by Laure Raynaud

published in the Météo-France 2018 Research Report (ISSN : 2116-4541)

The utilization of precipitation forecasts could be improved with an appropriate processing of model fields. An innovative and promising approach consists in extracting from the forecast fields information at a larger, and thus more predictable, scale than the model grid, under the form of “stochastic precipitating objects”. These objects are defined by fuzzy contours within which the distribution of rainfall, in terms of intensity and/or spatial variability (also known as texture), is homogeneous. Automatically detecting these objects is however complicated and different approaches are possible. A first heuristic solution has been developed in order to detect intensity-based objects from a similarity measure between the local precipitation distribution and reference distributions.

A novel detection method, based on the use of a convolutional neural network initially developed for the segmentation of medical images, has been evaluated. The network is trained on the outputs of the existing algorithm and it provides very similar results with significant gains regarding computation time when it is executed on graphical processors.

In addition, determining the texture of precipitations, in particular their continuous or intermittent nature, is important to properly characterize the sensible weather but it is not handled by the current algorithm. Preliminary experiments with a neural network trained on a very small set of manually labelled data provide encouraging results.

In the future neural networks could allow for the simultaneous detection of intensity and texture properties, and for a more detailed characterization of precipitation types including thunderstorms.

Up : 6 h-accumulated precipitation forecast from the AROME   French model (to the left) and objects corresponding to total (yellow), moderate (orange) and heavy (red) precipitation detected by the neural network (to the right).

Below : 6 h-accumulated precipitation forecast from the AROME   French model (to the left) and objects corresponding to continuous (blue) and intermittent (red) precipitation detected by the neural network (to the right).