Scholarlink

<p>Wetlands, as vital natural ecosystems, play a key role in environmental sustainability. However, in recent decades, they have come under pressure from human activities and climate change. The depletion of water and vegetation in the Hamoun Wetland, Iran-Afghanistan, has led to the occurrence of dust storms, which cause damage to agriculture, human settlements, air quality, and public health. This study investigates land use/land cover (LULC) changes in the Hamoun Wetland (2000–2020) using satellite-derived indices (NDVI, NDWI, LST, PDSI) and predicts annual changes in class areas using four multi-output machine learning models: XGBoost, Random Forest, Gradient Boosting (GBM), and K-Nearest Neighbors (KNN). Model evaluation employed Leave-One-Year-Out Cross-Validation (LOOCV) based on 21 annual observations. Results showed that barren land expanded significantly, while reed beds, water bodies, and agricultural lands declined. GBM achieved the highest accuracy (R² = 0.957, MAE = 86.4&#xa0;km²), followed by KNN (R² = 0.948, MAE = 93.4&#xa0;km²). Feature importance revealed that structural variables (areas of Agriculture, Barren Land, and Reed Beds) ranked highest, while environmental indices contributed approximately 14.4% in the GBM model. The median prediction error for GBM was ~ 50&#xa0;km², with a standard deviation of 69.5&#xa0;km² across LOOCV folds. Scenario simulations showed that agricultural expansion reduces water bodies (-108.2&#xa0;km²), severe drought affects water (-79.2&#xa0;km²) and reed beds (-17.3&#xa0;km²), and wetland restoration increases water (+ 94.7&#xa0;km²) but also barren land (+ 156.5&#xa0;km²). These findings demonstrate that machine learning models, particularly GBM with rigorous cross-validation, are effective tools for basin-scale LULC prediction and wetland 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

Retrospective simulation of land use area dynamics in the Hamoun Wetland using annual machine learning and management scenarios

  • 期刊论文

  • 发表日期：

    2026年7月3日

查看全文
  • Mozhgan Yarahmadi,
  • Philip David Hughes,
  • Leila Ghasemi,
  • Benli Liu

摘要

Wetlands, as vital natural ecosystems, play a key role in environmental sustainability. However, in recent decades, they have come under pressure from human activities and climate change. The depletion of water and vegetation in the Hamoun Wetland, Iran-Afghanistan, has led to the occurrence of dust storms, which cause damage to agriculture, human settlements, air quality, and public health. This study investigates land use/land cover (LULC) changes in the Hamoun Wetland (2000–2020) using satellite-derived indices (NDVI, NDWI, LST, PDSI) and predicts annual changes in class areas using four multi-output machine learning models: XGBoost, Random Forest, Gradient Boosting (GBM), and K-Nearest Neighbors (KNN). Model evaluation employed Leave-One-Year-Out Cross-Validation (LOOCV) based on 21 annual observations. Results showed that barren land expanded significantly, while reed beds, water bodies, and agricultural lands declined. GBM achieved the highest accuracy (R² = 0.957, MAE = 86.4 km²), followed by KNN (R² = 0.948, MAE = 93.4 km²). Feature importance revealed that structural variables (areas of Agriculture, Barren Land, and Reed Beds) ranked highest, while environmental indices contributed approximately 14.4% in the GBM model. The median prediction error for GBM was ~ 50 km², with a standard deviation of 69.5 km² across LOOCV folds. Scenario simulations showed that agricultural expansion reduces water bodies (-108.2 km²), severe drought affects water (-79.2 km²) and reed beds (-17.3 km²), and wetland restoration increases water (+ 94.7 km²) but also barren land (+ 156.5 km²). These findings demonstrate that machine learning models, particularly GBM with rigorous cross-validation, are effective tools for basin-scale LULC prediction and wetland management.