Incorporating cell evolution into object-based convective storm nowcasting

Radar-based nowcasting plays a crucial role in meeting the urgent demand for short-term, high-intensity convective rainfall predictions. Given the dynamic and clustering nature of convective storms, object-based nowcasting has emerged as an effective approach, characterised by its ability to identify, track, and extrapolate their motion. These models excel in identifying rainfall objects in radar images and constructing their temporal associations. However, a critical limitation in many of existing methods lies in their lack of mechanism to incorporate the evolution of rain cell intensity into the nowcasting process.

A recent study by Cheng et al. (2023) demonstrated the effectiveness of utilising convective core altitude – a property retrieved from three-dimensional radar data– to improve the prediction of the evolution of single-core convective cell lifecycle. Their results suggest that, compared to persistence nowcasts, the prediction errors in rainfall intensity can be reduced by 50% at 15-min forecast lead time. However, this model focused on predicting ‘mean’ cell properties, while promising, still falls short for operational forecasting.

This research aims to enhance object-based nowcasting by developing methods to integrate the cell evolution model proposed by Cheng et al. (2023) with an operational positional forecasting model. A recent development of a Kalman filter based object-based convective storm nowcasting model, co-developed by researchers from several international sectors and the UK Met Office (Wang et al., 2022), is employed here for positional prediction of convective cells. The key challenge of the integration lies in producing spatial-distributed convective cell nowcasts and the associated evolution prediction uncertainty that can be incorporated with the positional prediction uncertainty under a Kalman filter framework. 

To tackle this challenge, we test on two approaches. Inspired by Shehu and Haberlandt (2022), the first approach employs a cell analog method to identify historical cells with similar mean properties predicted by the cell evolution model. Those cell analogs with high similarity are then used to empirically constitute prediction uncertainty. For the second approach, we generate spatially-distributed cells via fitting bivariate Gaussian or Exponential shape (Willems, 2001; Féral et al., 2003) models using predicted mean properties, which can further cell samples with similar mean properties. Two approaches will then be integrated with the positional nowcasting model, respectively. Probabilistic nowcasting will be undertaken to generate ensemble nowcasts that account for both positional and evolution variations. Ensemble members from each approach at each forecasting time step can constitute a convective storm ‘hazard’ map. We will indirectly evaluate the performance of two proposed approaches via assessing these hazard maps with radar observations.