Scholarlink

Abstract <p>Accurate prediction of soil particle size distribution is essential for land use management, soil quality assessment, and erosion control, as fractions like silt and very fine sand are highly sensitive to water and wind erosion. This study aimed to develop an optimal Digital Soil Mapping (DSM) model to predict four soil texture fractions across the Sistan Plain using 400 surface soil samples. A total of 17 models were evaluated, encompassing three main approaches: geostatistical (GS) models, machine learning (ML) models, and one/two stage hybrid models. Soil properties including texture percentages, Organic Matter (OM), Calcium Carbonate Equivalent (CCE), EC, and pH were measured. The results showed that two-stage hybrid approaches (ensemble ML + GS) consistently achieved the highest accuracy across all predicted variables. For silt and VFsand, this was confirmed by the hybrid stacking model combined with Inverse Distance Weighting (IDW) residual modeling. For the test dataset, this model yielded <i>R</i><sup>2</sup> = 0.80 and NRMSE = 0.45 for VFsand, and <i>R</i><sup>2</sup> = 0.65 and NRMSE = 0.60 for silt. The Boosting Cokriging model and the Quantile Random Forest (QRF) + IDW model demonstrated the best performance for predicting clay and sand, respectively. Variable importance analysis indicated that OM and CCE percentages were the key input parameters across all models. In conclusion, two-stage hybrid models, particularly those incorporating the IDW technique to reduce spatial noise, significantly outperform individual models. This approach is highly suitable for creating high-resolution maps of soil texture fractions in areas with similar conditions without requiring additional fieldwork.</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

Optimal Mapping of Soil Texture Fractions: Two-Stage Hybrid Geostatistical and Machine Learning Models Outperform Individual Approaches in the Sistan Plain

  • 期刊论文

  • 发表日期：

    2026年6月23日

查看全文
  • A. Bameri,
  • F. Khormali,
  • F. Kiani,
  • M. R. Pahlavan-Rad,
  • A. Shahriari

摘要

Abstract

Accurate prediction of soil particle size distribution is essential for land use management, soil quality assessment, and erosion control, as fractions like silt and very fine sand are highly sensitive to water and wind erosion. This study aimed to develop an optimal Digital Soil Mapping (DSM) model to predict four soil texture fractions across the Sistan Plain using 400 surface soil samples. A total of 17 models were evaluated, encompassing three main approaches: geostatistical (GS) models, machine learning (ML) models, and one/two stage hybrid models. Soil properties including texture percentages, Organic Matter (OM), Calcium Carbonate Equivalent (CCE), EC, and pH were measured. The results showed that two-stage hybrid approaches (ensemble ML + GS) consistently achieved the highest accuracy across all predicted variables. For silt and VFsand, this was confirmed by the hybrid stacking model combined with Inverse Distance Weighting (IDW) residual modeling. For the test dataset, this model yielded R2 = 0.80 and NRMSE = 0.45 for VFsand, and R2 = 0.65 and NRMSE = 0.60 for silt. The Boosting Cokriging model and the Quantile Random Forest (QRF) + IDW model demonstrated the best performance for predicting clay and sand, respectively. Variable importance analysis indicated that OM and CCE percentages were the key input parameters across all models. In conclusion, two-stage hybrid models, particularly those incorporating the IDW technique to reduce spatial noise, significantly outperform individual models. This approach is highly suitable for creating high-resolution maps of soil texture fractions in areas with similar conditions without requiring additional fieldwork.