Scholarlink

<p>In arid and semi-arid countries, accurate agricultural drought characterization is crucial for climate-resilient planning and efficient early warning systems. Therefore, by integrating remote sensing data, CMIP6 projections, and machine learning (ML) models, the study assesses the spatio-temporal dynamics, driving factors, and future evolution of agricultural drought in the Borena Zone, southern Ethiopia. The Vegetation Health Index (VHI) was primarily targeted using several drought-related indicators, such as the Normalized Difference Vegetation Index (NDVI), Modified Normalized Difference Water Index (MNDWI), Land surface temperature (LST), and Temperature Condition Index (TCI) to detect the drought. Additionally, Shapley Additive Explanations (SHAP) and ML models were utilized to differentiate drought dynamics and performance for two seasons: March-May (MAM) and August-November (ASON). The findings show that the observed distribution of rainfall during both seasons spans from less than 20&#xa0;mm month⁻¹ in the southern lowlands throughout a large portion of the southern basin, and that LST variation increases to 35.1&#xa0;°C during MAM and rises to 38.1&#xa0;°C during ASON, making it greatly susceptible to evapotranspiration. Consequently, the southern zone’s root-zone soil moisture decreases from 0.30 to 0.35&#xa0;m³ m⁻³ in MAM to 0.25–0.30&#xa0;m³ m⁻³ in ASON, along with notable decreases in NDVI and MNDWI. The ML models show strong drought prediction performance when targeting VHI; the Receiver Operating Characteristic Area Under the Curve (ROC–AUC) and Cohen’s kappa of XGBoost achieve 0.997, and 0.72 during MAM, and Random Forest achieves approximately 0.988 and 0.78 during ASON, respectively. TCI and NDVI are the most important predictors of drought severity according to SHAP-based feature attribution. Integrating future drought forecasts from a bias-corrected CMIP6 model (CNRM-CM6-1) under SSP2-4.5 and SSP5-8 scenarios indicates a significant expansion and intensification of drought. Mild drought is projected to affect up to 36.2% of the zone under SSP2-4.5, whereas under SSP5-8.5, moderate, severe, and extreme drought are projected to increase by 18.87%, 5.03%, and 3%, respectively, by the late 21st century. These robust results highlight escalating risks to rain-fed agriculture, rangeland productivity, and water resources, underscoring the urgency of machine-learning–based drought early-warning systems and targeted adaptation strategies in the Borena Zone.</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

AI-driven projection of seasonal agricultural drought using CMIP6 and remote sensing data in Borena Zone, Southern Ethiopia

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年6月24日

查看全文
  • Mikhael G. Alemu,
  • Abdulkerim Bedewi Serur,
  • Fraol Abebe Wudineh,
  • Bahar Adem Beker

摘要

In arid and semi-arid countries, accurate agricultural drought characterization is crucial for climate-resilient planning and efficient early warning systems. Therefore, by integrating remote sensing data, CMIP6 projections, and machine learning (ML) models, the study assesses the spatio-temporal dynamics, driving factors, and future evolution of agricultural drought in the Borena Zone, southern Ethiopia. The Vegetation Health Index (VHI) was primarily targeted using several drought-related indicators, such as the Normalized Difference Vegetation Index (NDVI), Modified Normalized Difference Water Index (MNDWI), Land surface temperature (LST), and Temperature Condition Index (TCI) to detect the drought. Additionally, Shapley Additive Explanations (SHAP) and ML models were utilized to differentiate drought dynamics and performance for two seasons: March-May (MAM) and August-November (ASON). The findings show that the observed distribution of rainfall during both seasons spans from less than 20 mm month⁻¹ in the southern lowlands throughout a large portion of the southern basin, and that LST variation increases to 35.1 °C during MAM and rises to 38.1 °C during ASON, making it greatly susceptible to evapotranspiration. Consequently, the southern zone’s root-zone soil moisture decreases from 0.30 to 0.35 m³ m⁻³ in MAM to 0.25–0.30 m³ m⁻³ in ASON, along with notable decreases in NDVI and MNDWI. The ML models show strong drought prediction performance when targeting VHI; the Receiver Operating Characteristic Area Under the Curve (ROC–AUC) and Cohen’s kappa of XGBoost achieve 0.997, and 0.72 during MAM, and Random Forest achieves approximately 0.988 and 0.78 during ASON, respectively. TCI and NDVI are the most important predictors of drought severity according to SHAP-based feature attribution. Integrating future drought forecasts from a bias-corrected CMIP6 model (CNRM-CM6-1) under SSP2-4.5 and SSP5-8 scenarios indicates a significant expansion and intensification of drought. Mild drought is projected to affect up to 36.2% of the zone under SSP2-4.5, whereas under SSP5-8.5, moderate, severe, and extreme drought are projected to increase by 18.87%, 5.03%, and 3%, respectively, by the late 21st century. These robust results highlight escalating risks to rain-fed agriculture, rangeland productivity, and water resources, underscoring the urgency of machine-learning–based drought early-warning systems and targeted adaptation strategies in the Borena Zone.