Spatial control of rain gauge precipitations using radar data (Contribution to WP1) F Mounier, P Lassègues, A-L Gibelin, J-P Céron, J-M Veysseire Meteo-France.

Spatial control of rain gauge precipitations using radar data

(Contribution to WP1)

F Mounier, P Lassègues, A-L Gibelin, J-P Céron, J-M Veysseire

Meteo-France DCLIM / CNRM-GAME

Main topic

MAIN PROBLEMS of rain gauges: density, quality, instrument types

•Real-time Meteo-France (~1500-1800)

•volunteers (~2800)

Construct a 2007-2010 reference estimate of spatialized precipitation for rain gauges control and validation at fine scale.

PROPOSED SOLUTION

Use radar network in the spatialization process of the

precipitation estimate

Radar data need Radar data need also to be qualifiedalso to be qualified

Diagnostic of feasibility

Daily data over 2007-2010

• Average of station based correlation (rain gauge / radar) data over France in average above 0.8

• The Tschuprow coef. per quantile classes of rainfall intensity always above 0.35 that implies a strong link between rain gauge and radar estimate data for the selected stations

Using radar data to control rain gauge precipitations is relevant to construct a frame of reference to better spatialize precipitations.

First half to calibrate radar data

Methodology overview

Controlled rainfall observation processControlled rainfall observation process

RAIN GAUGE Precipitations

•Real-time (~1500-1800)

•volunteers (~2800)

divided into two roughly equal lots by carrying out a totally random draw

Production of an independent estimate of rainfall from rain gauges of First & Second halves (except the one controlled) using calibrated radar data via spatialisation method Rainfall Estimates

Second half to be controlled

Observations

4 spatialization methods / 2 used

TPS: Thin Plate Spline in a 3D space

use a smoothing coef. adjusted to minimize the RMSE and the radar data as a third dimension to estimate rain gauge value

KED: Kriging of rain gauge with radar oriented external drift

It is the radar data that define the trend part of the model to guide the estimation of the primary variable (rainfall) at the rain gauge.

Have been also explored but not retained:Have been also explored but not retained: Neural network Optimal interpolation

Rain data filtering control

Period of study: 2007 to 2010

Only daily results are presented

Rain data should be above 0.6 mm

Only radar or rain gauge data with a good quality parameter are

taken (84) into account

Only sample with a minimum of 100 radar/rain gauge couple of data per station are employed.

Results 2007-2010

Not differences easily readable!!

KED TPS

Cross-method (bootstrap + student test) comparison

Estimation 1

Estimation 2

Results 2007-2010: cross-method comparisont-values mapping

Mapping of the student t-value

(data within +/-1.96 are in white)

&

Kernal density plot to view the distribution of the three scores (data within +/-1.96 are set to zero)

TPS better

KED better

-60 0 60

TPS better KED better

RMSE

CORR

BIAS

Results 2007-2010 by season I

Winter Summer

RMSE

CORR

BIAS RMSE

CORR

BIAS

TPS betterTPS better

Krig betterKrig better

-60 0 60-60 0 60

Results 2007-2010 by season II

Autumn Spring

RMSE

CORR

BIAS

RMSE

CORR

BIAS

TPS betterTPS betterKrig better Krig better

-60 0 60 -60 0 60

Possible explanations for the results

orography (not significant) Rain intensity (not significant) Radar type C or S (not significant) Rain type convective/non-convective (significant)

Two tools to classify rain type: The instantaneous Cape (Convective Available Potential Energy) from

Aladin model: An air parcel need sufficient potential energy for convection, above 20j/Kg of Cape value the rain gauge is associated with a convective situation.

Antilope convective index: Generated from the Antilope radar product of Meteo-France, convective index is based on radar reflectivity gradients in the immediate vicinity of the pixel associated with controlled rain gauge; Above a 0 value the rain gauge is associated with a convective situation.

Classification following Rain type convective/non-convective

Non-convective situations Convective situations

Aladin cape values

Antilope convective index

TPS better TPS betterKrig better Krig better

RMSE

CORR

BIAS

RMSE

CORR

BIAS

control of daily precipitation using radar data - I

For each rain gauge:For each rain gauge:

rain gauge observation O

Estimate of rainfall E

RMSE and Bias

standard deviation

22 biasRMSESd

If |O – E| < 3Sd|O – E| < 3Sd

Observation plausible

If |O – E| |O – E| 3Sd 3Sd

Doubtful observation

Map of the % of doubtful observations

The largest circles are for the 10% of stations that have the worst performance

control of daily precipitation using radar data - IITot rainfall

observations testeddoubtful using

KEDdoubtful using

TPSdoubtful common to

both methods

6 356 775 (0,076% of tot)

4866(0,098% of tot)

6242 3 479

number of rainfall values

Number of rainfall observations available during the control process, with the number of doubtful ones following the method employed to obtain the

estimates.

KDE controlTPS control

control of daily precipitation using radar data - III

KDE control TPS control

Conclusions & Perspectives I

•TPS and kriging perform well to produce estimate of rain gauge data using radar data.

•TPS tends to perform better for non-convective situations while Kriging better for convective ones.

Type Case 1 Case 2of situation Convective Non-convective

of season Summer Winter

Spatialization method to be favored during control

Kriging TPS

The operational development of this WP1 contribution should be taken into account in the “best practice selection instructions”.

Conclusions & Perspectives II

Further analysis of the control method results & proceed to a human expertise of the controlled data.

Evaluate the possibility to apply this control method outside of France following the establishment of a critical study of network density of rain gauges and treatments related to radar data (collaboration possible).

Construction of a control method for situations of rain / no-rain and establishment of special treatment for the snow situations.

Further work on hourly data who faces various problems such as a sparse network of hourly rain gauges data (automatic station only) and also rainy data rarest and with a greater variability.

Continue collaboration with MeteoSwiss on the intercomparison of spatialization methods on specific areas (Alps…)

Acknowledgements

The research leading to these results has received funding from the European Union, Seventh Framework Programme (FP/2007-2013) under grant agreement no 242093.

Methodology overview

control of daily precipitation using radar data

rainfall valuerainfall

observations tested

doubtful using KED

doubtful using TPS

O. Krig. without radar

local mean

Equal to 0 3 230 695(0,071% of =0)

2312(0,074% of =0)

2394(0,07% of =0)

2282(0,06% of =0)

2057

Greater than 0 3 126 080(0,081% of >0)

2554(0,123% of >0)

3848(0,19% of >0)

6058(0,24% of >0)

7611

Total 6 356 775 (0,076% of tot)

4866(0,098% of tot)

6242(0,13% of tot)

8340(0,15% of tot)

9668

Number of rainfall observations available during the control process, with the number of doubtful ones following the method employed to obtain the

estimates.