Björn Thomas, Gunnar Lischeid, Jörg Steidl, Ottfried Dietrich

• Runoff, especially during summer months, and low flows have decreased in Central and Eastern Europe during the last decades. A detailed knowledge on predictors and dependencies between meteorological forcing, catchment properties and low flow is necessary to optimize regional adaption strategies to sustain minimum runoff. The objective of this study is to identify low flow predictors for 16 small catchments in Northeast Germany and their long-term shifts between 1965 and 2006. Non-linear regression models (support vector machine regression) were calibrated to iteratively select the most powerful low flow predictors regarding annual 30-day minimum flow (AM(30)). The data set consists of standardized precipitation (SPI) and potential evapotranspiration (SpETI) indices on different time scales and lag times. The potential evapotranspiration of the previous 48 and 3 months, as well as the precipitation of the previous 3 months and last year were the most relevant predictors for AM(30). Pearson correlation (r(2)) of the final model is 0.49Runoff, especially during summer months, and low flows have decreased in Central and Eastern Europe during the last decades. A detailed knowledge on predictors and dependencies between meteorological forcing, catchment properties and low flow is necessary to optimize regional adaption strategies to sustain minimum runoff. The objective of this study is to identify low flow predictors for 16 small catchments in Northeast Germany and their long-term shifts between 1965 and 2006. Non-linear regression models (support vector machine regression) were calibrated to iteratively select the most powerful low flow predictors regarding annual 30-day minimum flow (AM(30)). The data set consists of standardized precipitation (SPI) and potential evapotranspiration (SpETI) indices on different time scales and lag times. The potential evapotranspiration of the previous 48 and 3 months, as well as the precipitation of the previous 3 months and last year were the most relevant predictors for AM(30). Pearson correlation (r(2)) of the final model is 0.49 and if for every year the results for all catchments are averaged r(2) increases to 0.80 because extremes are smoothing out. Evapotranspiration was the most important low flow predictor for the study period. However, distinct long-term shifts in the predictive power of variables became apparent. The potential evapotranspiration of the previous 48 months explained most of the variance, but its relevance decreased during the last decades. The importance of precipitation variables increased with time. Model performance was higher at catchments with a more damped discharge behavior. The results indicate changes in the relevant processes or flow paths generating low flows. The identified predictors, temporal patterns and patterns between catchments will support the development of low flow monitoring systems and determine those catchments where adaption measures should aim more at increasing groundwater recharge. (C) 2014 Elsevier B.V. All rights reserved.…