Scholarlink

<p>Soil erosion is a major form of land degradation affecting agricultural productivity, ecological balance, and watershed sustainability, especially in monsoon-dominated subtropical regions. This study employs an integrated geospatial and machine learning approach to delineate soil erosion susceptibility zones within a 12,959&#xa0;ha watershed. Sixteen geo-environmental parameters—spanning topographic, hydrological, climatic, vegetative, and geological domains—were utilized as input variables. Three modeling techniques were applied: eXtreme Gradient Boosting (XGBoost), Artificial Neural Network (ANN), and Geographically Weighted Regression (GWR). Importantly, the three models reveal contrasting trade-offs between accuracy, interpretability, and spatial generalization: XGBoost achieved the highest statistical fit within the validation framework (R<sup>2</sup> = 0.94, NSE = 0.94), ANN (R<sup>2</sup> = 0.86) captured hydrological patterns but tended to overpredict, and GWR (R<sup>2</sup> = 0.62) provided spatially varying coefficient insights unavailable from the other approaches. ROC-AUC values for XGBoost and ANN were 0.99, and 0.89 for GWR, within the available hold-out validation framework. Key variables contributing to model outputs included the LS factor, stream power index (SPI), rainfall erosivity, normalized difference vegetation index (NDVI), and bare soil index (BSI). Spatially, the XGBoost model identified 6554.9&#xa0;ha as moderately to highly susceptible to erosion, aligning with observed field indicators such as gully formation.</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

Identification of soil erosion hotspot zones using integrated machine learning and geospatial models

  • 期刊论文

  • 发表日期：

    2026年6月22日

查看全文
  • Deepak Singh,
  • Nisha Singh,
  • Harendra Singh

摘要

Soil erosion is a major form of land degradation affecting agricultural productivity, ecological balance, and watershed sustainability, especially in monsoon-dominated subtropical regions. This study employs an integrated geospatial and machine learning approach to delineate soil erosion susceptibility zones within a 12,959 ha watershed. Sixteen geo-environmental parameters—spanning topographic, hydrological, climatic, vegetative, and geological domains—were utilized as input variables. Three modeling techniques were applied: eXtreme Gradient Boosting (XGBoost), Artificial Neural Network (ANN), and Geographically Weighted Regression (GWR). Importantly, the three models reveal contrasting trade-offs between accuracy, interpretability, and spatial generalization: XGBoost achieved the highest statistical fit within the validation framework (R2 = 0.94, NSE = 0.94), ANN (R2 = 0.86) captured hydrological patterns but tended to overpredict, and GWR (R2 = 0.62) provided spatially varying coefficient insights unavailable from the other approaches. ROC-AUC values for XGBoost and ANN were 0.99, and 0.89 for GWR, within the available hold-out validation framework. Key variables contributing to model outputs included the LS factor, stream power index (SPI), rainfall erosivity, normalized difference vegetation index (NDVI), and bare soil index (BSI). Spatially, the XGBoost model identified 6554.9 ha as moderately to highly susceptible to erosion, aligning with observed field indicators such as gully formation.