基于多源数据和Stacking集成学习的气象干旱监测模型

Meteorological Drought Monitoring Model Based on Multi-source Data and Stacking Ensemble Learning

摘要:

干旱作为一种具有时空变异性的复合型自然灾害，其频发性和破坏性对社会经济发展和生态系统稳定构成了严重威胁，精确监测干旱事件具有重要现实意义。本研究以陕西省为研究区域，通过整合植被-地表-气候多维干旱因子建立综合特征变量体系，选取最优气象干旱指数作为目标变量，基于Stacking集成学习与多种机器学习算法构建了陕西省2003—2020年堆叠集成干旱指数（Stacked ensemble drought index，SEDI），同时评估其在气象干旱监测中的适用性。结果表明：气象干旱综合指数（Meteorological drought composite index，MCI）、标准化降水指数（Standardized precipitation index，SPI）与标准化降水蒸散发指数（Standardized precipitation evapotranspiration index，SPEI）在月尺度上变化趋势总体一致，但MCI对干旱事件识别具有更高准确性和灵敏性，选为气象干旱监测模型的目标变量。在3种集成模型与5种单一模型中，基于XGBoost构建的集成模型XGBall在陕西省各区域的监测效果最佳，决定系数R2为0.934~0.945，均方根误差（RMSE）为0.208~0.256。2003—2020年，SEDI与MCI在榆林站、秦都站、石泉站干旱等级一致率分别为87.04%、83.80%、85.65%，2种指数反映的干旱趋势基本一致，且R2均大于0.91，表明SEDI能有效识别不同站点的干旱类型及变化趋势。利用两次典型干旱事件（2005年春季与2015年夏季）进行验证，SEDI在区域尺度干旱监测中具有良好的适用性，其与MCI在空间分布特征上具有较高一致性，不同干旱等级站点比例相似度高，能够较为准确地反映干旱过程的时空演变特征。空间自相关分析表明，陕西省气象干旱呈现显著的空间集聚性，全局莫兰指数为0.69，Z得分为3.58，P<0.001。其中，高-高集聚区主要分布在关中西南部和陕南地区，在这些区域干旱事件的发生频率及强度相对较低。低-低集聚区主要分布在关中东北部和陕北地区，在这些区域干旱事件的发生频率及强度相对较高。研究结果可为生态环境评估与保护、干旱状态监测及预警提供科学指导。

Abstract:

As a complex natural disaster exhibiting marked spatiotemporal heterogeneity, drought threatens socio-economic systems and ecosystem resilience through its frequent occurrence and cumulative destructive impacts. Therefore, accurate monitoring of drought events is of great practical significance. Focusing on Shaanxi Province as the research area and a comprehensive feature variable system was established by integrating vegetation, surface, and climate multi-dimensional drought factors. Using the optimal meteorological drought index as the target variable, the stacked ensemble drought index (SEDI) for Shaanxi Province during 2003—2020 was constructed based on Stacking ensemble learning and multiple machine learning algorithms, and its applicability in meteorological drought monitoring was evaluated. The results demonstrated that the monthly-scale variation trends of the meteorological drought composite index (MCI), standardized precipitation index (SPI), and standardized precipitation evapotranspiration index (SPEI) were generally consistent. However, MCI exhibited high accuracy and sensitivity in identifying drought events, thus it was selected as the target variable for the meteorological drought monitoring model. Among the three ensemble models and five single models, the ensemble model XGB-all, constructed based on XGBoost, demonstrated the best monitoring performance across different regions of Shaanxi Province, with coefficient of determination (R2) ranging from 0.934 to 0.945 and root mean square error (RMSE) ranging from 0.208 to 0.256. During 2003—2020, SEDI and MCI showed drought level matching rates of 87.04%, 83.80%, and 85.65% at Yulin, Qindu, and Shiquan stations respectively, with highly consistent drought trends and simulated R2 values all exceeded 0.91, indicating SEDI’s effectiveness in identifying drought types and variation trends across different stations. Validation through two drought events (spring 2005 and summer 2015) confirmed SEDI’s strong applicability for regional-scale drought monitoring, exhibiting high consistency with MCI in spatial distribution characteristics and similarity in proportions of different drought severity levels, effectively reflecting spatiotemporal evolution patterns of drought processes. Spatial autocorrelation analysis demonstrated significant spatial clustering of meteorological drought in Shaanxi Province, with a global Moran’s I of 0.69 (Z-score=3.58, P<0.001). High-high clusters were predominantly distributed in the southwestern Guanzhong Plain and southern Shaanxi regions, corresponding to areas with relatively lower drought frequency and intensity. Conversely, low-low clusters were concentrated in the northeastern Guanzhong Plain and northern Shaanxi regions, which exhibited high drought occurrence rates and severity. This finding can provide scientific guidance for ecological environment assessment, drought monitoring and early warning systems.