A Strategy to set up Test Dataset and Evaluation Benchmark for Radar Nowcasting of Precipitation in Italy

Prediction of extreme precipitation with high spatial resolution on short time scales (i.e., nowcasting) is still challenging, and data-driven approaches such as artificial intelligence tools are increasingly being used. In this respect, two factors are undoubtedly important: 1) The need of robust databases of temporal and spatial evolution of rain precipitation patterns to train new nowcasting routines, and 2) the establishment of an exhaustive benchmark to evaluate the improvements brought by new prediction algorithms. This article aims to contribute to the two points just mentioned by describing a novel practical-to-use radar data screening method and by analyzing the performance of nine existing radar nowcasting techniques to establish a minimum acceptable performance (MAP) level. The radar dataset used consists of 111 955 frames (1.5 year of data) at 1 × 1 $\text{km}^{2}$ resolution sampled 5 min apart over Italy, a country scarcely investigated so far in terms of radar nowcasting compared to other areas, and that may be considered a climate change hotspot. The results reveal a large variability in the performance of each tested nowcasting method, an aspect often not considered in similar studies but that evidences as the differences among the skills of various nowcasting methods can be systematically masked by the natural variability of each prediction outcome. The seasonal dependence of the nowcasting capacity is also shown and linked to the presence of orographic convective precipitation that occurs in Italy during the summer season. This last point also suggests for the future the need to implement convection-specific nowcasting modules, possibly optimized on local areas to improve the forecasting skills in case of atmospheric instability.