Comparing four radar rainfall nowcasting algorithms for 1481 events

Accurate and timely hydrological forecasts highly depend on their meteorological input. Current numerical weather predictions (NWP) do not have sufficiently high spatial and temporal resolutions for adequate use for short lead times (less than six hours ahead) in fast-responding mountainous, lowland and polder catchments. Therefore, radar rainfall nowcasting, the process of statistically extrapolating the most recent radar rainfall observation, is increasingly used. However, as most studies consist of analyses based on a relatively small sample of generally 2–15 events, best practices for the use and choice of these algorithms within operational forecasting systems are not yet available. In this study, we aim to determine the predictive skill of radar rainfall nowcasting algorithms for the short-term predictability of rainfall, in which we focus on different lowland catchments in the Netherlands. We concentrate particularly on the dependency of the forecast skill on catchment and environmental characteristics, such as event type and duration, seasonality, catchment size and location with regard to the radar location and prevailing wind direction. For this purpose, we performed a large-sample analysis of 1481 events spread over four event durations and twelve lowland catchments (6.5–957 km²). Four algorithms were tested and compared with Eulerian Persistence: Rainymotion Sparse and DenseRotation, pySTEPS deterministic and pySTEPS probabilistic with 20 ensemble members. Maximum skillful lead times increased for longer event durations, due to the more persistent character of these events. In all cases, pySTEPS deterministic attained the longest maximum skillful lead times: 25 min for 1-h, 39 min for 3-h, 56 min for 6-h and 116 min for 24-h durations. During winter, when more persistent stratiform precipitation is present, we found three times lower mean absolute errors than for nowcasts during summer with more convective precipitation. For the fractions skill score, higher skill was obtained with increasing grid cell sizes. This was advantageous for larger catchments, whereas some catchments became smaller than the grid size after upscaling. Catchment location mattered as well: up to two times higher skillful lead times were found downwind of the radars than upwind, given the prevailing wind direction. The pySTEPS algorithms outperformed Rainymotion benchmark algorithms due to rainfall field evolution estimations with cascade decomposition and an autoregressive model. We found that most errors still originate from growth and dissipation processes which are not or only partially (stochastically) accounted for.