landsat and Sentinel bands, spectral indices, terrain and bioclimatic variables) selected by Boruta feature selection technique were used to map the SOC in the region. For prediction of SOC, Random Forest (RF), Extreme Gradient Boosting (XGBoost), Cubist and Support Vector Machine (SVM) ML algorithms were employed. The SOC content varied from 0.13 to 3.20% with a mean of 0.95%. The prediction accuracy was obtained superior in the RF model in both calibration and validation datasets (R<sup>2</sup> = 0.923, RMSE = 0.162% and R<sup>2</sup> = 0.474 and RMSE = 0.339%, respectively) followed by the XGBoost model (R<sup>2</sup> = 0.814, RMSE = 0.216% and R<sup>2</sup> = 0.418 and RMSE = 0.355%, respectively). The lowest performance was shown by the SVM and Cubist models, capturing approximately 34.4% and 30.5% of the variance in the calibration dataset, respectively, and 28.4% and 23.7% of the variance in test sets, respectively. Mean Temperature of Wettest Quarter (BIO8) was a key predictor in the RF, XGBoost and SVM models for SOC prediction. The predicted SOC values varied from 0.32 to 2.25%, 0.02 to 2.67%, 0.31 to 2.11% and 1.64 to 7.79% in RF, XGBoost, Cubist and SVM models, respectively. The RF model demonstrated the highest accuracy and reliability for SOC prediction in the study area with a PICP value of 92.2%. These high-resolution SOC maps help in monitoring soil health and provide significant assistance for sustainable land-use management and planning.</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

Regional-scale predictive mapping of soil organic carbon in South Gujarat, India using machine learning algorithms

  • 期刊论文

  • 发表日期：

    2026年3月16日

查看全文
  • Abhishek Jangir,
  • Mahaveer Nogiya,
  • Brijesh Yadav,
  • Lal Chand Malav,
  • Pravash Chandra Moharana,
  • Roshan Lal Meena,
  • Ram Prasad Sharma,
  • Gopal Tiwari,
  • Benukantha Dash,
  • Ravindra Naitam,
  • Duraisamy Vasu,
  • Bansi Lal Mina,
  • Nitin Gorakh Patil

摘要

Soil organic carbon (SOC) is essential for evaluating soil health and ecosystem functionality. Despite the increasing application of machine learning (ML) algorithms to predict the SOC, their accuracy in high-resolution mapping remains underexplored. Accurate spatial SOC prediction is crucial for advising soil/land management and carbon sequestration strategies. The present study was carried out to assess the prediction performance of different ML algorithms for SOC prediction in surface soil (0–30 cm) of South Gujarat region of India at a regional scale. A total of 507 soil samples and 58 environmental covariates (Landsat and Sentinel bands, spectral indices, terrain and bioclimatic variables) selected by Boruta feature selection technique were used to map the SOC in the region. For prediction of SOC, Random Forest (RF), Extreme Gradient Boosting (XGBoost), Cubist and Support Vector Machine (SVM) ML algorithms were employed. The SOC content varied from 0.13 to 3.20% with a mean of 0.95%. The prediction accuracy was obtained superior in the RF model in both calibration and validation datasets (R2 = 0.923, RMSE = 0.162% and R2 = 0.474 and RMSE = 0.339%, respectively) followed by the XGBoost model (R2 = 0.814, RMSE = 0.216% and R2 = 0.418 and RMSE = 0.355%, respectively). The lowest performance was shown by the SVM and Cubist models, capturing approximately 34.4% and 30.5% of the variance in the calibration dataset, respectively, and 28.4% and 23.7% of the variance in test sets, respectively. Mean Temperature of Wettest Quarter (BIO8) was a key predictor in the RF, XGBoost and SVM models for SOC prediction. The predicted SOC values varied from 0.32 to 2.25%, 0.02 to 2.67%, 0.31 to 2.11% and 1.64 to 7.79% in RF, XGBoost, Cubist and SVM models, respectively. The RF model demonstrated the highest accuracy and reliability for SOC prediction in the study area with a PICP value of 92.2%. These high-resolution SOC maps help in monitoring soil health and provide significant assistance for sustainable land-use management and planning.