Jussi Leinonen, Daniele Nerini, Alexis Berne
Generative adversarial networks (GANs) have been recently adopted for super-resolution, an application closely related to what is referred to as "downscaling" in the atmospheric sciences: improving the spatial resolution of low-resolution images. The ability of conditional GANs to generate an ensemble of solutions for a given input lends itself naturally to stochastic downscaling, but the stochastic nature of GANs is not usually considered in super-resolution applications. Here, we introduce a recurrent, stochastic super-resolution GAN that can generate ensembles of time-evolving high-resolution atmospheric fields for an input consisting of a low-resolution sequence of images of the same field. We test the GAN using two datasets, one consisting of radar-measured precipitation from Switzerland, the other of cloud optical thickness derived from the Geostationary Earth Observing Satellite 16 (GOES-16). We find that the GAN can generate realistic, temporally consistent super-resolution sequences for both datasets. The statistical properties of the generated ensemble are analyzed using rank statistics, a method adapted from ensemble weather forecasting; these analyses indicate that the GAN produces close to the correct amount of variability in its outputs. As the GAN generator is fully convolutional, it can be applied after training to input images larger than the images used to train it. It is also able to generate time series much longer than the training sequences, as demonstrated by applying the generator to a three-month dataset of the precipitation radar data. The source code to our GAN is available at https://github.com/jleinonen/downscaling-rnn-gan.
Blandine Bianchi, Peter Jan van Leeuwen, Robin J. Hogan, Alexis Berne
The objective of this work is to provide high-resolution rain rate maps at short lead-time forecasts (nowcasts) necessary to anticipate flooding and properly manage sewage systems in urban areas by combining radars, rain gauges, and operational microwave links, and taking into account their respective uncertainties. A variational approach (3D-Var) is used to find the best estimate for the rain rate, and its error covariance, from the different rain sensors. Short-term rain rate forecasts are then produced by assuming Lagrangian persistence. A velocity field is obtained from the operational radar-derived rain fields, and the rain rate field is advected using the Total Variance Diminishing (TVD) scheme. The error covariance associated to the estimated rain rate is also propagated, and we use these two in the 3D-Var at the next observation time step. This approach can be seen as a Variational Kalman Filter (VKF), in which the covariance of the prior is not constant but dependent on time. The proposed approach has been tested using data from 14 rain gauges, 14 microwave links and the operational radar rain product from MeteoSwiss in the area of Zurich (Switzerland). During the applications the assumption of the Lagrangian persistence appears to be valid up to 20 min (a bit longer for stratiform events). During convective events, the algorithm is less powerful and shorter lead times should be considered (i.e., 15 min). Although such lead times are short, they are still useful to various hydrological and outdoor applications.
Milton Gomez, Louis Poulain--Auzeau, Alexis Berne, Tom Beucler
Numerical Weather Prediction (NWP) models that integrate coupled physical equations forward in time are the traditional tools for simulating atmospheric processes and forecasting weather. With recent advancements in deep learning, AI-based Weather Prediction models that rely on neural network architectures$\unicode{x2013}$Neural Weather Models (NeWMs)$\unicode{x2013}$have emerged as competent medium-range NWP emulators, with performances that compare favorably to state-of-the-art NWP models. However, they are commonly trained on reanalyses with limited spatial resolution (e.g., 0.25° horizontal grid spacing), which smooths out key features of weather systems. For example, tropical cyclones (TCs)$\unicode{x2013}$among the most impactful weather events due to their devastating effects on human activities$\unicode{x2013}$are challenging to forecast, as extrema are smoothed in deterministic forecasts at 0.25° resolution. To address this, we use our best observational estimates of wind gusts and minimum sea level pressure to train a hierarchy of post-processing models on NeWM outputs. Applied to Pangu-Weather and FourCastNet v2, the post-processing models produce accurate and reliable forecasts of TC intensity up to five days ahead. Our post-processing algorithm is tracking-independent, preventing full misses, and we demonstrate that even linear models extract predictive information from NeWM outputs beyond what is encoded in their initial conditions. While spatial masking improves probabilistic forecast consistency, we do not find clear advantages of convolutional architectures over simple multilayer perceptrons for our NeWM post-processing purposes. Overall, by combining the efficiency of NeWMs with a lightweight, tracking-independent postprocessing framework, our approach improves the accessibility of global TC intensity forecasts, marking a step toward their democratization.