This keynote will focus on recent research and developments at Météo-France's Nowcasting department. Firstly, the APIC warning system for exceptional rainfall at municipal level will be presented. This system was historically built on the comparison between observed rainfall rates of different durations and local climatological values. After aggregation at municipal level, warnings are displayed and sent automatically to over 13,000 institutional users. A 3-hour rainfall forecast, based on PIAF forecast ensembles that mix QPE and QPF extrapolation from the AROME-NWC model, has recently been added to these warnings. The way in which preferred values are deduced from these ensembles and how this information is communicated to subscribers will be described.At the same time, much effort has been put into improving and completing the OPIC radar-based storm objects, which are detected every 5 minutes in the French radar mosaic, tracked on consecutive images and whose severity level is characterized on the basis of its attributes. Two different approaches were first tested to predict object contours up to an hour in advance : (1) by advection of points sampled from the contour and (2) using a convolutional neural network learned over 3 years of OPIC observation. An interesting complementarity was found, since the probability of occurrence of storm objects deduced from the second method includes the contours forecasted by the first. Machine learning algorithms were also calibrated on 7 months of object attributes collected during their tracking, such as morphological parameters, co-located observations of different types, environmental variables deduced from NWP models. The results and validation of the resulting one-hour forecasts of general severity levels and hazards such as precipitation and hail will be discussed.
Montmerle et al. (Fri,) studied this question.