N2  - Quantification of the temporally varying streamflow intermittence at continental scales provides an important basis for evaluating biodiversity, ecosystem functions and ecosystem services in rivers as well as water resources for humans. As streamflow intermittence is often more prevalent in small upstream river reaches than in large downstream rivers, quantification needs to be done with a high spatial resolution. Aggregated to five classes (0, 1-2, 3-15, 16-29, 30-31 no-flow days), the number of no-flow days of approximately 1.5 million river reaches in Europe was estimated for each of the 468 months in the period 1981-2019 using a two-step Random Forest modeling approach. The model was developed based on a custom version of the 15 arc-sec HydroSHEDS drainage direction dataset. Data for 18 predictor variables (on hydrology, climate, physiography, geology, and land cover) as well as daily streamflow observed at 1,915 streamflow gauging stations were prepared as input to the RF model. In addition to upstream drainage area and slope, predictors based on time series of streamflow in 15 arc-sec grid cells were found to be most important for the RF model. These time series were generated by downscaling the 0.5 arc-deg runoff of the global hydrological model WaterGAP (downscaled streamflow is also already available for South America). In Europe but not in South America, the performance of downscaled monthly WaterGAP v2.2e streamflow as compared to streamflow observations is, on average, satisfactory also for small drainage basins of less than 10 km2. While 99% and 95% of the observed perennial station-months are predicted correctly for the calibration and validation periods, respectively, the RF approach tends to overestimate intermittence Considering only the intermittent station-months, the frequency of predicting the correct class among the four classes is about 56% and 47% for the calibration and the validation period, respectively. 9% of all reach-months are simulated to be intermittent. The temporal and spatial patterns of simulated intermittence classes are plausible. The simulated intermittence class in each reach-month will be used by the other DRYvER Work Packages to upscale models developed at the Drying River Network scale.
UR  - http://resolver.sub.uni-goettingen.de/purl?gldocs-11858/11122
ER  -