Evaluation of the new French operational radar product for urban hydrology

Even though the weather radar has early appeared as a promising tool for urban hydrology, its operational applications are quite unfrequent, mainly because the characteristics of radar images are often not suited to the requirements of urban hydrology. Nevertheless, the processing methods of radar images have significantly improved during the last years and it is interesting to test their capability. This study evaluates the radar data processed by the new operational French radar product. This evaluation is based on a radar-raingauge data comparison for 4 different time steps (5, 15, 30 and 60 minutes).

Radar data is measured by the C-band weather radar of Trappes (Region Ile de France, 30 km south-west from Paris). The operation protocol consists of volume scans performed every 15 minutes. The three lowest elevation angles (0.4, 0.8 and 1.5°) of this protocol are repeated every 5 minutes. The upper elevation angles are mostly used to identify vertical profiles of reflectivity (VPR). The radar measurements in polar format of 240 m × 0.5° are projected onto a Cartesian grid of 1×1 km2 spatial resolution. The studied radar data is processed by the operational radar rainfall processor developed by Meteo France. This processor includes a series of modules aimed at correcting for ground clutter, partial beam blocking, and VPR effects, as well as the nonsimultaneity of radar measurements. The Marshall-Palmer Z-R relationship is used for the reflectivity-to-rain-rate conversion. The surface rainfall is estimated as a weighted mean of the corrected tilts. Those surface rainfall estimations are finally adjusted. The adjustment method consists in computing every hour an adjustment factor from raingauge and radar data. This factor is then applied to the entire radar image during the next hour. In addition, quality indicators are associated to each radar pixel. An indicator of 0 is bad, an indicator of 100 is excellent. Those indicators take into account the fact that the quality of measurement decreases when the altitude of the radar beam increases and when beam blocking increases. The stratiform, convective or mixed nature of each pixel was identified according to the presence or absence of a bright band and on Steiner's first two criteria. Raingauge data is derived from 87 raingauges located between 0 and 135 km from the weather radar of Trappes. A total of 50 days of varied data recorded in 2007 and 2008 were used to compare radar to raingauge data.

For the 4 time steps considered (5, 15, 30 and 60 minutes), results were stratified according to the distance to the radar (7 classes: 0-30 km / 30-45 km / 45-60 km / 60-70 km / 70-80 km / 80-100 km and higher than 100 km), to the quality indicators (5 classes: 76-84 / 84-88 / 88-92 / 92-96 and 96-100) and according to the nature of the event (summer/winter and stratiform, convective or mixed). For each study the normalized bias, normalized mean absolute error (MAE), normalized root mean square error (RMSE), determination coefficient (r2) and the Nash criterion have been computed.

Results tendencies in function of distance, quality indicators and the nature of the event are verified independent of the time step considered. Results are significantly better with increasing time step (lower MAE and RMSE and higher r2 and Nash). Only the bias is independent of the time step. Bias appears clearly dependent on distance to the radar. With increasing distance, the weather radar compared to raingauge changes from an overestimation of rainfall to an underestimation of it. The other four statistical scores do not follow a particular tendency in function of the distance. Results are globally better for the 80-100 km class (RMSE=3.56 and Nash=0.39 at Δt=15 min, RMSE=2.06 and Nash=0.6 at Δt=60 min). The study in function of the quality indicators gives a supplementary information: the determination coefficient is the highest for optimal quality indicators (r2=0.32 at Δt=5min and r2=0.7 at Δt=60 min for the 96-100 class). The nature of the event is the most relevant criteria. Results are much better for stratiform pixels with a determination coefficient (r2) reaching 0.72 at the 60 minutes time step (against 0.45 for convective pixels and 0.53 for mixed ones). Convective pixels have the worst results with notably very high RMSE and bias (RMSE=9.21 and bias=0.6 against RMSE=1.22 and bias=0.09 for stratiform pixels at Δt=30 min). This is coherent with the results of the study in function of the season: winter events have better results than summer events. Results seem to be very sensitive to rain variability.

In conclusion, there was surprisingly no degradation of results with increasing distance. This might be partly due to the high bias which can be observed for low distances from the radar. It is also noticeable that results for the time step of 5 minutes tended to be rather bad. In the next step of this research, it is envisioned to quantify the measure noise relative to raingauge measurements.