Statistical Regionalization Models Intercomparisons and Hydrological Impacts Project – StaRMIP

This project works in two successive phases for CTRL and future climate.
Phase 1 (in progress) consists in implementing the statistical downscaling models (SDMs), developing the indicators of quality and simulating the HR fields of precipitation (PR) and temperature (T) in context of CTRL climate. The simulated fields will be assessed (uncertainties, indicators, comparison with RCM, etc.) by comparisons to HR gridded data from the “European Climate Data & Assessment” (ECA&D) project. The simulated fields will then be used as inputs into two daily hydrological models (GR4J and HydroStrahler), to compare observed and simulated water flows over the CTRL period for three Mediterranean catchments. The hydrological models will be preliminary calibrated and validated over long time periods based on the gridded PR and T data from the ECA&D project. A new statistical downscaling model will be developed to fill the main gaps of the state-of-the-art approaches.
Phase 2 of this project will rely on phase 1 to select the most efficient GCMs and statistical downscaling models on CTRL climate. Based on these choices, HR fields of future PR and T will be generated, evaluated in turn, and used as inputs in the selected hydrological models to analyze the hydrological impacts of future climate change.
Project management and coordination (PMC) will include an international scientific committee to supervise the good advancement of the work in a consistent way with the related international projects, to serve as a bridge towards and from these projects. PMC will promote the results of the project (e.g., articles, conferences) and manage the specific StaRMIP web site to provide the statistical simulations and the different SDM algorithms and indicators developed. PMC will also be in charge of the international workshop on regionalization.

The project goes on time. Four SDMs have already been implemented or applied for precipitation over the Euro-CORDEX domain:
One model with analogues; one stochastic model with representation of occurrences through a logistic regression and intensities through a Gamma distribution whose parameters depend linearly on atmospheric predictors; One Generalized Additive Model (GAM); the Cumulative Distribution Function – transform (CDFt).
Each model has been applied so far on ERA-I reanalyses for two distinct periods (summer: 15 April – 14 October, winter: 15 October – 14 April) according to a cross-validation procedure.
The first simulations from those models are currently evaluated through the implementation of different statistical indicators and criteria. For rainfall occurrences: occurrence (or non-occurrence) probabilities, mean persistence, dry or wet spells probabilities, etc. For intensities: mean intensities, percentage of explained variance, ratio of variances, etc.

The first results already allow to distinguish some properties proper to each of the models. For illustration, the percentage of explained variance of the winter precipitation for the 1989-2008 period is given in figure 1 for three of the SDMs (Analogues, GAM and the stochastic model). This figure indicates for example that GAM simulations have an inadequate variability with an overestimation of the variance on the Mediterranean circumference and an underestimation on the northern European half. Other models will then quickly enrich those first results in the coming months.

Moreover, discussions and thoughts have started on the SDM that we want to develop in the context of StaRMIP. For the moment, the priority is given to a non-stationary stochastic spatial model.

The current perspectives are simple regarding the given planning for the coming 42 months:
- Continuation of the SDMs implementation and their evaluation through the developed indicators: first on ERA-I reanalyses, then on CTRL simulations from GCMs involved in the IPCC exercise.
- Sensitivity analyses of the hydrological models with those data as inputs.
- Development of a non-stationary stochastic spatial model.
- Moving to phase 2 for the climate and hydrological simulations.

Those results will be presented to the CORDEX conference (4-7 November, 2013, Brussels, Belgium).
An article is also in progress to detail all the first results through the statistical indicators already implemented.

The main aim of this project is to study the uncertainties of statistical regionalization (or downscaling) of precipitation (PR) and temperature (T) at high resolution (HR, here 0.25°x0.25°) and their hydrological impacts in the Mediterranean region. Indeed, this region has geographical and environmental specificities (e.g., mountains, ocean and sea, high density of population) making necessary but particularly complex the downscaling of precipitation and its HR modeling. Moreover, this region has been identified as a “hot-spot” by IPCC (2007) for future climate changes. Indeed, while the signs of the trends seem defined, strong uncertainties remain in the intensity, the patterns and distributions of those changes. In this context, the StaRMIP project aims at intercomparing the whole set of the main statistical approaches for the regionalization of PR ant T (still too often considered as “black boxes”) in their uses and conceptual differences. Ensembles of HR statistically downscaled PR and T will be generated and the simulated fields will be integrated into hydrological models in order to evaluate their quality in space and time via sensitivity analyses. Based on those evaluations, a new statistical downscaling model will be developed to fill in the most important weaknesses and gaps of the state-of-the-art models. The HR projections for future, plugged into the hydrological models, will also serve to forecast future water availability under the constraint of various climatic projections.

Hence, this project will provide:
(i) Ensembles of control (CTRL, 1989-2010) and future (2021-2050 and 2051-2080) HR statistical simulations of PR and T according to different climate models and scenarios over the Mediterranean region;
(ii) Guidelines to their relevant uses and interpretations;
(iii) Indicators quantifying the quality (in space, time and extreme representation) of the HR simulations;
(iv) An as-generic-as-possible statistical downscaling model improving the state-of-the art models;
(v) Retrospective and prospective hydrological scenarios of water availability on several Mediterranean catchments; ands
(vi) Uncertainty assessment of the climate and hydrological simulations, using well dedicated statistical models.
The proposed intercomparisons and studies will also be extended to a part of the datasets of the dynamical simulations available from the Med-CORDEX project. The results brought by the different studies will provide theoretical and practical “guidelines” for the applications of the statistical regionalization approaches, whose relative strengths, weaknesses, and potential improvements are still not well known while they are more and more applied.

Moreover, the statistical downscaling algorithms, the high-resolution simulations, as well as the evaluation techniques of those, will be freely provided to the scientific community, and a website will be created to promote the various deliverables.