Improving rain cell tracking for convective rainfall nowcasting: a two-stage analogue model approach

Rain cell tracking methods are essential to the object-based rainfall nowcasting of convective storms. These methods identify rain cells from radar images and the tracks associate cells between any two successive time steps. Based upon the identified cells and tracks, the positions of the cells in the next few time steps can be forecasted. Many existing nowcasting methods assume Lagrangian persistence. That is, they generally lack the mechanisms to predict the temporal evolution of cell properties and their types. This deficiency may have a great impact to the accuracy of the convective storm nowcasting. To improve cell tracking methods, a two-stage analogue model is proposed to address the limits of existing cell tracking methods.

• Predicting cell type: three machine learning classifiers –KNN, logistic regression and random forest—are employed to predict the cell types based on rain cell properties.
• Predict temporal evolution of cell properties: an ensemble forecast (0-1h lead times) of cell mean intensity, maximum intensity, size and major axis length is obtained using an analogue method. This method assumes that rainfall cells with similar conditions will evolve similarly. Analogues are chosen based on the predicted cell type from the previous step.

In this study, a dataset of rainfall cells from a total of 165 convective storms between 2005 and 2017 is used. These rainfall cells are identified using enhanced TITAN. The study area is centred at Birmingham city, with an area of 512 × 512 km². Results show that the random forest classifier has the best performance in predicting track types. As the temporal profile of the selected cell properties is incorporated into the prediction process, the prediction accuracy of the random forest classifier can be higher than 80%. Results also show that predicting cell type prior to the selection of analogues improves the forecasting of temporal evolution of cell properties at a lead time of 5 minutes. Overall, the analogue method enhances the prediction of temporal evolution of cell properties compared with assuming Lagrangian persistence. At the moment, cell types are predicted for a 5-minute lead time.