Scholarlink

<p>This study is aimed to predict and map agricultural soil organic carbon (SOC) to support decision-makers in implementing sustainable agriculture. Google Earth Engine (GEE) and machine learning algorithms (MLA) are employed to predict SOC distribution in the study area using Sentinel-2&#xa0;A (S2A) and Landsat 8 data. One dataset of 1000 SOC sample points was collected from SoilGrids250 2.0 and filled with topographic attributes, spectral indicators, and SOC parameters to strengthen model performance. Multi-machine learning models (MLMs), such as random forest (RF), extreme gradient boosting (XGB), neural networks (NN), support vector regression (SVR), extra trees (ET), and gradient boosting machine (GBM), were compared using a comparative analysis. The ET model had superior predictive performance, with the highest coefficient of determination (R² of 0.77) and lowest “root mean square error” (RMSE of 3.89) in both data sources. The results indicate that S2A imagery outperforms Landsat in SOC prediction. This means higher-quality satellite images enhance the SOC prediction as they contain more details. The study also validated the efficiency of hybrid ensemble models, such as RF, XGB, SVR, NN, GBM, and ET, in handling complex SOC variations. Notably, ET, XGB, and NN demonstrated better accuracy with S2A data, which shows the significance of multisource data integration in enhancing SOC estimation.</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

Exploring soil organic carbon variability: a machine learning approach and comparison of Sentinel-2 A and Landsat 8 in google earth engine

  • 期刊论文

  • 发表日期：

    2026年4月8日

查看全文
  • Amanullah Adeel,
  • Siddhi Rajendra Ingawale,
  • A. S. Jadhav

摘要

This study is aimed to predict and map agricultural soil organic carbon (SOC) to support decision-makers in implementing sustainable agriculture. Google Earth Engine (GEE) and machine learning algorithms (MLA) are employed to predict SOC distribution in the study area using Sentinel-2 A (S2A) and Landsat 8 data. One dataset of 1000 SOC sample points was collected from SoilGrids250 2.0 and filled with topographic attributes, spectral indicators, and SOC parameters to strengthen model performance. Multi-machine learning models (MLMs), such as random forest (RF), extreme gradient boosting (XGB), neural networks (NN), support vector regression (SVR), extra trees (ET), and gradient boosting machine (GBM), were compared using a comparative analysis. The ET model had superior predictive performance, with the highest coefficient of determination (R² of 0.77) and lowest “root mean square error” (RMSE of 3.89) in both data sources. The results indicate that S2A imagery outperforms Landsat in SOC prediction. This means higher-quality satellite images enhance the SOC prediction as they contain more details. The study also validated the efficiency of hybrid ensemble models, such as RF, XGB, SVR, NN, GBM, and ET, in handling complex SOC variations. Notably, ET, XGB, and NN demonstrated better accuracy with S2A data, which shows the significance of multisource data integration in enhancing SOC estimation.