Blending radar, NWP and satellite data for real-time ensemble rainfall analysis: a scale-dependentmethod

[摘要] This report is the second in a series on the development path towards a national-scale real-time rainfall analysis system for Australia. The document provides a description of a method suitable for real-time and near real-time rainfall analysis and ensemble generation. A preliminary evaluation of the method for a study area and period is also presented. The method developed uses components of the short-term ensemble prediction system (STEPS; Seed et al. 2013) adapted for rainfall analysis (as opposed to forecasting). Specifically the method employs a multiplicative cascade representation of rainfall, blends multiple sources of rainfall data at the cascade level, derives weights for the blending based on mutual correlation between data at the cascade level, and computes ensembles of rainfall analyses in which each member possess the expected scaling properties of rainfall across a wide range of spatial scales.We call the method scale-dependent blending.We developed, applied and tested the scale-dependent blending method for three sources of real-time and near real-time gridded estimates of hourly accumulated rainfall from (a) Rainfields v3.0 mosaic of radar rainfall (radar), (b) ACCESS-R forecasts of rainfall accumulations (NWP) and (c) merged multi-satellite rainfall product from the GPM IMERG system (satellite). The study focussed on a 1600 x 1600 km 2 area of southeast Australia and the month of September 2016. Our investigation found that:•all three of data sets examined were very good at capturing rainfall spatial structure over a range of scales, with radar and satellite most similar in their ability, while NWP consistently showed less structure in the range 25 - 125 km scale;•the weights derived from the correlation of cascade components for each data related strongly to the ability of the data to represent structure at the corresponding spatial scale; and•the preliminary evaluation against independent data showed that the blended results often compensated for poor performance of individual data, and were almost always better than satellite data alone in areas of the country without radar coverage.Recommended future work focusses primarily on more thorough evaluation of the method, over different parts of the country (e.g. Snowy Mountains), and detailed examination of the ensembles as a true representation of rainfall uncertainty. Specific items include:(Immediate tasks in order of priority)•Assessing the impact of a revised structured noise perturbation scheme on blended ensemble means, especially in areas of transition from radar coverage to no radar coverage.•Extending the verification statistics to include metrics on detection (categorical statistics) and skill scores on related to how well the ensembles represent uncertainty.•Conducting localised verification targeting (i) coast zones with high levels of gauge coverage; (ii) in-land areas with poor or no radar coverage; and (iii) mountainous areas that pose challenge to radar and satellite retrievals.•Experimenting with temporally static weights versus those derived instantaneously, or dynamically, to quantify the expected difference in performance in real-time and near real-time settings.(Further development, especially ahead of national-scale implementation)•Exploring impact of non-stationarity on method and consider the approaches of concatenating a series of localised analyses, or using spatially varying weights in the blending procedure.•Incorporating the result of independent evaluation of the source data though either an a priori bias correction or guidance on the appropriate use choice of data to serve as the reference of each cascade level.

[发布日期] 2017-03-30 [发布机构] CSIRO

[学科分类] 地球科学(综合)