Enhancing Radar-Based Ensemble Nowcasting with CSGD: A UK Study

Probabilistic radar-based precipitation nowcasting is increasingly vital for real-time hydrological applications because not only it produces timely rainfall input but also the informative ensemble nowcasts may facilitate decision making. There are two primary sources uncertainty while using radar-based nowcasts for hydrological applications. The first one lies in nowcasting algorithm itself; for example, inaccurately predicted rainfall magnitudes and rainfield advection displacement errors, both exacerbated as the lead time increases. The second one is the ‘measurement’ error. There is a notable discrepancy between radar-derived precipitation estimates and measurements from rain gauges, underscoring the inherent uncertainties, including systematic and random errors, in radar data. This discrepancy necessitates aligning indirect radar measurements with actual ground-level precipitation for practical hydrological applications and analyses.

In this study, we focus on tackling the ‘measurement’ uncertainty, such that the applicability of ensemble nowcasts to hydrological practices can be improved. In the proposed method, rain gauge observations are treated as the ground truth. The Censored and Shifted Gamma Distribution (CSGD) model is then constructed using these gauge data and the co-located radar rainfall estimates. The use of CSGD model lies in its ability to not only condition actual rainfall estimates on radar data values but also account for precipitation climatology at gauge locations. Based on the CSGD parameters at know locations, we can further interpolate parameters for any locations within our study domain. We then employed the STEPS (Short-Term Ensemble Prediction System) to generate radar-based ensemble nowcasts, which are then adjusted at each radar pixel locations using CSGD model with the corresponding parameters. This leads to CSGD-enhanced ensemble nowcasts.

The United Kingdom, with its comprehensive weather data, served as the experimental area for this study. The 1-km UK C-band radar composite from the Met Office Nimrod System and the Met Office Integrated Data Archive System (MIDAS) gauge data were utilised. By aggregating these datasets into hourly scales, climatological and conditional CSGD parameters from 2015 to 2020 were estimated. The evaluation involved two stage. Initially, about 10% of rain gauges were excluded from the CSGD model fitting, with parameters estimated via Kriging interpolation. This is to ensure the quality of interpolated CSGD parameters. Then, a total of 30 storm events from 2021 to 2023 were selected to test the proposed method. Preliminary results show that the CSGD-enhanced ensemble nowcasts show a higher agreement with rain gauge observations as compared to the original nowcasts.

The proposed method is of great practical potential to provide not only timely but also enhanced precipitation nowcasts to critical hydrological applications, such as landslide or flooding warnings.