Scholarlink

<p>The attributes of soils in the Krishna Delta have become increasingly unbalanced due to high agricultural productivity, which may negatively affect their long-term productivity and soil health. Still, the variation in their spatial distribution is still underexplored. This research attempts to estimate six main soil attributes, including organic carbon (OC), electrical conductivity (EC), pH, nitrogen (N), phosphorus (P₂O₅), and potassium (K₂O) using four machine learning algorithms: Random Forest, k-Nearest Neighbour, Artificial Neural Network, and Cubist. A hybrid method of feature selection with the application of Variance Inflation Factor (VIF) and Recursive Feature Elimination (RFE) reduced the total number of 36 environmental covariates to the optimal subset. Among all ML models, the RF model showed superiority in all simulations, especially in OC (R² = 0.43) and K₂O (R² = 0.51) predictions, since it can better characterize non-linear associations between soils and their environment. The residuals proved that the model predicted values with relatively small errors without spatial dependency. Ecological validity was attained when LULC data were used in the models; it was evident that forests had a high amount of organic carbon and nutrients, while crops had acid soil pH (5.7) and depleted potassium content (520.956&#xa0;kg/ha). The built environment had a relatively high nitrogen content (215.111&#xa0;kg/ha). Overall, this study demonstrates that the combination of feature selection, machine learning, and spatial validation is highly efficient for DSM even in places where data availability is an issue.</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 assessment of soil health in the Krishna Delta using machine learning and land use patterns

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年6月17日

查看全文
  • Amol Kumar Garg,
  • V. Siva Jyothi,
  • Sk. Mustak,
  • Prashant K. Srivastava,
  • Jitendra Kumar Pattanaik,
  • Pritam Mitra

摘要

The attributes of soils in the Krishna Delta have become increasingly unbalanced due to high agricultural productivity, which may negatively affect their long-term productivity and soil health. Still, the variation in their spatial distribution is still underexplored. This research attempts to estimate six main soil attributes, including organic carbon (OC), electrical conductivity (EC), pH, nitrogen (N), phosphorus (P₂O₅), and potassium (K₂O) using four machine learning algorithms: Random Forest, k-Nearest Neighbour, Artificial Neural Network, and Cubist. A hybrid method of feature selection with the application of Variance Inflation Factor (VIF) and Recursive Feature Elimination (RFE) reduced the total number of 36 environmental covariates to the optimal subset. Among all ML models, the RF model showed superiority in all simulations, especially in OC (R² = 0.43) and K₂O (R² = 0.51) predictions, since it can better characterize non-linear associations between soils and their environment. The residuals proved that the model predicted values with relatively small errors without spatial dependency. Ecological validity was attained when LULC data were used in the models; it was evident that forests had a high amount of organic carbon and nutrients, while crops had acid soil pH (5.7) and depleted potassium content (520.956 kg/ha). The built environment had a relatively high nitrogen content (215.111 kg/ha). Overall, this study demonstrates that the combination of feature selection, machine learning, and spatial validation is highly efficient for DSM even in places where data availability is an issue.