Transformer-based nowcasting of radar composites from satellite images for severe weather

Abstract

Weather radar data are critical for nowcasting and an integral component of numerical weather prediction models. While weather radar data provide valuable information at high resolution, their ground-based nature limits their availability, which impedes large-scale applications. In contrast, meteorological satellites cover larger domains but with coarser resolution. However, with the rapid advancements in data-driven methodologies and modern sensors aboard geostationary satellites, new opportunities are emerging to bridge the gap between ground- and space-based observations, ultimately leading to more skillful weather prediction with high accuracy. Here, we present a Transformer-based model for nowcasting ground-based radar image sequences using satellite data up to two hours lead time. Trained on a dataset reflecting severe weather conditions, the model predicts radar fields occurring under different weather phenomena and shows robustness against rapidly growing/decaying fields and complex field structures. Model interpretation reveals that the infrared channel centered at 10.3 $μm$ (C13) contains skillful information for all weather conditions, while lightning data have the highest relative feature importance in severe weather conditions, particularly in shorter lead times. The model can support precipitation nowcasting across large domains without an explicit need for radar towers, enhance numerical weather prediction and hydrological models, and provide radar proxy for data-scarce regions. Moreover, the open-source framework facilitates progress towards operational data-driven nowcasting.

Figures (23)

• Figure 1: A thunderstorm wind event from test dataset (ID=S857496 the in SEVIR catalog, also accessible through the NCEI catalog via https://www.ncdc.noaa.gov/stormevents/eventdetails.jsp?id=857496). Input data are shown in the first three rows for the channels C09 (6.9 $\mu m$), C13 (10.3 $\mu m$), and lightning counts, respectively. Target VIL fields are shown in the fourth row, while model predictions for our model (EF Sat2Rad) and the baseline model (U-Net) are shown in the fifth and sixth rows, respectively. Time steps are given at the top of both input and output data, separated with a dashed line. Note that temporal intervals for input and output are 5 minutes, but plotted with varying intervals for visualization purposes.
• Figure 2: Summary of model performance using FSS at different spatial scales for our model and the baseline, indicated by solid and dashed lines, respectively.
• Figure 3: Relative importance of channel and length of input data for different VIL thresholds, quantified via skill score of each permuted chunk compared to the original model performance as $1 - FSS_{X}/FSS_{ref}$ where $FSS_{ref}$ is the score of the original test set (summarized in Fig. \ref{['fig:modelPerf']}) and $FSS_{X}$ is the score of the same dataset with the $X$ chunk permuted: (a) For input channels (b) For temporal extent of input data.
• Figure 4: Same as Fig. \ref{['fig:modelOut']}, but for a hail event with ID=S857225 in the SEVIR catalog.
• Figure 5: Same as Fig. \ref{['fig:modelOut']}, but for an event with ID=R19061502217821 in the SEVIR catalog.