A Deep Learning approach for MCS detection: application for storm tracking and nowcasting

Storms deeply impact human safety and economic activities by generating severe hazards such as large hail, strong wind gusts or floods. The more critical phenomena often imply mesoscale convective systems (MCS) that originate from the aggregation of intense cells. The induced circulation reinforces the system and maintains it for a long time.

Convective-scale numerical weather prediction (NWP) models are now able to simulate realistically such storms. Nevertheless, model and displacement errors often alter the forecast accuracy and ultimately the weather alerts efficiency.

In order to study the predictability of MCS, present work attempts to automatically detect, track and characterize such system in the outputs of the operational high resolution nowcasting model used at Météo-France, Arome-NWC. Such system provides 6h forecast every hour by updating the last AROME-France forecast towards the latest conventional and radar observations, thanks to a 3D variational (3DVar) data assimilation system.

Three methods are developed and compared. The first one applies a segmentation algorithm on images of maximum simulated reflectivity (Zmax). The second one applies a watershed transformation algorithm on Zmax, using as seed the 10,8 µm brightness temperature (TB10.8) from the SEVIRI imager onboard geostationary satellite MSG, that is diagnosed by applying the RTTOV radiative transfer algorithm to simulated variables. Finally, the last one trains a convolutional neural network (CNN) from MCS that are hand-labelled from simulated Zmax and TB10.8 images. Subjective analyses and objective evaluation based on objects-oriented scores depict the third approach as the more reliable.

The operational benefits of this object-oriented approach are investigated on several convective cases using synthesis plots. Superimposing the MCS detected in the different available hourly runs of Arome-NWC, as well as comparing tracked parameters within the MCS throughout the different forecasts, help in assessing the behavior of the successive runs. That could facilitate the analysis of the forecaster in tricky situations, leading to a positive impact on the weather alert efficiency.