Optimising predictor domains for spatially coherent precipitation downscaling

Abstract

Statistical downscaling is widely used to overcome the scale gap between predictors from Numerical Weather Prediction (NWP) models or General Circulation Models (GCMs) and predictands like local precipitation, required for example for medium-term operational forecasts or climate change impact studies. The predictors are considered over a given spatial domain which is rarely optimised with respect to the target predictand location. In this study the geopotential predictor domains used by an analogue downscaling method are optimised for 608 target zones covering France. An extended version of the growing rectangular domain algorithm provides an ensemble of five near-optimum domains for each target zone. All five near-optimum domains are consistently equally skillful based on the Continuous Rank Probability Score. Relevance maps calculated for selected target zones first reveal high skill geopotential regions with specific shapes for locations in south-eastern France compared to the rest of the country. In all cases, the optimised domains tend to include the most relevant area on the relevance maps. The domain centers of the optimised domains are mainly distributed following the geographical location of the target location, but there are apparent differences between the windward and the lee side of mountain ridges. Moreover, domains for target zones located in south-eastern France are centered more east and south than the ones for target locations on the same longitude. The size of the optimised domains tends to be larger in the southeastern part of the country, while domains with a very small meridional extent can be found in a east-west band around 47° N. Sensitivity tests on the archive length for the analogue method show a general robustness except for zones with high interannual variability like in the Cévennes area. Moreover, results appear to be rather unsensitive to the starting point of the optimisation algorithm except for zones located in the transition area north of the zones having optimized domains with a small meridional extent. This study paves the way for defining regions with homogeneous geopotential predictor domains for precipitation downscaling over France

Linking Large‐Scale Circulation Descriptors to Precipitation Variability in the Northern French Alps

“…Seasons featuring a large occurrence of Atlantic circulations are associated with large precipitation accumulation in the Northern French Alps. This result appears to be consistent since the precipitation in the Northern French Alps is mostly related to westerlies bringing moist air from the Atlantic (Garavaglia et al, 2010;Horton et al, 2012;Radanovics et al, 2013;Sodemann & Zubler, 2010). Large seasonal pressure differences over Western Europe (large MPDs) are associated with large seasonal precipitation in the Northern French Alps.…”

Characterizing large‐scale circulations driving extreme precipitation in the Northern French Alps

“…The method was applied for the first time by Duband (1981) to produce probabilistic forecast of daily precipitation over South‐Eastern France. Since then the method has been widely applied and has been subject to various improvements (Obled et al ., 2002; Matulla et al ., 2008; Marty et al ., 2012; Radanovics et al ., 2013; Daoud et al ., 2016; Raynaud et al ., 2017; Chardon et al ., 2018). Unlike weather pattern classifications, the analog method manages to link LSC to regional precipitation in a continuous way, since the link between specific LSCs and the associated precipitation is preserved.…”

Evaluation of different large‐scale predictor‐based statistical downscaling models in simulating zone‐wise monsoon precipitation over India

“…Therefore, too large a domain size can add unnecessary noise and create spurious results (Fealy and Sweeney, ). In addition, some predictors like temperature or moisture variables have their main influence close to the target location (Ben Daoud, ; Radanovics et al ., ). Therefore, small predictor domain close to the target location may be more suitable.…”