Pysteps is an open-source and community-driven Python library for probabilistic precipitation nowcasting, that is, very-short-range forecasting (0–6 h). The aim of pysteps is to serve two different needs. The first is to provide a modular and well-documented framework for researchers interested in developing new methods for nowcasting and stochastic space–time simulation of precipitation. The second aim is to offer a highly configurable and easily accessible platform for practitioners ranging from weather forecasters to hydrologists. In this sense, pysteps has the potential to become an important component for integrated early warning systems for severe weather. The pysteps library supports various input/output file formats and implements several optical flow methods as well as advanced stochastic generators to produce ensemble nowcasts. In addition, it includes tools for visualizing and post-processing the nowcasts and methods for deterministic, probabilistic and neighborhood forecast verification. The pysteps library is described and its potential is demonstrated using radar composite images from Finland, Switzerland, the United States and Australia. Finally, scientific experiments are carried out to help the reader to understand the pysteps framework and sensitivity to model parameters.

Weather radar has become an invaluable tool for monitoring rainfall and studying its link to hydrological response. However, when it comes to accurately measuring small-scale rainfall extremes responsible for urban flooding, many challenges remain. The most important of them is that radar tends to underestimate rainfall compared to gauges. The hope is that by measuring at higher resolutions and making use of dual-polarization radar, these mismatches can be reduced. Each country has developed its own strategy for addressing this issue. However, since there is no common benchmark, improvements are hard to quantify objectively. This study sheds new light on current performances by conducting a multinational assessment of radar's ability to capture heavy rain events at scales of 5 min up to 2 h. The work is performed within the context of the joint experiment framework of project MUFFIN (Multiscale Urban Flood Forecasting), which aims at better understanding the link between rainfall and urban pluvial flooding across scales. In total, six different radar products in Denmark, the Netherlands, Finland and Sweden were considered. The top 50 events in a 10-year database of radar data were used to quantify the overall agreement between radar and gauges as well as the bias affecting the peaks. Results show that the overall agreement in heavy rain is fair (correlation coefficient 0.7–0.9), with apparent multiplicative biases on the order of 1.2–1.8 (17 %–44 % underestimation). However, after taking into account the different sampling volumes of radar and gauges, actual biases could be as low as 10 %. Differences in sampling volumes between radar and gauges play an important role in explaining the bias but are hard to quantify precisely due to the many post-processing steps applied to radar. Despite being adjusted for bias by gauges, five out of six radar products still exhibited a clear conditional bias, with intensities of about 1 %–2 % per mmh⁻¹. As a result, peak rainfall intensities were severely underestimated (factor 1.8–3.0 or 44 %–67 %). The most likely reason for this is the use of a fixed Z–R relationship when estimating rainfall rates (R) from reflectivity (Z), which fails to account for natural variations in raindrop size distribution with intensity. Based on our findings, the easiest way to mitigate the bias in times of heavy rain is to perform frequent (e.g., hourly) bias adjustments with the help of rain gauges, as demonstrated by the Dutch C-band product. An even more promising strategy that does not require any gauge adjustments is to estimate rainfall rates using a combination of reflectivity (Z) and differential phase shift (Kdp), as done in the Finnish OSAPOL product. Both approaches lead to approximately similar performances, with an average bias (at 10 min resolution) of about 30 % and a peak intensity bias of about 45 %.

Nowcasts (i.e., short-term forecasts from 5 minutes to 6 hours) of heavy rainfall are important for applications such as flash flood predictions. However, current precipitation nowcasting methods based on the extrapolation of radar echoes have a limited ability to predict the growth and decay of rainfall. While deep learning applications have recently shown improvement compared to extrapolation-based methods, they still struggle to correctly nowcast small-scale high-intensity rainfall. To address this issue, we present a novel model called the Lagrangian Convolutional Neural Network (L-CNN) that separates the growth and decay of rainfall from motion using the advection equation. In the model, differences between consecutive rain rate fields in Lagrangian coordinates are fed into a U-Net-based convolutional neural network (CNN), known as RainNet, that was trained with the Root-Mean-Squared-Error loss function. This results in a better representation of rainfall temporal evolution compared to the RainNet and the extrapolation-based LINDA model that were used as reference models. On Finnish weather radar data, L-CNN underestimates rainfall less than RainNet, demonstrated by greater POD (29% at 30 min at 1 mm h $^{-1}$ threshold) and smaller bias (98% at 15 min). The increased ETS values over LINDA for leadtimes under 15 min, with maximum increases of 7% (5 mm h $^{-1}$ threshold) and 10% (10 mm h $^{-1}$ ), show that the L-CNN represents the growth and decay of heavy rainfall more accurately than LINDA. This implies that nowcasting of heavy rainfall is improved when growth and decay are predicted using a deep learning model.

Short-range forecasts (nowcasts) of rainfall facilitate providing early warning of severe rainfall and flooding, which is particularly important in densely populated urban areas. Nowcasts are conventionally obtained by the extrapolation of radar echoes from a constant altitude plan position indicator (CAPPI) or lowest-angle plan position indicator (PPI). Lacking a model for growth or decay, this approach has a limited ability to forecast the summertime convective storms. In a previous attempt to address this shortcoming, called RadVil, the predicted surface rain rate is obtained from mass balance equations of vertically integrated liquid (VIL) retrieved from volumetric reflectivity measurements. Predicting the growth and decay has also been attempted by using an autoregressive (AR) model, which led to the development of spectral prognosis (S-PROG). A novel combination of these two methodologies, called ANVIL, is proposed. In this approach, the growth and decay of VIL are modeled by an autoregressive integrated (ARI) process. It is shown that the predictability of growth and decay is scale-dependent. Thus, the key idea of ANVIL is to decompose the VIL into multiple spatial scales and apply a separate ARI model to each scale. The operational feasibility of ANVIL is evaluated using the Next-Generation Radar (NEXRAD)/WSR-88D radar that covers the Dallas-Fort Worth metropolitan area. The evaluation is done using ten convective events in 2018 and 2019. Using several verification metrics, it is shown that ANVIL has up to 25% improved skill compared to conventional nowcasting techniques. The improvement is consistent for a wide range of spatial scales (1-11 km) and rain rate thresholds (5-20 mmh ^-1 ).

Precipitation nowcasting (forecasting locally for 0–6 h) serves both public security and industries, facilitating the mitigation of losses incurred due to, e.g., flash floods and is usually done by predicting weather radar echoes, which provide better performance than numerical weather prediction (NWP) at that scale. Probabilistic nowcasts are especially useful as they provide a desirable framework for operational decision-making. Many extrapolation-based statistical nowcasting methods exist, but they all suffer from a limited ability to capture the nonlinear growth and decay of precipitation, leading to a recent paradigm shift towards deep-learning methods which are more capable of representing these patterns. Despite its potential advantages, the application of deep learning in probabilistic nowcasting has only recently started to be explored. Here we develop a novel probabilistic precipitation nowcasting method, based on Bayesian neural networks with variational inference and the U-Net architecture, named DEUCE. The method estimates the total predictive uncertainty in the precipitation by combining estimates of the epistemic (knowledge-related and reducible) and heteroscedastic aleatoric (data-dependent and irreducible) uncertainties, using them to produce an ensemble of development scenarios for the following 60 min. DEUCE is trained and verified using Finnish Meteorological Institute radar composites compared to established classical models. Our model is found to produce both skillful and reliable probabilistic nowcasts based on various evaluation criteria. It improves the receiver operating characteristic (ROC) area under the curve scores 1 %–5 % over STEPS and LINDA-P baselines and comes close to the best-performer STEPS on a continuous ranked probability score (CRPS) metric. The reliability of DEUCE is demonstrated with, e.g., having the lowest expected calibration error at 20 and 25 dBZ reflectivity thresholds and coming second at 35 dBZ. On the other hand, the deterministic performance of ensemble means is found to be worse than that of extrapolation and LINDA-D baselines. Last, the composition of the predictive uncertainty is analyzed and described, with the conclusion that aleatoric uncertainty is more significant and informative than epistemic uncertainty in the DEUCE model.