One of the most crucial applications of radar-based precipitation nowcasting systems is the short-term forecast of extreme rainfall events such as flash floods and severe thunderstorms. While deep learning nowcasting models have recently shown to provide better overall skill than traditional echo extrapolation models, they suffer from conditional bias, sometimes reporting lower skill on extreme rain rates compared to Lagrangian persistence, due to excessive prediction smoothing. This work presents a novel method to improve deep learning prediction skills in particular for extreme rainfall regimes. The solution is based on model stacking, where a convolutional neural network is trained to combine an ensemble of deep learning models with orographic features, doubling the prediction skills with respect to the ensemble members and their average on extreme rain rates, and outperforming them on all rain regimes. The proposed architecture was applied on the recently released TAASRAD19 radar dataset: the initial ensemble was built by training four models with the same TrajGRU architecture over different rainfall thresholds on the first six years of the dataset, while the following three years of data were used for the stacked model. The stacked model can reach the same skill of Lagrangian persistence on extreme rain rates while retaining superior performance on lower rain regimes.

Downscaling based on deep learning (DL) is a key application in Earth system modeling, enabling the generation of high-resolution fields from coarse numerical simulations at reduced computational costs compared to traditional regional models. Additionally, generative DL models can potentially provide uncertainty quantification through ensemble-like scenario generation, a task prohibitive for conventional numerical approaches. In this study, we apply a latent diffusion model (LDM) to demonstrate that recent advancements in generative modeling enable DL to deliver results comparable to those of numerical dynamical models, given the same input data, preserving the realism of fine-scale features and flow characteristics at reduced computational costs. We apply our LDM to downscale ERA5 data over Italy up to a resolution of 2 km. The high-resolution target data consist of 2 m temperature and 10 m horizontal wind components from a dynamical downscaling performed with COSMO-CLM. A selection of predictors from ERA5 is used as input, and a residual approach against a reference U-Net is leveraged in applying the LDM. The performance of the generative LDM is compared with reference baselines of increasing complexity: a quadratic interpolation of ERA5, a U-Net, and a generative adversarial network (GAN) built on the same reference U-Net. Results highlight the improvements introduced by the LDM architecture combined with the residual approach, outperforming all the baselines in terms of spatial error, frequency distributions, and power spectra. These findings point out the potential of LDMs as cost-effective, robust alternatives for downscaling applications (e.g., downscaling of climate projections), where computational resources are limited but high-resolution data are critical.

Citizen science has shown great potential for bringing large groups of people closer to science, thanks in part to cooperation with universities and research centers. In this context, amateur weather networks played a major role in the last few decades thanks to a constant growth in technology. An example is given by the Meteonetwork association, born in 2002, and mainly composed by atmospheric science enthusiasts, who built up in time a huge weather station network in Italy. In recent years, they have enlarged their horizons over Europe, displaying real time observations and daily maps coming from both personal weather stations and official networks. This study described how Meteonetwork has set up an open crowdsourced weather data system, how data are validated, and which products are generated and freely accessible for scientists and stakeholders for their own purposes. Two concrete use cases were described as examples: the Weatherness Project, which selects a subnet of Meteonetwork data for biometeorological and health purposes and the data assimilation process implemented to improve the initial conditions into the WRF meteorological model for daily weather forecasts.

We introduce , a high-resolution radar reflectivity dataset collected by the Civil Protection weather radar of the Trentino South Tyrol Region, in the Italian Alps. The dataset includes 894,916 timesteps of precipitation from more than 9 years of data, offering a novel resource to develop and benchmark analog ensemble models and machine learning solutions for precipitation nowcasting. Data are expressed as 2D images, considering the maximum reflectivity on the vertical section at 5 min sampling rate, covering an area of 240 km of diameter at 500 m horizontal resolution. The distribution also includes a curated set of 1,732 sequences, for a total of 362,233 radar images, labeled with precipitation type tags assigned by expert meteorologists. We validate as a benchmark for nowcasting methods by introducing a deep learning model to forecast reflectivity, and a procedure based on the UMAP dimensionality reduction algorithm for interactive exploration. Software methods for data pre-processing, model training and inference, and a pre-trained model are publicly available on GitHub (https://github.com/MPBA/TAASRAD19) for study replication and reproducibility.

The Journal of Applied Remote Sensing (JARS) is an online journal that optimizes the communication of concepts, information, and progress within the remote sensing community to improve the societal benefit for monitoring and management of natural disasters, weather forecasting, agricultural and urban land-use planning, environmental quality monitoring, ecological restoration, and numerous other commercial and scientific applications.

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.