Evaluation of radar rainfall nowcasting techniques to forecast synthetic storms of different processes

Early hydrological hazard warning demands precise weather forecasts to accurately predict the timing and the location of intense precipitation events which can cause severe floods/landslides and present risks to urban and natural environments. Extrapolation of precipitation by radar rainfall products at high space and time scales with short lead times outperforms forecasts of numerical weather prediction. Therefore, developing and improving of rainfall nowcasts systems are essential. Rainfall nowcasting is the process of forecasting precipitation field movement and evolution at high spatial and temporal resolutions with short lead times(<6h) in which the advection of the precipitation fields is estimated by extrapolating real-time remotely sensed observations. Radar rainfall nowcasting is increasingly applied because of the high potential of radar products in short-term rainfall forecasting due to their high spatiotemporal resolutions (typically, 1 km and 5 min). It consists of two procedures in tracking precipitation features to calculate the velocity from a series of consecutive radar images and propagating the most recent precipitation observation into the future using the obtained velocity. Optical flow represents one of the most used methods for tracking the motion fields from consecutive images. Deep learning techniques are those machine learning methods that utilise deep artificial neural networks. Deep learning has become one of the most popular and rapidly spreading methods in different scientific disciplines including water-related research. Deep learning applications in radar-based precipitation nowcasting is still in its early stage with many knowledge gaps and their full potential in rainfall nowcasting requires more investigation. This work evaluates the performance of a deep convolutional neural network (called rainnet) and three optical flow algorithms (called Rainymotion Sparse, Rainymotion Dense, Rainymotion DenseRotation) compared with Eulerian Persistence to assess their predictive skills in nowcasting. Synthetic precipitation scenarios have been created with different motion fields (linear and rotational motions), velocities, intensities, sizes, and locations. The models have been evaluated to forecast different precipitation processes that contribute mainly to model errors such as constant and accelerated linear and rotational motions, growth and decay in both size and intensity. Different verification metrics have been used to evaluate the skill of the forecasts.

 

Keywords: radar rainfall nowcasting; deep learning; optical flow; extrapolation; rainnet; rainymotion