Integrating Weather Station Data and Radar for Precipitation Nowcasting: SmaAt-fUsion and SmaAt-Krige-GNet

TL;DR

This study tackles the limitation of radar-only precipitation nowcasting by integrating discrete, multivariate weather-station data. It introduces two novel architectures, SmaAt-fUsion and SmaAt-Krige-GNet, that fuse weather-station measurements with radar inputs—via bottleneck fusion and Kriging-based spatial maps, respectively. Evaluated on four years of KNMI Netherlands data, both methods outperform radar-only baselines, with SmaAt-fUsion offering robust gains across precipitation intensities and SmaAt-Krige-GNet delivering particular benefits for low-intensity rainfall. The results demonstrate the practical value of incorporating localized station data into deep learning-based nowcasting and point to opportunities for extending fusion to more datasets and longer lead times.

Abstract

Short-term precipitation nowcasting is essential for flood management, transportation, energy system operations, and emergency response. However, many existing models fail to fully exploit the extensive atmospheric information available, relying primarily on precipitation data alone. This study examines whether integrating multi-variable weather-station measurements with radar can enhance nowcasting skill and introduces two complementary architectures that integrate multi-variable station data with radar images. The SmaAt-fUsion model extends the SmaAt-UNet framework by incorporating weather station data through a convolutional layer, integrating it into the bottleneck of the network; The SmaAt-Krige-GNet model combines precipitation maps with weather station data processed using Kriging, a geo-statistical interpolation method, to generate variable-specific maps. These maps are then utilized in a dual-encoder architecture based on SmaAt-GNet, allowing multi-level data integration . Experimental evaluations were conducted using four years (2016--2019) of weather station and precipitation radar data from the Netherlands. Results demonstrate that SmaAt-Krige-GNet outperforms the standard SmaAt-UNet, which relies solely on precipitation radar data, in low precipitation scenarios, while SmaAt-fUsion surpasses SmaAt-UNet in both low and high precipitation scenarios. This highlights the potential of incorporating discrete weather station data to enhance the performance of deep learning-based weather nowcasting models.

Figures (6)

• Figure 1: An overview map of all the nodes included in the weather station dataset. The dataset contains a total of 22 weather stations. Some stations share identical coordinates, resulting in 17 nodes in the figure.
• Figure 2: Overview of the SmaAt-fUsion architecture. Weather station data is integrated into the model by concatenating its learned representation with that of the precipitation maps at the bottleneck of the SmaAt-UNet model. The precipitation stream processes 12 radar maps through the standard SmaAt-UNet encoder–decoder: depthwise-separable convolutions (blue arrows), max-pooling (brown), bilinear upsampling (green), and skip connections (grey). Features progressively compress through the encoder and are then expanded during decoding, forming the main spatial–temporal pathway of the network. In parallel, the station-data stream transforms the 22×8×12 tensor using two 3D depthwise-separable convolutions (pink) and an AdaptiveMaxPool3D layer to produce a compact embedding. This embedding flows into the fusion stage, where it is concatenated with the UNet bottleneck, creating a shared latent representation used for the final precipitation prediction.
• Figure 3: An illustration of the Kriging process and Overview of the SmaAt-Krige-GNet architecture.
• Figure 4: An example of Kriging map input of the eight meteorological variables at a selected timestamp. To improve visibility, Atmospheric pressure uses a percentile-based color scale due to its narrow dynamic range, while the other variables share a common normalization ($-2$ to 2).
• Figure 5: Examples precipitation input of the dataset on 2019-03-10, 2019-06-15, 2019-09-29, 2019-12-14.