Analysis and application of CNN to improve deterministic optical flow nowcasting at DWD

Optical flow based nowcasting is a powerful technique to compute forecasts for small lead times up to a couple of hours. Forecasts of radar data are very useful as standalone products and serve as input data for multiple operational products, e.g., for the seamless combination of radar and NWP data and for cell-based analysis and prediction products. However, the optical flow technique has several drawbacks. It assumes stationarity in both the data values as well as the advection information. Further, it is a deterministic technique and the nowcasts have no dynamic properties. Recently, machine learning techniques have shown promising results for producing nowcasts with dynamic properties. However, for radar reflectivity and precipitation data, the predictions often lack high-intensity values and tend to become blurry for larger lead times.

In the current work we explore the potential of deterministic convolutional neural networks (CNN) to improve the operational optical flow nowcasting at DWD. A two-year dataset consisting of radar, NWP and orography data is used for training modified UNet based neural networks. Each network predicts radar reflectivity composites for a specific lead time, in 5-minute steps. Several optimization techniques are combined, both for the input data and the network architecture. The input data contains optical flow based nowcasts of previous radar timesteps that are mapped to the target lead time. The NWP input parameters are chosen for their known importance in convective processes. To understand their impact in this application, an ablation study is performed. The network architecture is optimized. The classic UNet architecture is augmented with additional horizontal computation blocks. This adds more nonlinearity to the finer scales of the network and reduces the validation error. Individual encoders are used for the radar and NWP data and combined with affine linear transformations. Experiments with classical pointwise loss functions as well as losses with spatial context (e.g., FSS) are conducted. The new forecasts are compared with the operational nowcasting at DWD as well as a closed-loop NN approach (RainNet).