3.1.2. 6 h Accumulated Precipitation

Further, the 6 h accumulated precipitation forecasts have been also verified in order to assess the impact of the ZTD data assimilation depending on the forecast lead time (Figures 5–10). The statistical measures for the dry season were computed for the period between 0600 UTC and 1800 UTC. The periods from 0000-0600 UTC and 1800-0000 UTC were discarded from the analysis because less precipitation was observed and fewer observation-model pairs were available compared to the 0600–1800 UTC time window. Figures 5 and 6 illustrate that the ZTD assimilation leads to a marked improvement of the precipitation forecasting in the dry season by increasing the probability of detection and the prediction quality. The improvement is more profound during the afternoon hours (12000–18000 UTC), when statistically significant increases at the 95% confidence level are evident for POD and ETS, which reach 27.8% for higher than 10 mm threshold and 21.4% for higher than 5 mm threshold, respectively (Figure 5b). In the same 6 h interval, the ZTD experiment results in higher FBIAS values, which are greater than 1 when rainfall is lower than 20 mm. This finding partially explains the overall overestimation of the observed events frequency found for the below 20 mm 24 h precipitation in the dry period (Figure 3a). FAR is mainly decreased during the ZTD experiment between 0600 and 1800 UTC in the dry period, especially for the higher rainfall threshold (Figure 5). This is also true for MAE, as shown in Figure 6. In particular, statistically significant reductions of 13% (19.9%) at the 95% (90%) confidence level are introduced by the ZTD simulation for precipitation above 20 mm (between 5 and 10 mm) from 1200–1800 (0600–1200) UTC (Figure 6). Figure 6 also illustrates that, when considering the lowest three rainfall thresholds (<5 mm), the WRF model mainly overestimates the observed precipitation, whereas it underestimates the higher than 5 mm observed rainfall during the examined forecast lead times. From 1200-1800 UTC, the model overestimation between 1 and 2 mm of precipitation is higher by 57.4% (statistically significant at the 90% confidence interval) during the ZTD simulation (Figure 6). The forecast errors for the rainfall thresholds up to 10 mm are smaller than ~9 mm for all 6 h intervals, while in contrast, they are greater than 19 mm for the highest precipitation threshold (Figure 6).

**Figure 5.** Qualitative model performance statistics averaged for the dry period events for precipitation accumulations between (**a**) 0600–1200 UTC and (**b**) 1200–1800 UTC under six rainfall thresholds during the CTL and ZTD numerical experiments. Percentages indicate the relative difference of the statistical measures between the conducted experiments (one asterisk shows statistical significance at the 90% confidence interval, while two asterisks show statistical significance at the 95% confidence interval).

**Figure 6.** Quantitative model performance statistics averaged for the dry period events for precipitation accumulations between (**a**) 0600-1200 UTC and (**b**) 1200-1800 UTC under six rainfall thresholds during the CTL and ZTD numerical experiments. Percentages indicate the relative difference of the statistical measures between the conducted experiments (one asterisk shows statistical significance at the 90% confidence interval, while two asterisks show statistical significance at the 95% confidence interval).

**Figure 7.** Qualitative model performance statistics averaged for the wet period events for precipitation accumulations between (**a**) 0000–0600 UTC and (**b**) 0600–1200 UTC under six rainfall thresholds during the CTL and ZTD numerical experiments. Percentages indicate the relative difference of the statistical measures between the conducted experiments (one asterisk shows statistical significance at the 90% confidence interval, while two asterisks show statistical significance at the 95% confidence interval).

**Figure 8.** Qualitative model performance statistics averaged for the wet period events for precipitation accumulations between (**a**) 1200–1800 UTC and (**b**) 1800–0000 UTC under six rainfall thresholds during the CTL and ZTD numerical experiments. Percentages indicate the relative difference of the statistical measures between the conducted experiments (one asterisk shows statistical significance at the 90% confidence interval, while two asterisks show statistical significance at the 95% confidence interval).

**Figure 9.** Quantitative model performance statistics averaged for the wet period events for precipitation accumulations between (**a**) 0000–0600 UTC and (**b**) 0600–1200 UTC under six rainfall thresholds during the CTL and ZTD numerical experiments. Percentages indicate the relative difference of the statistical measures between the conducted experiments (one asterisk shows statistical significance at the 90% confidence interval, while two asterisks show statistical significance at the 95% confidence interval).

**Figure 10.** Quantitative model performance statistics averaged for the wet period events for precipitation accumulations between (**a**) 1200–1800 UTC and (**b**) 1800–0000 UTC under six rainfall thresholds during the CTL and ZTD numerical experiments. Percentages indicate the relative difference of the statistical measures between the conducted experiments (one asterisk shows statistical significance at the 90% confidence interval, while two asterisks show statistical significance at the 95% confidence interval).

During the wet season (Figures 7–10), the positive impact of ZTD assimilation on the 6 h accumulated precipitation, especially when exceeding 20 mm, is clearly shown. More specifically, during the first 6 h of the numerical forecasts, the FAR is decreased by 3.2% (statistically significant at the 95% confidence interval) for the highest rainfall threshold when ZTD data are assimilated in the WRF model (Figure 7a). For the same period and threshold, FBIAS is closer to 1 during the ZTD experiment, whereas no significant divergences between the conducted simulations are found for POD and ETS (Figure 7a). Marked differences are also not evident between 0600 UTC and 1200 UTC for all qualitative statistical measures and precipitation thresholds, except for POD, which is higher by 2.3% (statistically significant at the 95% confidence interval) during the ZTD experiment, when rainfall is above 5 mm (Figure 5b). From 1200-000 UTC, the POD and ETS scores are higher for the ZTD compared to the CTL experiment for the majority of the precipitation thresholds. Especially for the greater rainfall threshold, the improvement provided by the ZTD assimilation in POD and ETS is 10% in the intervals 1200-1800 UTC and 1800-0000 UTC, respectively (Figure 8). Additionally, a statistically significant (95% confidence level) reduction of ETS by 11% is evident during the ZTD simulation for the higher than 10 mm precipitation threshold between 1200 UTC and 1800 UTC (Figure 8a). Concerning the categorical statistical measures, the ZTD assimilation results in the increase of the MAE by ~1 mm (~12%) for the highest rainfall threshold from 0000–0600 UTC (Figure 9a). Statistically significant reductions by 5.4% (90% confidence level) and 8.5% (95% confidence level) provided by the ZTD experiment are found for the precipitation intervals [2,5) and [10,20) from 0000–0600 UTC and 0600–1200 UTC, respectively (Figure 9). MAE is also decreased by 8.4% (statistically significant at the 95% confidence interval) during the ZTD simulation when rainfall is greater than 20 mm in the 6 h forecast from 1200–1800 UTC (Figure 10a). For most of the lower than 20 mm precipitation thresholds, the ZTD simulation between 1200 and 0000 UTC provides improvement. The MB values show that the WRF model overestimates the three lowest rainfall thresholds, with error magnitudes lower than ~4 mm for all 6 h intervals. In contrast, the precipitation amounts that are higher than 5 mm are underestimated by the model and the extent of under-prediction increases with the forecast lead time (Figures 9 and 10).
