Comparison of the performance of three types of multiple regression for phenology in Bavaria in a dynamical-statistical model approach


  • Felix Pollinger
  • Katrin Ziegler
  • Heiko Paeth



Bavaria, regression analysis, climate change, statistic methods, regional climate models, phenology


Some of the most obvious consequences of anthropogenic climate change are observed changes in the dates of the occurrence of phenological events. Most prominently, observations from the Northern Hemisphere’s extratropics indicate an earlier occurrence of spring events. Recent climate models include land surface schemes that provide representation of the vegetation. However, they are limited in simulating the plants’ response to climate change. In this study we present results of a dynamical-statistical modeling approach for phenology in southeastern Germany, combining climate change simulations provided by a high resolution, state-of-the-art regional climate model (RCM) with three different types of regression methods: ordinary least squares (OLS), least absolute deviation (LAD) and random forest (RFO). We focus on changes in the day of the year (DOY) of Forsythia suspensa flowering, the earliest phenophase of the growing season in Bavaria. Based on roughly 2600 observations, collected at 94 phenological and 26 meteorological stations between 1952 and 2013, we compare the regressions via a bootstrap, using once 13 and once 4 meteorological variables as predictors. Altogether, we find the regressions with less variables to be more robust, while the regression estimates are nearly identical. Explained variance and RMSE (root mean square error) are 54.8 % and 8.8 days for RFO and 51.2 % and 9.1 days for the other regressions. These trained and cross validated statistical models are used to estimate the effects of future climate change on the DOY by applying them to the RCM simulations. For OLS or LAD, under a low (high) greenhouse gas emission scenario, we find a mean advance of the DOY of 8 (15) days by the end of the 21th century compared to the base period from 1961 to 1990. The spatial pattern of the change resembles the topography, with the strongest trends in the DOY over mountainous regions as a consequence of a simultaneous rise in temperatures and reduction in snow depth. RFO is restricted to the range of the observations and hence the response to the simulated climate is damped, resulting in an advance of DOY of only 5 (8) days and a reduction in variance. There is no apparent spatial pattern identifiable. Altogether, we find OLS and LAD to be more suitable for dynamical-statistical modeling of phenology than RFO.




How to Cite

Pollinger, F., Ziegler, K., & Paeth, H. (2017). Comparison of the performance of three types of multiple regression for phenology in Bavaria in a dynamical-statistical model approach. ERDKUNDE, 71(4), 271–285.