Scholarlink

<p>Accurate agricultural drought prediction remains a major challenge in semi-arid regions due to complex climate interactions and data variability. While machine learning has been increasingly applied in this context, few studies have systematically compared multiple models across varying timescales, regions, and feature selection strategies while incorporating model interpretability techniques. This study addresses that gap by evaluating eight machine learning algorithms; LASSO, k-Nearest Neighbor (kNN), Decision Trees (DT), Random Forest (RF), Gradient Boosting Machines (GBM), Support Vector Machines (SVM), Adaptive Boosting (AdaBoost), and Artificial Neural Networks (ANN), for predicting the Agricultural Standardized Precipitation Index (aSPI) at 3, 6, and 9-month timescales in Algeria’s four primary cereal-producing regions: Oum El Bouaghi, Setif, Sidi Bel Abbes, and Tiaret, over the period 1982–2021. The study employs three distinct feature selection scenarios based on RReliefF rankings to assess the trade-off between model complexity and accuracy. SHapley Additive exPlanations (SHAP) were used to interpret model outputs and identify key drought-driving variables. Results revealed that optimal model performance was highly region- and timescale-specific. While Artificial Neural Networks (ANN) demonstrated strong overall performance, particularly for aSPI9 prediction (R<sup>2</sup> &gt; 0.96), other models like Gradient Boosting, SVM, and Random Forest were optimal for specific regions and forecasting horizons. SHAP analysis confirmed precipitation as the primary drought driver but also revealed model-specific sensitivities to other variables like Sunshine (S) and Relative Humidity (RH) during mispredictions. Notably, models using reduced input features often retained or improved accuracy, emphasizing the value of efficient feature selection. The findings provide a valuable framework for interpretable and regionally calibrated drought forecasting, supporting early warning systems and resource planning in Algeria’s cereal-growing zones.</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 machine learning models for regional agricultural drought prediction in Algeria using SHAP analysis

  • 期刊论文

  • 发表日期：

    2025年11月17日

查看全文
  • Tarek Bouregaa

摘要

Accurate agricultural drought prediction remains a major challenge in semi-arid regions due to complex climate interactions and data variability. While machine learning has been increasingly applied in this context, few studies have systematically compared multiple models across varying timescales, regions, and feature selection strategies while incorporating model interpretability techniques. This study addresses that gap by evaluating eight machine learning algorithms; LASSO, k-Nearest Neighbor (kNN), Decision Trees (DT), Random Forest (RF), Gradient Boosting Machines (GBM), Support Vector Machines (SVM), Adaptive Boosting (AdaBoost), and Artificial Neural Networks (ANN), for predicting the Agricultural Standardized Precipitation Index (aSPI) at 3, 6, and 9-month timescales in Algeria’s four primary cereal-producing regions: Oum El Bouaghi, Setif, Sidi Bel Abbes, and Tiaret, over the period 1982–2021. The study employs three distinct feature selection scenarios based on RReliefF rankings to assess the trade-off between model complexity and accuracy. SHapley Additive exPlanations (SHAP) were used to interpret model outputs and identify key drought-driving variables. Results revealed that optimal model performance was highly region- and timescale-specific. While Artificial Neural Networks (ANN) demonstrated strong overall performance, particularly for aSPI9 prediction (R2 > 0.96), other models like Gradient Boosting, SVM, and Random Forest were optimal for specific regions and forecasting horizons. SHAP analysis confirmed precipitation as the primary drought driver but also revealed model-specific sensitivities to other variables like Sunshine (S) and Relative Humidity (RH) during mispredictions. Notably, models using reduced input features often retained or improved accuracy, emphasizing the value of efficient feature selection. The findings provide a valuable framework for interpretable and regionally calibrated drought forecasting, supporting early warning systems and resource planning in Algeria’s cereal-growing zones.