Improving radar rainfall estimation by merging point rainfall measurements within a model combination framework

While the value of correcting raw radar rainfall estimates using simultaneous ground rainfall observations is well known, approaches that use the complete record of both gauge and radar measurements to provide improved rainfall estimates are much less common. We present here two new approaches for estimating radar rainfall that are designed to address known limitations in radar rainfall products by using a relatively long history of radar reflectivity and ground rainfall observations. The first of these two approaches is a radar rainfall estimation algorithm that is nonparametric by construction. Compared to the traditional gauge adjusted parametric relationship between reflectivity (Z) and ground rainfall (R), the suggested new approach is based on a nonparametric radar rainfall estimation method (NPR) derived using the conditional probability distribution of reflectivity and gauge rainfall. The NPR method is applied to the densely gauged Sydney Terrey Hills radar network, where it reduces the RMSE in rainfall estimates by 10%, with improvements observed at 90% of the gauges. The second of the two approaches is a method to merge radar and spatially interpolated gauge measurements. The two sources of information are combined using a dynamic combinatorial algorithm with weights that vary in both space and time. The weight for any specific period is calculated based on the error covariance matrix that is formulated from the radar and spatially interpolated rainfall errors of similar reflectivity periods in a cross-validation setting. The combination method reduces the RMSE by about 20% compared to the traditional Z-R relationship method, and improves estimates compared to spatially interpolated point measurements in sparsely gauged areas.