Correction of hourly radar precipitation data based on rain-gauges values: what is the most efficient method for hydrologic modeling purposes?

The effectiveness of a hydrologic model is largely driven by the availability and nature of the input data. Among these, many studies proved precipitation to be the most important because it regulates the amount of water entering the system. Spatially continuous precipitation data can be obtained from radar technology. However, radar precipitation values are an indirect measure, and it is widely believed that their use in hydrologic modelling is complicated due to the presence of bias. The use of radar data is increasingly problematic in mountain regions where elevation plays a key role on precipitation, creating significant variations in few kilometers. Also, mountains can lead to a shadow effect of the radar beam.

The research objective is to integrate precipitation data derived from the radar into a partially distributed hydrologic model, running in an area with complex morphology. The study area is a portion of Upper Valtellina valley (about 2300 km²), located within the Alpine belt on the border between Italy and Switzerland, and characterized by an elevation range between 350 and 3400 m a.s.l. The hourly series of 22 rain-gauges (18 Italian and 4 Swiss stations) and hourly precipitation from a radar dataset (1km x 1km resolution, from MeteoSWISS) from 2010 to 2020 are used. The mean bias between the series extracted in the radar cells at the station locations and the series measured by rain-gauge is around -28%, indicating a general underestimation of the radar data. The targets of the correction techniques are the precipitation series at the centroids of the sub-basins defined by the hydrologic model.

For the correction, two approaches are tested: (i) the radar precipitation is corrected in every centroid of the hydrologic model subbasins (point-based correction); (ii) the radar precipitation is adjusted by spatializing the radar-station error (interpolation-based correction). The first approach is based on finding the statistical relations between the radar-station series of the three closet stations to the target centroid and applying the statistical correction (Copula or Cumulative Distribution Function (CDF) matching bias correction) to the precipitation series in the centroid cell. The result of the correction is a combination of the statistical relationships weighted according to a Triangular Irregular Network. The second technique focusses instead on the interpolation of the error (residuals) calculated as the difference between radar and rain-gauge values, which is subsequently added to the original radar raster. Two different interpolation techniques are used: Thin Plate Splines and Inverse Distance Weighting. All methods are evaluated through performance indices (KGE and RMSE) at the station locations by Leave One Out cross validation.

Point-based applications are cost-effective and require less computational effort than spatial interpolations. Preliminary results show that the point-based corrections through Copula and CDF have similar performances. In detail, the KGE increases from 0.18 to 0.52 and 0.55 for Copula and CDF, respectively. RMSE decreases from 0.78 mm to 0.53 mm (Copula) and 0.62 mm (CDF). Interpolation-based corrections are still ongoing, therefore there are no definite results regarding the comparative effectiveness of one type of correction over the other.