Drought impact links to meteorological drought indicators and predictability in Spain

Abstract

Drought affects many regions worldwide, and future climate projections imply that drought severity and frequency will increase. Hence, the impacts of drought on the environment and society will also increase considerably. Monitoring and early warning systems for drought rely on several indicators; however, assessments of how these indicators are linked to impacts are still lacking. Here, we explore the links between different drought indicators and drought impacts within six sub-regions in Spain. We used impact data from the European Drought Impact Report Inventory database and provide a new case study to evaluate these links. We provide evidence that a region with a small sample size of impact data can still provide useful insights regarding indicator–impact links. As meteorological drought indicators, we use the Standardised Precipitation Index and the Standardised Precipitation Evapotranspiration Index; as agricultural and hydrological drought indicators, we use a Standardised Soil Water Content Index and a Standardised Streamflow Index and a Standardised Reservoir Storage Index. We also explore the links between drought impacts and teleconnection patterns and surface temperature by conducting a correlation analysis, and then we test the predictability of drought impacts using a random forest model. Our results show that meteorological indices are best linked to impact occurrences overall and at long timescales between 15 and 33 months. However, we also find robust links for agricultural and hydrological drought indices, depending on the sub-region. The Arctic Oscillation, Western Mediterranean Oscillation, and the North Atlantic Oscillation at long accumulation periods (15 to 48 months) are top predictors of impacts in the northwestern and northeastern regions, the community of Madrid, and the southern regions of Spain, respectively. We also find links between temperature and drought impacts. The random forest model produces skilful models for most sub-regions. When assessed using a cross-validation analysis, the models in all regions show precision, recall, or R2 values higher than 0.97, 0.62, and 0.68, respectively. Thus, our random forest models are skilful in predicting drought impacts and could potentially be used as part of an early warning system.