Extended Data Fig. 3: Case study of performance on a challenging precipitation event starting 2019-07-30 at 15:15 UK, showing a pattern of precipitation around a low-pressure area which is slow moving, resulting in the cyclonic banded structures over England, DGMR captures extent of precipitation overall over the area, though slightly overdoes rain coverage between bands, and is significantly preferred by meteorologists (73% first choice, N = 56, p < 2 × 10−4). | Nature

Extended Data Fig. 3: Case study of performance on a challenging precipitation event starting 2019-07-30 at 15:15 UK, showing a pattern of precipitation around a low-pressure area which is slow moving, resulting in the cyclonic banded structures over England, DGMR captures extent of precipitation overall over the area, though slightly overdoes rain coverage between bands, and is significantly preferred by meteorologists (73% first choice, N = 56, p < 2 × 10−4).

From: Skilful precipitation nowcasting using deep generative models of radar

Extended Data Fig. 3: Case study of performance on a challenging precipitation event starting 2019-07-30 at 15:15 UK, showing a pattern of precipitation around a low-pressure area which is slow moving, resulting in the cyclonic banded structures over England, DGMR captures extent of precipitation overall over the area, though slightly overdoes rain coverage between bands, and is significantly preferred by meteorologists (73% first choice, N = 56, p < 2 × 10−4).

a, Geographic context for the predictions. b, A single prediction at T + 30, T + 60, and T + 90 min lead time for different models. CSI at thresholds 2 mm h−1 and 8 mm h−1 and CRPS for an ensemble of four samples shown in a bottom left corner. For axial attention we show the mode prediction and the single sample. Images are 256 km × 256 km. c, Expert meteorologist preference for the visualized prediction (axial attention uses the mode prediction; we report the percentage of meteorologists for their first-choice rating as well as the Clopper–Pearson 95% confidence interval). Maps produced with Cartopy and SRTM elevation data46.

Back to article page