3.1.1. Daily Precipitation

Figure 3 presents the categorical scores for each 24 h accumulated precipitation threshold, aggregated for the dry and wet season rainfall events for the CTL and ZTD numerical experiments. Overall, the model performs adequately in capturing the occurrence (non-occurrence) of precipitation, as indicated by the POD (FAR), which is higher (lower) than 0.57 (0.38) and 0.67 (0.26) for all thresholds during the dry (Figure 3a) and wet (Figure 3b) period, respectively. The ETS values show a satisfying precipitation forecast quality, especially for the wet season (Figure 3b), when they range from 0.53 to 0.88. During both periods, the FBIAS values demonstrate that the model underestimates the observed frequency of higher than 20 mm daily precipitation, whereas it slightly overestimates the frequency of the observed daily rainfall for the lower than 20 mm thresholds, except those that are greater than 2 mm (5 mm and 10 mm) during the CTL experiment in the wet (dry) season (Figure 3).

Assimilating ZTD observations into the 3D-var/WRF model leads to increased probability of precipitation detection during the dry period (Figure 3a) across all rainfall thresholds. Especially for the highest precipitation threshold (>20 mm), the ZTD assimilation induced relative improvement is 10.6% (statistically significant at the 95% confidence interval; Figure 3a). Concerning FBIAS, the ZTD experiment leads to larger frequency biases compared to the CTL simulation, when precipitation is lower than 20 mm in the dry period, reaching 10.3% relative difference for the above 10 mm rainfall threshold (statistically significant at the 90% confidence interval; Figure 3a). FAR in the dry period is also higher for the same threshold (>10 mm) during the ZTD experiment, whereas a decrease in FAR is evident for greater than 20 mm 24 h precipitation when ZTD assimilation is applied (Figure 3a). No marked differences between the CTL and ZTD experiments are evident during the dry season for ETS, except from the considerable improvement by 11.4% (statistically significant at the 90% confidence interval) provided by the ZTD assimilation for the highest precipitation threshold (Figure 3a). For the same threshold, the ZTD experiment also leads to statistically significant improvements for all statistical measures, except FBIAS during the wet season (Figure 3b). In particular, POD and ETS are increased by 3.5% and 8.1% at the 95% confidence level, while FAR is reduced by 16.9% at the 90% confidence interval (Figure 3b).

**Figure 3.** Qualitative model performance statistics averaged for the (**a**) dry and (**b**) wet period events for daily precipitation 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).

In terms of quantitative statistics, the model overestimates, to a small extent, the 24 h accumulated precipitation values that are lower than 10 mm (20 mm) during the dry (wet) period, whereas it significantly underestimates the high rainfall accumulations for both dry and wet season events (Figure 4). More specifically, the CTL and ZTD mean absolute errors exceed 15 mm for the larger precipitation threshold, showing that both numerical experiments cannot capture the magnitude of severe rainfall (Figure 4). A similar extent of errors for intense precipitation thresholds have also been found in previous studies (e.g., see References [53,65,66]), showing that quantitative precipitation forecasting (QPF) remains a challenge for regional NWP systems due to uncertainties associated with physics parameterizations, primarily microphysics and convection, domain configuration (e.g.,

resolution and size), and initial conditions [67,68]. The improvement of a model's initial state through data assimilation results in more accurate QPF. This is evident in the present study, as the ZTD simulations reduce the deviations from the observations in the precipitation interval [20,...) mm by ~1.10 mm during both dry and wet periods (Figure 4). In the latter season, this reduction corresponds to a 5.5% (statistically significant at the 95% confidence interval) relative improvement in MAE (Figure 4b). For the rest of the rainfall thresholds, lower (higher) ZTD MAEs can be seen between 2 (0 mm) and 20 mm (2 mm) of precipitation during the wet (dry) period (Figure 4).

**Figure 4.** Quantitative model performance statistics averaged for the (**a**) dry and (**b**) wet period events for daily precipitation under six rainfall intervals 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).
