5.5 Evaluation of Radarnet implementation
5.5.1 Case study evaluation
A 24 hour rainfall event beginning at 16:00 UTC on 15th September 2016 was used to analyse the impacts of LDR-based VPR classification on surface rain rates. This convective event included rainfall intensities that would be considered extreme in the UK, from a relatively stationary weather system that caused very high accumulations in some locations (figure 5.12). The position of maximum rainfall intensities with respect to
Figure 5.13: Bias (left) and RMSE (right) of hourly radar accumlations over the 24 hour trial compared to those of colocated gauges, where the gauge accumulation exceeds a certain threshold. Event counts are shown by the dashed grey line. There is a small but noticable reduction in both bias and RMSE for the LDR algorithm trial at all intensities.
Wardon Hill, coupled with a 0oisotherm height of around 3 km, provided ample sampling of the melting layer for the duration of the event.
Reflectivity volumes from Wardon Hill were processed first using the current Radarnet convective diagnosis algorithm (Z1 > 30 dBZ) only, and then with the new algorithm:
using melting layer LDR (in regimes 1 and 2) to identify the presence or absence of bright band. In each case, where the operational or LDR-based criterion identified “no bright band”, a constant VPR (Z(h) =Zs) was assumed.
A pilot investigation using this case study discovered some instances in which the LDR- based criterion caused a reduction in surface rain rates. On inspection of other radar PPIs covering this event, which included high reflectivity values above and lowρhv values
below the melting layer, it was deduced that these results were likely attributable to hail being sampled at the melting level. In this case high reflectivity measurements persisted at high levels above the freezing level, but the LDR in sampling regimes 1 and 2 (section 5.2.2) did not distinguish correctly between melting hail and stratiform bright band, leading to a bright band correction being applied. For this reason it was decided to allow positive identification of non-bright band precipitation from either LDR or Z1 in
regimes 1 and 2, in order to diagnose convection correctly in the specific case where hail is present.
The results of the prototype LDR-based algorithm (including the hail contingency) com- pared to the operational reflectivity-based convective diagnosis scheme are presented in figures 5.13 and 5.14. Figure 5.13 shows the bias and RMSE (appendix B) of hourly radar accumulations from the event. The LDR algorithm reduces negative biases, partic-
Figure 5.14: Left: total radar accumulation differences (trial minus control) with all rain gauge positions marked, to a maximum range of 255 km. Black crosses show the 58 gauges with near- complete timeseries at which the total radar event accumulations differed. Right: scatterplot of the hourly accumulations at these 58 gauges where LDR-based VPR classification produced a change in the radar accumulation.
ularly at high rain rates. This is expected, since the primary impact of LDR should be to reduce the inappropriate correction for bright band which contributes to these large negative biases. The impact of this on RMSE is noticable, with the error on the highest intensity events reduced by almost 1 mm h−1.
It should be remarked at this point that the biases in this case, although large, are not atypical of statistics for the UK network. This is partly due to the use of a fixed Marshall-Palmer (Marshall and Palmer, 1948) reflectivity-rain rate (ZR) relationship to calculate rain rate. The true ZR relation is dependent on rain drop size distribution (DSD, section 1.2.2), and the Marshall-Palmer relation is known to underestimate rain rates in compact ice and convective conditions, which contain a relatively large population of small rain drops (Bringi et al., 2009; Matrosov et al., 2016). For the event of 15th-16th September, independent disdrometer measurements suggest the DSD was more uniform than Marshall-Palmer (Thompson 2016, personal communication). This goes some way to explaining the particularly large negative biases during this event.
Figure 5.14 evaluates the changes in total event rainfall accumulation when LDR is used to inform the VPR correction. The left hand panel shows the change in accumulation (in mm h−1) resulting from LDR-based classification. Differences are confined to the region (approximately 100-200 km in range) in which the radar sampled the melting layer. As expected, LDR-based VPR classification generally results in increased rain rates through avoiding inappropriate correction for bright band. The left hand panel of figure 5.14 also shows the positions of all rain gauges used in the evaluation of QPEs from Wardon Hill. Black crosses indicate the 58 gauges at which radar accumulations differed for this case
study, and which had near-complete timeseries of hourly accumulations throughout the event (at least 22 out of the 24 hours available). Green stars show the locations of all other reporting rain gauges.
Alongside the preservation of high rain intensities from non-bright band events, there is a small reduction in this case in rainfall accumulations at longer ranges. These represent an area over which LDR diagnoses “non-bright band” conditions, but the current Radarnet non-bright band VPR shape does not account for any decrease in reflectivity above the melting layer. This result, coupled with the average shapes of VPRs observed in the high resolution dataset (chapter 4, figure 4.6), indicates a need to reconsider the shape used for VPR correction in non-bright band conditions. The potential to correct more effectively for non-bright band VPRs is explored in chapter 6.
The right hand panel of figure 5.14 plots the hourly radar accumulations that differ between trial and control against the colocated rain gauge accumulation. This is to determine whether the changes to QPE accumulations - both positive and negative - represent an improvement in radar accuracy with respect to “ground truth”. The ma- jority of changes due to LDR, particularly at the higher hourly values, move the radar accumulation closer to that of the colocated gauge. This can be seen both in individual points on the scatterplot and the increase in radar-gauge correlation for the LDR trial.