There is a growing interest in emerging opportunistic sensors for precipitation, motivated by the need to improve its quantitative estimates at the ground. The scope of this work is to present a preliminary assessment of the accuracy of commercial microwave link (CML) retrieved rainfall rates in Northern Italy. The CML product, obtained by the open-source RAINLINK software package, is evaluated on different scales (single link, 5 km×5 km grid, river basin) against the precipitation products operationally used at Arpae-SIMC, the regional weather service of Emilia-Romagna, in Northern Italy. The results of the 15 min single-link validation with nearby rain gauges show high variability, which can be caused by the complex physiography and precipitation patterns. Known sources of errors (e.g. the attenuation caused by the wetting of the antennas or random fluctuations in the baseline) are particularly hard to mitigate in these conditions without a specific calibration, which has not been implemented. However, hourly cumulated spatially interpolated CML rainfall maps, validated with respect to the established regional gauge-based reference, show similar performance (R² of 0.46 and coefficient of variation, CV, of 0.78) to adjusted radar-based precipitation gridded products and better performance than satellite-based ones. Performance improves when basin-scale total precipitation amounts are considered (R² of 0.83 and CV of 0.48). Avoiding regional-specific calibration therefore does not preclude the algorithm from working but has some limitations in probability of detection (POD) and accuracy. A widespread underestimation is evident at both the grid box scale (mean error of −0.26) and the basin scale (multiplicative bias of 0.7), while the number of false alarms is generally low and becomes even lower as link coverage increases. Also taking into account delays in the availability of the data (latency of 0.33 h for CML against 1 h for the adjusted radar and 24 h for the quality-controlled rain gauges), CML appears as a valuable data source in particular from a local operational framework perspective. Finally, results show complementary strengths for CMLs and radars, encouraging joint exploitation.

The assimilation of radar reflectivity data is of great interest for numerical weather prediction (NWP) models at the convective scale since these observations are very dense both in space and time and they are related to microphysical prognostic variables. So far, the assimilation of reflectivity in operational frameworks has been limited to derived products, like precipitation rates and humidity profiles. This approach is also adopted at ARPAE‐SIMC, the Hydro‐Meteo‐Climate Structure of the Regional Agency for Prevention, Environment and Energy of Emilia‐Romagna region in Italy, where latent heat nudging (LHN) is employed to assimilate precipitation rates in the COSMO‐2I model, the convection‐permitting version of the regional model developed by the COnsortium for Small‐scale MOdelling (COSMO) which is run operationally to provide high‐resolution forecasts over Italy. However, to fully exploit the information contained in these observations, the whole reflectivity volume should be directly assimilated. Nevertheless, despite several promising studies, its implementation in an operational framework has not yet been achieved. In this study, a set‐up designed to assimilate reflectivities operationally in the COSMO‐2I model through a local ensemble transform Kalman filter (LETKF) has been evaluated over 37 days in 2018 and performed 303 forecasts. The comparison with the current operational set‐up based on LHN reveals an average improvement in quantitative precipitation forecasts (QPFs) up to 7 hr, even if the impact on convective cases is much stronger than that observed in conditions of stratiform precipitation. Moreover, a small but positive impact is noticed for the RMSE of upper‐air variables, while the impact on bias is mixed. Mixed results are also obtained when considering surface variables. In the light of the results of this study, a pre‐operational parallel system in which COSMO‐2I analyses are generated replacing LHN by the direct assimilation of reflectivity volumes was implemented in April 2020.

The aim of this study is to provide an evaluation of the impact of two largely used data assimilation techniques, namely three‐ and four‐dimensional variational data assimilation systems (3D‐Var and 4D‐Var), on the forecasting of heavy precipitation events using the Weather Research and Forecasting (WRF) model. For this purpose, two flash flood events in central Italy are analysed. The first occurred on September 14, 2012 during an Intensive Observation Period of the Hydrological cycle in the Mediterranean experiment (HyMeX) campaign, while the other occurred on May 3, 2018. Radial velocity and reflectivity acquired by C‐band weather radars at Mt. Midia (central Italy) and San Pietro Capofiume (northern Italy), as well as conventional observations (SYNOP and TEMP), are assimilated into the WRF model to simulate these damaging flash flood events. In order to evaluate the impact of the 3D‐Var and 4D‐Var assimilation systems on the estimation of short‐term quantitative precipitation forecasts, several experiments are carried out using conventional observations with and without radar data. Rainfall evaluation is performed by means of point‐by‐point and filtering methodologies. The results point to a positive impact of the 4D‐Var technique compared to results without assimilation and with 3D‐Var experiments. More specifically, the 4D‐Var system produces an increase of up to 22% in terms of the Fractions Skill Score compared to 3D‐Var for the first flash flood event, while an increase of about 5% is achieved for the second event. The use of a warm start initialization results in a considerable reduction in the spin‐up time and a significant improvement in the rainfall forecast, suggesting that the initial precipitation spin‐up problem still occurs when using 4D‐Var.

Quantitative precipitation forecast (QPF) is still a challenge for numerical weather prediction (NWP), despite the continuous improvement of models and data assimilation systems. In this regard, the assimilation of radar reflectivity volumes should be beneficial, since the accuracy of analysis is the element that most affects short-term QPFs. Up to now, few attempts have been made to assimilate these observations in an operational set-up, due to the large amount of computational resources needed and due to several open issues, like the rise of imbalances in the analyses and the estimation of the observational error. In this work, we evaluate the impact of the assimilation of radar reflectivity volumes employing a local ensemble transform Kalman filter (LETKF), implemented for the convection-permitting model of the COnsortium for Small-scale MOdelling (COSMO). A 4-day test case on February 2017 is considered and the verification of QPFs is performed using the fractions skill score (FSS) and the SAL technique, an object-based method which allows one to decompose the error in precipitation fields in terms of structure ( S ), amplitude ( A ) and location ( L ). Results obtained assimilating both conventional data and radar reflectivity volumes are compared to those of the operational system of the Hydro-Meteo-Climate Service of the Emilia-Romagna Region (Arpae-SIMC), in which only conventional observations are employed and latent heat nudging (LHN) is applied using surface rainfall intensity (SRI) estimated from the Italian radar network data. The impact of assimilating reflectivity volumes using LETKF in combination or not with LHN is assessed. Furthermore, some sensitivity tests are performed to evaluate the effects of the length of the assimilation window and of the reflectivity observational error ( roe ). Moreover, balance issues are assessed in terms of kinetic energy spectra and providing some examples of how these affect prognostic fields. Results show that the assimilation of reflectivity volumes has a positive impact on QPF accuracy in the first few hours of forecast, both when it is combined with LHN or not. The improvement is further slightly enhanced when only observations collected close to the analysis time are assimilated, while the shortening of cycle length worsens QPF accuracy. Finally, the employment of too small a value of roe introduces imbalances into the analyses, resulting in a severe degradation of forecast accuracy, especially when very short assimilation cycles are used.

An analysis to evaluate the impact of multiple radar reflectivity data with a three-dimensional variational (3-D-Var) assimilation system on a heavy precipitation event is presented. The main goal is to build a regionally tuned numerical prediction model and a decision-support system for environmental civil protection services and demonstrate it in the central Italian regions, distinguishing which type of observations, conventional and not (or a combination of them), is more effective in improving the accuracy of the forecasted rainfall. In that respect, during the first special observation period (SOP1) of HyMeX (Hydrological cycle in the Mediterranean Experiment) campaign several intensive observing periods (IOPs) were launched and nine of which occurred in Italy. Among them, IOP4 is chosen for this study because of its low predictability regarding the exact location and amount of precipitation. This event hit central Italy on 14 September 2012 producing heavy precipitation and causing several cases of damage to buildings, infrastructure, and roads. Reflectivity data taken from three C-band Doppler radars running operationally during the event are assimilated using the 3-D-Var technique to improve high-resolution initial conditions. In order to evaluate the impact of the assimilation procedure at different horizontal resolutions and to assess the impact of assimilating reflectivity data from multiple radars, several experiments using the Weather Research and Forecasting (WRF) model are performed. Finally, traditional verification scores such as accuracy, equitable threat score, false alarm ratio, and frequency bias – interpreted by analysing their uncertainty through bootstrap confidence intervals (CIs) – are used to objectively compare the experiments, using rain gauge data as a benchmark.

This work introduces GPTCast, a generative deep learning method for ensemble nowcasting of radar-based precipitation, inspired by advancements in large language models (LLMs). We employ a generative pre-trained transformer (GPT) model as a forecaster to learn spatiotemporal precipitation dynamics using tokenized radar images. The tokenizer is based on a Variational Quantized Autoencoder (VQGAN) featuring a novel reconstruction loss tailored for the skewed distribution of precipitation that promotes faithful reconstruction of high rainfall rates. This approach produces realistic ensemble forecasts and provides probabilistic outputs with accurate uncertainty estimation. The core architecture operates deterministically during the forward pass; ensemble variability arises from sampling the categorical probability distribution predicted by the forecaster during inference, rather than requiring external random inputs such as noise injection common in other generative models. All forecast variability is thus learned solely from the data distribution. We train and test GPTCast using a 6-year radar dataset over the Emilia-Romagna region in northern Italy, showing superior results compared to state-of-the-art ensemble extrapolation methods.