Scholarlink

<p>Soil salinity represents a critical environmental challenge that undermines crop yield and long-term land viability in warm-temperate monsoon climates, underscoring the importance of developing reliable monitoring methods that remain effective across different seasons. This study comparatively evaluated geostatistical and machine learning methods for seasonal soil salinity mapping across the agricultural lands of central Shaanxi Province, China. A total of 580 soil samples were collected during winter and summer seasons, and soil electrical conductivity (EC), major soluble cations, and sodium adsorption ratio (SAR) were analyzed. Kriging was applied to characterize the spatial structure of soil salinity, while Random Forest (RF), Artificial Neural Networks (ANN), and Support Vector Machines (SVM) were implemented using a comprehensive set of salinity-related spectral indices derived from Landsat-8 and Sentinel-2 imagery. Model performance was assessed using the coefficient of determination (R<sup>2</sup>), root mean square error (RMSE), and relative RMSE (%RMSE), and spatial agreement between kriging and machine learning outputs was evaluated using overall accuracy (OA) and the Kappa coefficient. The results revealed pronounced seasonal variability in soil salinity, with higher EC values, greater spatial heterogeneity, and expanded high-salinity zones during the summer season. Kriging exhibited moderate predictive accuracy, with R<sup>2</sup> values of 0.41 in winter and 0.49 in summer, indicating its effectiveness in capturing broad spatial patterns but limited ability to represent fine-scale variability. In contrast, machine learning models substantially outperformed kriging, particularly when Sentinel-2 imagery was employed. The ANN model achieved the highest accuracy under winter conditions (R<sup>2</sup> = 0.76, RMSE = 0.15), whereas the RF model combined with Sentinel-2 data provided the best overall performance in summer (R<sup>2</sup> = 0.89, RMSE = 0.11). Overall, the findings demonstrate that integrating high-resolution satellite data with season-specific machine learning algorithms offers a robust and reliable framework for accurate soil salinity assessment and supports improved agricultural land and water management.</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

Comparative evaluation of geostatistical and machine learning approaches for seasonal mapping of soil salinity in agricultural lands

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月26日

查看全文
  • Song Geng,
  • Zhuolin Li,
  • Chenhao Zhang,
  • Yonggang Yi

摘要

Soil salinity represents a critical environmental challenge that undermines crop yield and long-term land viability in warm-temperate monsoon climates, underscoring the importance of developing reliable monitoring methods that remain effective across different seasons. This study comparatively evaluated geostatistical and machine learning methods for seasonal soil salinity mapping across the agricultural lands of central Shaanxi Province, China. A total of 580 soil samples were collected during winter and summer seasons, and soil electrical conductivity (EC), major soluble cations, and sodium adsorption ratio (SAR) were analyzed. Kriging was applied to characterize the spatial structure of soil salinity, while Random Forest (RF), Artificial Neural Networks (ANN), and Support Vector Machines (SVM) were implemented using a comprehensive set of salinity-related spectral indices derived from Landsat-8 and Sentinel-2 imagery. Model performance was assessed using the coefficient of determination (R2), root mean square error (RMSE), and relative RMSE (%RMSE), and spatial agreement between kriging and machine learning outputs was evaluated using overall accuracy (OA) and the Kappa coefficient. The results revealed pronounced seasonal variability in soil salinity, with higher EC values, greater spatial heterogeneity, and expanded high-salinity zones during the summer season. Kriging exhibited moderate predictive accuracy, with R2 values of 0.41 in winter and 0.49 in summer, indicating its effectiveness in capturing broad spatial patterns but limited ability to represent fine-scale variability. In contrast, machine learning models substantially outperformed kriging, particularly when Sentinel-2 imagery was employed. The ANN model achieved the highest accuracy under winter conditions (R2 = 0.76, RMSE = 0.15), whereas the RF model combined with Sentinel-2 data provided the best overall performance in summer (R2 = 0.89, RMSE = 0.11). Overall, the findings demonstrate that integrating high-resolution satellite data with season-specific machine learning algorithms offers a robust and reliable framework for accurate soil salinity assessment and supports improved agricultural land and water management.