Nowcasting by analogues: a generalization using radar data sequences

Reliable quantitative precipitation forecasts are the essential inputs to hydrological modeling and early warning systems. The assimilation of high resolution radar data in numerical models is computationally demanding and needs corresponding parameterizations of the physical processes at fine spatial scales to become effective. Therefore, simple approaches as rainfall extrapolation- and expert-based nowcasting systems with high spatial and temporal resolutions are widely used as they meet hydrological and user needs in resolution and timeliness and prove their efficiency against numerical models.

Lagrangian advection schemes are efficiently used for nowcasting precipitation from radar images in continental areas. But their applicability in Alpine terrains is limited by two main factors. First, the motion vector field, as estimated by optical flow on subsequent radar images or using Doppler-derived wind fields, is inadequate to advect the precipitation field in the presence of orographic blocking where the large scale precipitation features remain stationary. Secondly, they cannot handle the evolution of precipitation in terms of orographic enhancements, triggered convection or dissipation.

NORA (Nowcasting of Orographic Rainfall by means of Analogues) is a method overcoming these drawbacks. It is currently used for short-term forecasting of orographic rainfall in the southern side of the Swiss Alps. A 5-year historical archive representative of orographic precipitation events is winnowed to search for analogue situations in terms of similar mesoscale flows and radar precipitation patterns. The subsequent evolution of the analogues constitutes a natural ensemble of hourly rainfall forecasts with realistic statistical properties.

This study aims at generalizing the search of analogues by looking at similarities of radar image sequences instead of comparing single images in an attempt to capture the dynamics of atmospheric processes in an adequate way. The approach is based on the embedding of radar data sequences in a lower dimensional phase space characterizing the evolution of rainfall patterns. Principal component analysis (PCA) is applied to the archive of time-ordered radar images organized into a matrix where each row represents an image and each column a pixel in the image. The evidence of similar trajectories in the eigen-space motivates the application of the method of analogues. This research explores the predictability of the system according to different embedding dimensions (PCA and mesoscale flows), lead times, number of analogues and length of retrieved sequences. Initial results suggest that predictability of rainfall at a particular lead time is dependent on the number and length of retrieved sequences. In particular we observed that the forecast skill at short (long) lead times can be enhanced by retrieving short (long) sequences.