Comparison of weather radar, numerical weather prediction and gauge-based rainfall estimates

Abstract

Hydrological hazards have been extensive world-wide in recent times. In particular, there has been widespread flooding across much of Australia in response to extreme precipitation events. Hydrological modelling can be used to effectively manage the extensive effects of flood events, but the primary input for hydrological hind-, now- and fore-casting of these events is reliable knowledge of both observed and forecast precipitation. The Bureau of Meteorology's rain gauge network is the most common source of data on precipitation in Australia, and is the benchmark for validation of other estimates, despite its limitations. However, the rain gauge network is spatially limited for much of Australia, with interpolation commonly required. Moreover, the uncertainty of these estimates increases with distance from the gauge, there can be significant lag times in receiving the data (especially from daily gauges), and they give no indication of the likely precipitation in the coming hours. In contrast, weather radars present the opportunity to measure precipitation data with good spatial and temporal resolution. However, weather radars require repeated adjustment against the gauge network to decrease their inaccuracies. In addition, the accuracy of the rainfall declines with distance from the radar, because the radar scans are too high to see low-level precipitation at far range. Moreover, there is limited spatial coverage by quantitative weather radars in Australia, which are mostly located near the capital cities and along the eastern coastline. Numerical weather prediction (NWP) models can provide information on precipitation with national coverage for different spatial and temporal resolutions with lead times out to several days, but the accuracy of this information is uncertain. This uncertainty is due to high variability of rainfall in space and time, and the inadequacy of current numerical weather prediction physics. Many studies have focused on developing effective methods for error estimation of these predictions, because of the considerable potential for its use in hydrologic applications, but have been limited to individual storms. Consequently, the best estimation of precipitation data for flood forecasting in Australia is likely to include a combination of all three approaches (gauge network, weather radar observations and nowcasts, together with numerical weather prediction forecasts), but the uncertainty in each must be understood through longitudinal studies using independent precipitation data. This paper undertakes an extensive inter-comparison of these approaches using NWP outputs from the Australian Community Climate Earth-System Simulator (ACCESS) and the Yarrawonga radar data, using an independent rain gauge network in the Murrumbidgee catchment in south eastern Australia. Several statistical comparisons were made using hourly and daily accumulation of rainfall from January to August 2010 to estimate the errors between radar/NWP and the gauge network, with results presented for one representative station. It was found that the radar has fewer false alarms than NWP (26% versus 74 % for hourly rainfall). Most large errors for radar and NWP estimates were found to occur between Feb-April, which can be related to the particular weather events in these months. Based on average value of bias, radar mostly underestimated rainfall with the largest difference from gauge measurements in March (-0.78 mm/hr and -3.44 mm/day for hourly and daily data accumulations respectively) while NWP mostly overestimated the rainfall (0.07-0.09 mm/hr and 0.64-0.89 mm/day). Moreover, NWP with 12-24 and 24-36hr lead times (0.97 mm/hr and 3.39 mm/day RMSE for hourly and daily accumulations respectively) outperformed the 0-12 and 36-48hr lead times, while radar had larger errors (1.38 mm/hr and 4.17 mm/day for hourly and daily accumulations respectively). It should be noted that the gauge station was located a large distance from the radar, potentially leading to larger than usual errors. Additionally, comparison of a single point gauge rainfall estimate may also exaggerate the errors from NWP which has a coarse resolution of 12 km. Moreover, the usefulness of radar data in hydrological modelling is limited by the radar coverage for this catchment, the large amount of missing data, and the relatively poor performance compared to the model predictions. However, it is recommended that further studies of radar and model predictions be carried out for large experimental areas.