Scholarlink

<p>This study investigates vegetation dynamics in a semiarid mountain ecosystem in southwestern Saudi Arabia from 1990 to 2024 by integrating multisource remote sensing data, bioclimatic variables, and machine learning models. Trends in vegetation greenness (NDVI), water content (NDWI), and land surface temperature (LST) were quantified via Kendall’s τ and Sen’s slope estimators. To capture fine-scale climatic heterogeneity in complex terrains, CHELSA bioclimatic variables were downscaled from ~ 1&#xa0;km to 30&#xa0;m resolution via random forest regression. These predictors were used to model NDVI trend patterns through a comparative framework including a baseline artificial neural network (ANN), metaheuristic-optimized ANN variants (ANN–PSO and ANN–GWO), and ensemble models (random forest and XGBoost). Model performance was assessed via independent validation, error metrics (RMSE, MAE, R²), bootstrap uncertainty analysis, and spatial residual diagnostics. All the models exhibited strong predictive ability, with XGBoost achieving the highest accuracy (RMSE ≈ 0.051; R² ≈ 0.92) and the lowest residual spatial autocorrelation. The spatial results indicate dominant greening across ~ 73–74% of the area, whereas ~ 25% of the area exhibits degradation concentrated in thermal and moisture-stressed zones. These findings highlight the value of integrated spectral-climatic modeling for monitoring vegetation changes in semiarid mountainous environments.</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 neural networks and ensemble models for vegetation trend prediction in a semiarid mountain ecosystem, Saudi Arabia

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年6月19日

查看全文
  • Asma A. Al-Huqail,
  • Zubairul Islam,
  • Chigozie Edson Utazi

摘要

This study investigates vegetation dynamics in a semiarid mountain ecosystem in southwestern Saudi Arabia from 1990 to 2024 by integrating multisource remote sensing data, bioclimatic variables, and machine learning models. Trends in vegetation greenness (NDVI), water content (NDWI), and land surface temperature (LST) were quantified via Kendall’s τ and Sen’s slope estimators. To capture fine-scale climatic heterogeneity in complex terrains, CHELSA bioclimatic variables were downscaled from ~ 1 km to 30 m resolution via random forest regression. These predictors were used to model NDVI trend patterns through a comparative framework including a baseline artificial neural network (ANN), metaheuristic-optimized ANN variants (ANN–PSO and ANN–GWO), and ensemble models (random forest and XGBoost). Model performance was assessed via independent validation, error metrics (RMSE, MAE, R²), bootstrap uncertainty analysis, and spatial residual diagnostics. All the models exhibited strong predictive ability, with XGBoost achieving the highest accuracy (RMSE ≈ 0.051; R² ≈ 0.92) and the lowest residual spatial autocorrelation. The spatial results indicate dominant greening across ~ 73–74% of the area, whereas ~ 25% of the area exhibits degradation concentrated in thermal and moisture-stressed zones. These findings highlight the value of integrated spectral-climatic modeling for monitoring vegetation changes in semiarid mountainous environments.