A neural network-based observation operator for weather radar data assimilation

TL;DR

This work tackles the challenge of assimilating radar reflectivity into numerical weather prediction by learning a differentiable, nonlinear observation operator. A convolutional encoder–decoder neural network (ResUNet) is trained to map ALADIN state variables (temperature, humidity, winds at four pressure levels, plus surface fields) to radar reflectivity, producing model-equivalent observations $\hat{\mathbf{y}} = D(E(\mathbf{x}))$ that are embedded in a 3DVar framework with a lightweight background-error model. Across multiple regimes and a Slovenian floods case, the operator yields reflectivity patterns that closely resemble observations and generates localized, multivariate increments that align with convective structures, improving the analysis quality in radar space. While promising, the study notes limitations including lack of hydrometeor variables in the control vector and the need for full-cycle, multi-radar experiments to quantify forecast impacts and scalability.

Abstract

In three-dimensional variational data assimilation (3DVar) for numerical weather prediction (NWP), the observation operator $\mathcal{H}$ plays a central role by mapping model state variables to an observation equivalent. For weather radar, however, specifying $\mathcal{H}$ is particularly challenging: reflectivity is a nonlinear, microphysics-dependent diagnostic quantity that only indirectly relates to the model's prognostic variables, making traditional parameterised radar operators complex, regime-dependent and difficult to tune. In this study, we propose a neural-network (NN)-based observation operator for radar reflectivity and apply it within a 3DVar framework. Using five years (2019-2023) of radar reflectivity data from the Lisca radar and 4.4 km-resolution short-range forecasts from ALADIN model over Slovenia, we train a convolutional encoder-decoder neural network to map model temperature, humidity, horizontal wind components and surface pressure fields to radar reflectivity. Across independent test cases spanning clear-sky, stratiform, and convective regimes, the NN-based operator accurately reproduces the spatial structure and intensity of observed reflectivity, relying primarily on the model state near the observation point. In the extreme precipitation case, which caused widespread floods in Slovenia on August 4, 2023, assimilating the full radar disc reduces the domain-averaged reflectivity root-mean-square error from 5.99 dBZ to 3.47 dBZ and improves the alignment between the analysed and observed convective bands. Embedded within 3DVar, the Jacobian of the NN observation operator allows radar reflectivity observations to inform model state variables, producing corresponding analysis increments. The proposed NN radar observation operator offers a flexible alternative to traditional parameterised radar operators for improving convective-storm forecasts.

Figures (13)

• Figure 1: Spatial overview of the datasets used for training the neural-network observation operator. (a) spatial domain and radar sampling. The red square indicates the spatial extent of the ALADIN data used in the analysis, and the black dots indicate the coverage of the Lisca radar of $0.5^\circ$ elevation. Red areas on the radar disc indicate regions where terrain mountain shadow zones block the radar beam. (b) 3D geometry of the Lisca radar’s first elevation scan ($0.5^\circ$).
• Figure 2: Spatial distribution of the global radar outlier mask derived from Mahalanobis-distance analysis of reflectivity data. The mask was computed only for data points within a 160 km radius of the Lisca radar that had already passed the initial filtering steps. Gates exceeding the 90th percentile distance threshold were classified as outliers (dark red) and excluded from subsequent use to enhance data quality.
• Figure 3: Example of the radar–data cleaning procedure. (a) raw reflectivity field showing radial artefacts, distant-range noise, and terrain-induced beam blockage effects. (b) cleaned field after applying the radial filtering, retaining only the observations in a disc of 160 km from the radar centre, and the Mahalanobis-distance outlier mask. (c) difference between the raw and cleaned fields, with filtered gates highlighted in dark red.
• Figure 4: Spatial correspondence between the ALADIN model and radar datasets: black dots mark the model grid, and red dots denote radar observations interpolated onto the same spatial domain.
• Figure 5: Schematic overview of the convolutional encoder–decoder neural network used to emulate the 3DVar observation operator. The model is trained with input $\mathbf{x}$ consisting of ALADIN temperature (t), relative humidity (r), and horizontal wind components (u and v) at four pressure levels (975 hPa, 925 hPa, 850 hPa, and 800 hPa), together with 2 m temperature (t2m), 2 m relative humidity (r2m), and mean sea-level pressure (msl). The output ${\mathbf{y}}$ is the reflectivity field measured by the radar, interpolated onto the model grid at the corresponding time.