Cell-tracking-based framework for assessing nowcasting model skill in reproducing growth and decay of convective rainfall

The rapid temporal evolution of convective rainfall poses a challenge for quantitative rainfall nowcasting models that forecast rainfall on timescales ranging from 5 min to 6 h. With the growing potential of machine learning models for precipitation nowcasting to produce realistic-looking nowcasts for long lead times, it is important to investigate whether the nowcasts also produce realistic development for convective rainfall. Common verification metrics traditionally used to validate nowcasting models are often dominated by large-scale stratiform rainfall, and averaging the metrics across entire precipitation fields obscures how accurately the models replicate individual convective cells, which makes it difficult to distinguish the model skill for the growth and decay of convective rainfall. In this study, we present a framework based on the tracking of convective cells to investigate how accurately nowcasting models reproduce the development of convective rainfall. In the framework, a cell identification and tracking algorithm is applied first to the input observation rainfall fields and then separately to the target observation and nowcast rainfall fields where the tracks identified in the input observations are continued. Features describing the cells and cell tracks, such as the cell volume rain rate and area, are then extracted. In addition to the errors in these feature values, the models' skill in reproducing the existence of convective cells is estimated by calculating several contingency table metrics, such as the critical success index. The results allow the analysis of how accurately the models reproduce the growth and decay of convective rainfall and quantify the differences between the models, for example, due to differences in how the models smooth the nowcasts (i.e. blurring). The framework also allows differentiation of the results based on the initial conditions of the cell tracks, demonstrated here by separating the tracks into decaying or growing cell tracks based on the cell status when the nowcast is created. We demonstrate the framework with four open-source nowcasting models: the advection nowcast, the S-PROG (Spectral Prognosis; Seed, 2003) and LINDA (Lagrangian Integro-Difference equation model with Autoregression; Pulkkinen et al., 2021) models from the pysteps library, and the L-CNN (Lagrangian Convolutional Neural Network; Ritvanen et al., 2023) model, with data from the Swiss radar network. The results indicate that the L-CNN model reproduced the existence of convective cells best among the models and had smaller errors in the cell volume rain rate than LINDA and S-PROG. LINDA had the smallest underestimation in the cell mean rain rate, whereas S-PROG significantly overestimated the cell volume rain rate and area because of blurring.