Spatially coherent probabilistic precipitation downscaling with meteorological analogues

Studying past and present day precipitation and its link to large scale circulation increases our understanding of precipitation characteristics and helps to anticipate their future behaviour. Downscaling techniques are being developed to bridge the gap between large-scale climate information from global reanalyses or GCM global projections and local meteorological information relevant for hydrology. The stepwise analogue downscaling method for hydrology (SANDHY) is extended to the whole mainland of France by optimising the geopotential predictor domains for 608 zones covering France using a multiple growing rectangular domain algorithm that allows to take equifinality into account. A high diversity of predictor domains has been found. To increase the spatial coherence three ways are explored to reduce the parameter space: assessing the skill for predictor domains found for other zones, form groups of zones using cluster algorithms and using a less skewed predictand variable during optimisation. Using information from neighbouring zones allows to counterbalance in part limitations of the optimisation algorithm. A feature based spatial verification method (SAL) is adapted for probabilistic precipitation simulation as provided by SANDHY. Skill scores derived from the probabilistic SAL are used to assess different strategies for spatially coherent precipitation downscaling at catchment scale. Locally optimised predictor domains lead to a better localisation of precipitation in the catchment and higher local skill while uniform predictor domains for the whole catchment lead to a more realistic spatial structure of the simulated precipitation. Streamflow simulations for the Durance catchment (Southern Alps) are most sensitive to the realistic localisation of precipitation which highlights the interest of locally optimising predictor domains.STAR