Scholarlink

<p>Droughts exert a profound influence on water availability, agriculture, and ecosystems worldwide, and Dobrogea, one of Romania’s driest regions, has been repeatedly affected by severe droughts over recent decades. Given the increasing frequency and intensity of such events under climate change, accurate drought prediction is vital for managing agriculture and water resources in this vulnerable region. This study evaluates the performance of statistical (SARIMA) and machine learning models (Rthe seasonal Mann-Kendallandom Forest, Gradient Boosting, Support Vector Regression) in forecasting the Standardized Precipitation Evapotranspiration Index (SPEI) at multiple time scales (1, 3, 6, 12, and 24 months). The models were developed using in situ data from six meteorological stations and satellite-based datasets covering 1967–2021, providing both local and regional perspectives on drought variability. Lagged predictors (1–12 months) were incorporated to capture temporal memory effects and enhance forecast accuracy. Residual diagnostics confirmed the statistical adequacy and proper specification of the selected models. Results reveal that Dobrogea has experienced widespread, recurrent moderate to severe drought conditions, with extreme events becoming more frequent since 2001. In terms of SPEI prediction, model performance improved with increasing timescale. Among the predictive approaches, SVR and SARIMA achieved the best overall results, showing higher NSE (~ 0.95) and lower RMSE and MAE values compared to the other models. Overall, the findings highlight Dobrogea’s persistent drought exposure and demonstrate the potential of machine learning —particularly SVR—to enhance drought forecasting precision, supporting sustainable agricultural practices and adaptive management under a changing climate.</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

Drought forecasting in Dobrogea, Romania using SPEI-based machine learning models

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年5月2日

查看全文
  • Cristina Serban,
  • Anata Flavia Ionescu,
  • Carmen Maftei

摘要

Droughts exert a profound influence on water availability, agriculture, and ecosystems worldwide, and Dobrogea, one of Romania’s driest regions, has been repeatedly affected by severe droughts over recent decades. Given the increasing frequency and intensity of such events under climate change, accurate drought prediction is vital for managing agriculture and water resources in this vulnerable region. This study evaluates the performance of statistical (SARIMA) and machine learning models (Rthe seasonal Mann-Kendallandom Forest, Gradient Boosting, Support Vector Regression) in forecasting the Standardized Precipitation Evapotranspiration Index (SPEI) at multiple time scales (1, 3, 6, 12, and 24 months). The models were developed using in situ data from six meteorological stations and satellite-based datasets covering 1967–2021, providing both local and regional perspectives on drought variability. Lagged predictors (1–12 months) were incorporated to capture temporal memory effects and enhance forecast accuracy. Residual diagnostics confirmed the statistical adequacy and proper specification of the selected models. Results reveal that Dobrogea has experienced widespread, recurrent moderate to severe drought conditions, with extreme events becoming more frequent since 2001. In terms of SPEI prediction, model performance improved with increasing timescale. Among the predictive approaches, SVR and SARIMA achieved the best overall results, showing higher NSE (~ 0.95) and lower RMSE and MAE values compared to the other models. Overall, the findings highlight Dobrogea’s persistent drought exposure and demonstrate the potential of machine learning —particularly SVR—to enhance drought forecasting precision, supporting sustainable agricultural practices and adaptive management under a changing climate.