Rainfall Classification of Kilometer-Scale Ensemble Forecasts Using Convolutional Neural Networks and SOMs

Abstract Ensemble forecasting has been developed for many years to take into account the chaotic behavior of the atmosphere. Ensembles provide large volumes of data, from which it is difficult to extract the full benefit for some downstream applications. This is why the use of summary products, such as clustering, is becoming increasingly valuable. However, grouping similar ensemble members requires defining relevant similarity metrics, which can become challenging for high-resolution intermittent fields such as rainfall. A new prototype method is presented to overcome this problem. It combines a compression step and a clustering step. The compression step consists of a convolutional autoencoder (CAE) that projects the rainfall fields into its latent space. Input data for the CAE are a 4-yr database of 1-h accumulated rainfall forecasts from the French convection-permitting ensemble prediction system. Different data preprocessing techniques and hyperparameters are tested to select the best configurations. CAEs are shown to objectively and subjectively outperform more traditional approaches such as principal component analysis (PCA) and wavelets. The clustering step consists of a self-organizing map (SOM) that incorporates the forecasters’ constraints in two stages. Different SOMs are tested on a large rainfall dataset projected into a CAE latent space. The best one in terms of topographic error is selected. After a merging step with forecasters, 22 classes remain. They serve as a basis for the design of a useful experimental product dedicated to forecasters, which is illustrated with several examples. Significance Statement Ensemble forecasts provide a range of possible weather outcomes for the future. To handle the different forecast possibilities in an operational context, similar forecasts can be grouped together. The novelty of this research lies in applying this method to ensemble forecasts of 1-h accumulated rainfall. The prototype method uses a neural network called a convolutional autoencoder and a self-organizing map. The neural network provides more satisfactory results than more classical approaches. It enables the creation of a finalized product for synthesizing rainfall forecasts.