Scholarlink

<p>Soil load-bearing capacity (SLBC) is a critical determinant of sustainable soil management and the efficient operation of agricultural machinery, yet continuous field-based assessment is constrained by high costs and labor demands. This study presents a novel, spatially explicit remote sensing–machine learning (RS–ML) framework for mapping SLBC across croplands in northeastern Bangladesh, particularly in haor regions. Field observations from 127 agricultural sites were collected using a Dynamic Cone Penetrometer (DCP), with penetration-per-blow values converted to DCP indices and subsequently transformed into California Bearing Ratio (CBR) and SLBC. A predictor set, including soil properties, topographic derivatives, and spectral indices, was preprocessed in Google Earth Engine and harmonized into raster layers at ~ 250&#xa0;m resolution. Extreme Gradient Boosting (XGBoost), Gradient Boosting Machine (GBM), and Random Forest (RF) models were trained using hyperparameter tuning, 10-fold cross-validation, and a 70/30 train–test split. During 10-fold cross-validation, XGBoost (MSE 3.03–43.57, RMSE 1.74–6.60, MAE 1.49–3.92, R² 0.567–0.955) and GBM (MSE 2.81–46.60, RMSE 1.68–6.83, MAE 1.32–3.75, R² 0.531–0.958) showed better predictive performance for SLBC than RF (MSE 6.59–60.91, RMSE 2.57–7.80, MAE 2.14–4.44,R² 0.230–0.901). However, RF achieved lower error on the test dataset, while the Wilcoxon signed-rank test showed no statistically significant differences among the models. Bulk Density was identified as the most influential factor shaping spatial variability in soil load-bearing capacity (SLBC). Higher SLBC in Sylhet Sadar and lower values in Tahirpur and Bishwambarpur were observed. Limitations include relatively homogeneous soil conditions, limited field coverage, lack of seasonal variability, and restricted model extrapolation beyond training data. Despite these constraints, the proposed ML–RS framework provides spatially explicit SLBC predictions that can inform data-driven decision-making and support mechanized farming in wetland-prone croplands of northeastern Bangladesh and similar agroecosystems.</p> Graphical Abstract <p></p> <p>This graphical abstract presents a spatially explicit, multi-feature-based machine learning–remote sensing (ML–RS) framework for mapping Soil Load Bearing Capacity (SLBC) across persistent croplands in northeastern Bangladesh. Field-based SLBC measurements were collected from 127 locations using a Dynamic Cone Penetrometer (DCP). Penetration-per-blow values were converted into DCP indices and subsequently transformed into California Bearing Ratio (CBR) and SLBC using established empirical relationships, which served as the model target. A comprehensive suite of multi-source remote sensing predictors was assembled, including soil intrinsic properties (bulk density, sand, silt, and clay), hydro-thermal variables (soil moisture and soil temperature), topographic derivatives (slope, aspect, curvature, and topographic wetness index), and vegetation- and water-related spectral indices (NDVI and NDWI). All predictor layers were harmonized to ~ 250&#xa0;m spatial resolution across croplands using integrated workflows in Google Earth Engine, ArcGIS, and Python. Extreme Gradient Boosting (XGBoost), Gradient Boosting Machine (GBM), and Random Forest (RF) models were trained with hyperparameter tuning, 10-fold cross-validation, and a 70/30 train–test split. XGBoost and GBM showed stronger and faster-converging performance during cross-validation and learning-curve analysis than RF. However, RF achieved lower error on the independent test dataset, while the Wilcoxon signed-rank test indicated no statistically significant differences among the models. Bulk Density was the most influential factor shaping SLBC, with higher values in Sylhet Sadar and lower values in Tahirpur and Bishwambarpur. With larger data coverage, incorporation of seasonal data, machinery interaction, and hybrid models, the ML–RS framework could become more scalable and provide more rigorous SLBC predictions.</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

Spatial Mapping of Cropland Soil Load Bearing Capacity in Northeastern Bangladesh: A Multi-Feature-based Prediction using Machine Learning-Remote Sensing Fusion

  • 期刊论文

  • 发表日期：

    2026年5月13日

查看全文
  • Sabyasachi Niloy,
  • Mohd Saifur Rahman,
  • Saif Izlal,
  • Fahim Mahafuz Ruhad,
  • Aniruddha Chanda,
  • Ajoy Kumar Saha,
  • Md. Altaf Hossain

摘要

Soil load-bearing capacity (SLBC) is a critical determinant of sustainable soil management and the efficient operation of agricultural machinery, yet continuous field-based assessment is constrained by high costs and labor demands. This study presents a novel, spatially explicit remote sensing–machine learning (RS–ML) framework for mapping SLBC across croplands in northeastern Bangladesh, particularly in haor regions. Field observations from 127 agricultural sites were collected using a Dynamic Cone Penetrometer (DCP), with penetration-per-blow values converted to DCP indices and subsequently transformed into California Bearing Ratio (CBR) and SLBC. A predictor set, including soil properties, topographic derivatives, and spectral indices, was preprocessed in Google Earth Engine and harmonized into raster layers at ~ 250 m resolution. Extreme Gradient Boosting (XGBoost), Gradient Boosting Machine (GBM), and Random Forest (RF) models were trained using hyperparameter tuning, 10-fold cross-validation, and a 70/30 train–test split. During 10-fold cross-validation, XGBoost (MSE 3.03–43.57, RMSE 1.74–6.60, MAE 1.49–3.92, R² 0.567–0.955) and GBM (MSE 2.81–46.60, RMSE 1.68–6.83, MAE 1.32–3.75, R² 0.531–0.958) showed better predictive performance for SLBC than RF (MSE 6.59–60.91, RMSE 2.57–7.80, MAE 2.14–4.44,R² 0.230–0.901). However, RF achieved lower error on the test dataset, while the Wilcoxon signed-rank test showed no statistically significant differences among the models. Bulk Density was identified as the most influential factor shaping spatial variability in soil load-bearing capacity (SLBC). Higher SLBC in Sylhet Sadar and lower values in Tahirpur and Bishwambarpur were observed. Limitations include relatively homogeneous soil conditions, limited field coverage, lack of seasonal variability, and restricted model extrapolation beyond training data. Despite these constraints, the proposed ML–RS framework provides spatially explicit SLBC predictions that can inform data-driven decision-making and support mechanized farming in wetland-prone croplands of northeastern Bangladesh and similar agroecosystems.

Graphical Abstract

This graphical abstract presents a spatially explicit, multi-feature-based machine learning–remote sensing (ML–RS) framework for mapping Soil Load Bearing Capacity (SLBC) across persistent croplands in northeastern Bangladesh. Field-based SLBC measurements were collected from 127 locations using a Dynamic Cone Penetrometer (DCP). Penetration-per-blow values were converted into DCP indices and subsequently transformed into California Bearing Ratio (CBR) and SLBC using established empirical relationships, which served as the model target. A comprehensive suite of multi-source remote sensing predictors was assembled, including soil intrinsic properties (bulk density, sand, silt, and clay), hydro-thermal variables (soil moisture and soil temperature), topographic derivatives (slope, aspect, curvature, and topographic wetness index), and vegetation- and water-related spectral indices (NDVI and NDWI). All predictor layers were harmonized to ~ 250 m spatial resolution across croplands using integrated workflows in Google Earth Engine, ArcGIS, and Python. Extreme Gradient Boosting (XGBoost), Gradient Boosting Machine (GBM), and Random Forest (RF) models were trained with hyperparameter tuning, 10-fold cross-validation, and a 70/30 train–test split. XGBoost and GBM showed stronger and faster-converging performance during cross-validation and learning-curve analysis than RF. However, RF achieved lower error on the independent test dataset, while the Wilcoxon signed-rank test indicated no statistically significant differences among the models. Bulk Density was the most influential factor shaping SLBC, with higher values in Sylhet Sadar and lower values in Tahirpur and Bishwambarpur. With larger data coverage, incorporation of seasonal data, machinery interaction, and hybrid models, the ML–RS framework could become more scalable and provide more rigorous SLBC predictions.