An improved statistical approach to merge satellite rainfall estimates and raingauge data.

Summary Deriving high quality daily rainfall estimates are required not only for successful hydrological modelling but also for its application in ungauged basins. At present, there are two resources to estimate rainfall fields: raingauge stations and remote sensing devices (such as satellites and radar). The observations obtained from raingauges are accurate and reliable, but are usually limited to interpolate daily rainfall as input into hydrological models by spatial coverage. Hence, the areal precipitation estimates constructed solely by raingauges exhibit a great deal of uncertainty where the raingauge network is sparse, especially for local convective events. Satellite-based gridded precipitation estimates are produced at high temporal (3 h) and reasonably high spatial resolution (0.25° × 0.25°). However, the accuracy of these surfaces is lower than using ground based observations. The main aim of this paper is to present a novel method of generating a synthesis of a discrete set of point raingauge observations and a satellite derived rainfall product. This method stems from some ideas of data assimilation and focuses on the difference between gridded and point data. A nonparametric kernel smoothing method is employed in this merging strategy with emphasis on discontinuity correction and spatial interpolation adapting for sparse design. A cross-validation study was undertaken to blend observations from the Australian raingauge network and satellite derived TRMM Multisatellite Precipitation Analysis (TMPA) 3B42. In order to test the proposed method, several methods including ordinary kriging and cokriging were trialled. The comparison showed that the bias results produced by the proposed method were far superior to the other methods trialled and other error statistics such as root mean square error were comparable or better (although the kriging methods tended to perform better in densely gauged areas). In particular, the bias of the proposed method was reduced consistently over the whole country, whilst the kriging methods trialled produced relatively large biases across the country leading to near overall zero bias. In addition to improving quantitative efficiency, the estimated rainfall field showed good visual performance in the presence of sparse gauge stations.