Precipitation nowcasting, the forecasting of short-term localized precipitation, is a crucial component of weather forecasting, particularly given the increasing frequency of extreme weather events due to climate change. Building on the critical role of precipitation nowcasting, this study focuses on enhancing deep learning models by optimizing hyper-parameters on well-established models such as convolutional long short-term memory (ConvLSTM) and small attention U-Net (SmaAt-UNet) rather than proposing new architectures. Our dataset consists of 10 years of precipitation maps at 1 3 1 km2, 5-min resolution from M & eacute;t & eacute;o-France's Mosaic product covering the Paris region, providing a robust testbed for training and evaluating models. To better characterize the performance of models over different types of rain events, clustering was carried out to partition the events into four different groups ranging from light rain to intense rain. Four main changes were explored to improve the training process. We show that increasing the temporal context up to 40 min signifi-cantly improves predictive accuracy for high-intensity rainfall events. Second, the introduction of wider spatial observation windows mitigates edge effects during training, enhancing spatial consistency and intensity estimation. Third, extending the output horizon to 90 min to predict precipitation at 60 min offered substantial gains for recurrent models like ConvLSTM, highlighting their capacity to better capture long-term dynamics. Last, the influence of convolution kernel size was studied to maximize the performance of the models. These changes collectively lead to significant forecast improvements especially for extreme precipitation events with metric values increasing compared to baseline configurations. We also show that the proposed model performs better when evaluated at coarser scales, relevant for hydrological applications.

Abstract Microwave brightness temperatures (TB) measured by satellites have played a significant role in observing various water cycle components. Until now, the measurement of TB has been used to retrieve many geophysical variables with Bayesian or, more recently, supervised deep learning approaches. These retrieval methods rely on the availability and the quality of the target variable and could be affected by changes in the relationship between TB and the target variable under climate change. This study aims to show that an unsupervised semantic image segmentation model of TB images offers meaningful segments that represent geophysical variables by interpreting and analyzing the resulting segmentation. The unsupervised segmentation methods explored include fully convolutional networks and pixelwise approaches with k -means. The main challenge tackled here is evaluating the model’s performance without any ground-truth labeling. The proposed evaluation requires external reference variables, including precipitation and sea surface temperatures (SSTs). The segmentation classes that represent these two geophysical variables are first compared to reference precipitation and SST data using segmentation metrics. The second step is to investigate the behavior of these classes. Results show that the precipitation classes capture the pattern of monsoon precipitation, while the ocean classes show the seasonality of SST. With an unsupervised approach, it is essential to verify the properties of the obtained segments extensively. The segmentation results in this study show that unsupervised learning is a key approach to extracting decades-long information from TB observations. Significance Statement Unsupervised learning has the potential to extract information from the climatic record in a large database of microwave brightness temperatures. However, the lack of labeled data means that the resulting clusters or representations require careful interpretation, as they are not inherently associated with known geophysical classes. While certain architectural and algorithmic choices may introduce useful constraints, validating the results against geophysical knowledge remains essential to ensure their physical relevance. This study aims to provide a rigorous evaluation process and a geophysical interpretation for the obtained segmentation. This is an important step before using these unsupervised segments to study climate change in the decades-long brightness temperature observations.

Weather radar mosaics provide high-resolution precipitation estimates but are often degraded by beam-blockage artefacts caused by terrain or infrastructure. These artefacts introduce systematic attenuation of radar reflectivity and distort precipitation patterns, particularly in complex topographic environments. Traditional correction approaches typically rely on geometric modeling or multi-radar fusion and often treat detection and correction as separate problems. In this work, we propose a guided conditional diffusion framework for correcting radar beam-blockage artefacts. The model reconstructs the precipitation field through a conditional reverse diffusion process guided by the corrupted radar observation and its associated quality map. During training, beam-blockage patterns are simulated on clean radar sectors to generate corrupted observations while preserving the original precipitation fields as references. This enables the model to learn both the spatial characteristics of blockage and the structure of precipitation fields. The method is evaluated on real radar mosaics containing both simulated and naturally occurring beam-blockage effects. Quantitative results show that the proposed approach reduces reconstruction errors within blocked regions while preserving precipitation values in unaffected areas. Qualitative analysis further demonstrates the recovery of coherent precipitation structures in degraded regions.

Precipitation nowcasting plays an essential role in operational weather forecasting services. Sudden precipitation events have significant socio-economic impacts, including natural disasters like flash floods. This challenge is becoming increasingly critical as climate change alters weather patterns and the frequency of extreme weather events continues to increase.Over the last decade, radar observations, offering high temporal and spatial resolution, have facilitated the development of machine learning methods for precipitation nowcasting. Once trained, these methods are well suited to processing large datasets with low latency, especially in a real-time context. Recent advances in the field of nowcasting have focused on optimizing model architectures, improving loss functions for imbalanced data, and integrating multivariate inputs, including radar and satellite observations.This study explores some critical hyperparameters, such as temporal context length, edge effect during training, influence of the output horizons prediction, and convolution kernel size. To do this, we investigate the performance of several models, including both machine learning approaches from different families, in particular SmaAt-Unet, ConvLSTM , and DGMR (trained on UK rains) , as well as non-machine learning methods such as STEPS. An eleven years consistent radar precipitation dataset covering the Paris region was set up from Météo-France mosaic. Nine years were used for training machine learning models, and two years were reserved to evaluate the models’ performances. To assess the model in different weather conditions, the data set is divided into four groups with distinct characteristics corresponding to various meteorological phenomena. To ensure consistent evaluation, we evaluated the models on the same two-year test dataset, focusing on three criteria, namely: spatial consistency (Pearson correlation coefficient), location accuracy (CSI), and precipitation intensity (MSE).Our analysis reveals that machine learning models consistently outperform traditional optical flow methods, with notable variations in performance across timescales and rainfall intensities. We also highlight that performance is nearly identical for all models in the presence of stratiform rain, while there are substantial differences in the convective rain group. Additionally, we show that for deep learning models, considering edge effects during training prevents the propagation of inevitable errors and helps avoid the appearance of ghost rain cells at the edges of the map. Furthermore, we show that the size of the kernels of the first layers plays an important role and must be large enough to allow correlation between distant pixels.Finally, our study provides guidelines for the development of precipitation nowcasting models.

Retrieval of rain from Passive Microwave radiometers data has been a challenge ever since the launch of the first Defense Meteorological Satellite Program in the late 70s. Enormous progress has been made since the launch of the Tropical Rainfall Measuring Mission (TRMM) in 1997 but until recently the data were processed pixel-by-pixel or taking a few neighboring pixels into account. Deep learning has obtained remarkable improvement in the computer vision field, and offers a whole new way to tackle the rain retrieval problem. The Global Precipitation Measurement (GPM) Core satellite carries similarly to TRMM, a passive microwave radiometer and a radar that share part of their swath. The brightness temperatures measured in the 37 and 89 GHz channels are used like the RGB components of a regular image while rain rate from Dual Frequency radar provides the surface rain. A U-net is then trained on these data to develop a retrieval algorithm: Deep-learning RAIN (DRAIN). With only four brightness temperatures as an input and no other a priori information, DRAIN is offering similar or slightly better performances than GPROF, the GPM official algorithm, in most situations. These performances are assumed to be due to the fact that DRAIN works on an image basis instead of the classical pixel-by-pixel basis.

This study proposes using a data-driven statistical model to freeze errors due to differences in environmental forcing when evaluating surface turbulent heat fluxes from weather and climate numerical models with observations. It takes advantage of continuous acquisition over approximately 10 years of near-surface sensible and latent heat fluxes (H and LE respectively) together with ancillary parameters at the M & eacute;t & eacute;opole flux station, a supersite of the Aerosol, Clouds and Trace Gases Research Infrastructure in France (ACTRIS-FR), located in Toulouse. The statistical model consists of several multi-layer perceptrons (MLPs) with the same architecture. A total of 13 variables characterizing environmental forcing in the surface layer on an hourly timescale are used as input parameters to estimate the observed H and LE simultaneously. The MLPs are trained using 5-year observational data under a 5-fold cross-validation. The remaining data are used to test the estimates under unknown conditions. The performance of the statistical model ranges within the state-of-the-art surface parameterization schemes on hourly and seasonal timescales. It also has a good generalization ability, but it hardly estimates negative H and large LE. A case study is conducted with data from a regional climate simulation. The statistical model is used to evaluate the simulated fluxes in the simulated environment to better examine the flaws of their numerical formulation throughout the simulation. Comparison of simulated fluxes with observed and MLP-based fluxes shows different results. According to MLP-based fluxes in the simulated environment, the land surface scheme of this climate model tends to underestimate large sensible heat flux. Thus, it incorrectly partitions between surface heating and evaporation during the late summer. Our innovative method provides insight into different techniques for evaluating simulated near-surface turbulent heat fluxes when a long period of comprehensive observations is available. It can usefully support ongoing efforts to improve surface parameterization schemes.

Earth–satellite microwave links such as TV-SAT can help for rainfall monitoring and could be a complement or an alternative to ground-based weather radars, rain gauges or Earth observation satellites. Rain-induced attenuation which is harmful for telecommunication is exploited here as an opportunistic way to estimate rain rate along the link path. This technique provides rain measurements at a fine temporal resolution (a few tens of seconds) and with a spatial resolution of a few kilometres, which is a good compromise for human activities such as civil security (watershed monitoring, flash flood), agriculture or transport. The advantages of this technique include the low cost of the equipment used, as well as the cost of on-site maintenance. However, the measured attenuation does not directly provide rain intensity, requiring the estimation of additional parameters. These include the contribution of natural radiation from the atmosphere. In this paper, we detail a theoretical framework allowing us to estimate rainfall from the measurements of a low-cost sensor operating simultaneously over two parts of the Ku frequency band. This framework is assessed in a densely instrumented area in the south of France, where very good results are obtained when compared to rain gauge measurements, in terms of both overall rain accumulation and rainfall rate distribution. Then we apply this dual-channel method in Côte d'Ivoire, in the metropolitan area of Abidjan, where such an approach is very promising. It is shown that this technique when compared to rain gauge measurements gives far better results than a naive single-channel approach neglecting the natural radiation of atmosphere but that significant errors remain in rainfall assessment, leading to a persistent underestimation of rain accumulation. Finally we discuss various effects that could lead to this remaining underestimation, opening the door for further studies.

In this paper, we investigate the performance of several models, including both machine learning approaches, in particular SmaAt-Unet, ConvLSTM and DGMR, as well as non-machine learning methods such as Lagrangian persistence, S-PROG, and STEPS, using a consistent radar precipitation dataset covering the Paris region over a period of two years. In our analysis, the dataset is partitioned into four groups with distinct characteristics, shedding light on the diverse behaviors of these models. Our findings reveal that machine learning models outperform traditional optical flow models, with noticeable variations in performance across different time scales and rain intensities. This improvement by machine learning can be particularly observed in light rain events. On the other hand, when considering heavy rain events, the results are closer from a model to another, especially for horizons exceeding 15 minutes.

With a current population of over 12 million (Eurostat 2021), the Ile-de-France region is particularly vulnerable to climate change, air-quality and water scarcity. Socio-economic development is leading to a degradation of the urban environment as a result of all human activities and the artificialization of land. Mitigating these adverse effects requires the implementation of specific adaptation policies to make the region more resilient to extreme events (such as heat waves, floods, droughts, pollution episods), but also to better manage water resources. To achieve this, we need to better characterize and understand the dynamics of precipitation in this region.The presence of urban areas modifies the interactions between the surface and the atmosphere through the modifications of energy and water budgets inducing the urban heat island effect, an increased surface roughness, and anthropogenic aerosols emissions. Several studies investigated the impact of these urban areas on precipitation with various conclusions due to lots of disparity (methods, datasets, cities' characteristics, climate) between the studies (e.g Lalonde et al., 2023; Liu and Niyogi, 2019). By using a unique methodology and a radar product covering the whole Continental USA at 4km resolution or the whole Europe at 2km resolution over tens of years, Lalonde et al. (2024) still highlighted a large diversity of urban impacts between cities, concluding on the necessity of specific analysis for each city.In this study, we use the recently available long-term 1 km/hourly radar product COMEPHORE (Tabary et al., 2012) which provide precipitation rates over France since 1997 to characterize the spatial variability of precipitation over Ile de France with a special focus on the impact of urban area. We complement this dataset with vertical profiles of reflectivity and vertical velocities provided by one radar located in a southwest suburb of Paris (most of the time upwind from Paris city center) and another one located in the city center to detect potential differences in microphysical properties of precipitation between urban and upwind environments.In the next step, we aim to assess the role of aerosols and urban form in this variability.

<p>Due to climate change, understanding the changes in the water cycle has become a pressing issue. It is increasingly important to study prolonged periods of intense precipitation or dry spells to better manage water supply, infrastructure and agriculture. However, obtaining fine-scale precipitation data is challenging due to the intermittent nature of rain in time and space. Ground-based instruments could have mismatches between different regions due to spatial distribution, calibration, and complex topography. On the other hand, space-borne observations have uncertainties in their retrieval algorithms. This study proposes to deal directly with microwave images from space remote sensing, as this type of data makes it possible to study the evolution of the atmospheric water cycle on a global scale and with a temporal coverage of several decades by avoiding the uncertainties from retrieval methods. In recent years, convolutional neural networks have shown promising capabilities in identifying cyclones and weather fronts in large labelled climate datasets. However, these models required large labelled datasets for training and testing. The present study aims to test unsupervised segmentation approaches of microwave images, which are thus segmented into different classes. Instead of focusing only on one aspect, for example, precipitation, the obtained classes contain many physical properties. This is due to the fact that microwave brightness temperatures contain essential information relative to the atmospheric water cycle that can be used to derive many products such as rain intensity, water vapour, cloud fraction, and sea surface temperature. The unsupervised segmentation model consists of blocks of fully convolutional networks serving as feature extractors. Without labels, pseudo-targets from the feature extractors are used to train the model. The performance of the model in terms of intra-class and inter-class distances is compared with those of simpler models such as Kmeans. A major challenge in the unsupervised approach is validating and interpreting the resulting classes. Most of the obtained cluster patterns provide geographically coherent regions whose mode of variability of geophysical quantities can be highlighted. The presented study will then explore how the different classes computed by the unsupervised methods can be labelled and how the properties of the said classes change through time and space.</p>

We focus in this paper on the use of satellite attenuation in Ku-frequency band for measuring rainfall, using low-cost sensors allowing to measure only the total received power. This implies that the measured powers are not only composed of the satellite signal, but also include a radiometric contribution, which is the atmospheric radiation and the receiving system noise. In case of heavy rainfall, while the satellite signal weakens, the atmospheric radiation becomes more important and neglecting it can lead to a significant underestimation of rainfall. In this paper, we introduce the theoretical framework of these measurements, and derive a simple method for correcting this underestimation, based on the measurements of the received power on two channels (different frequencies and / or polarizations), one with strong satellite signal and the other one with no or weak satellite signal. Then we do a preliminary assessment of this method in Abidjan, Ivory Coast, where such a sensor has been deployed during 9 months close to a rain gauge. These preliminary results show that this method seems to lead to rainfall assessments better than using a single-channel approach, but only after a calibration phase to assess the difference in the gain of the system between the two channels we use.

The surface turbulent fluxes, namely sensible and latent heat fluxes, are keys factors governing the boundary layer processes. Therefore, their correct representation in numerical models is crucial for accurate weather forecasts and climate projections. However, the formulation of these fluxes in such models is the second source of uncertainty, leading to incorrect surface-atmosphere interactions in the simulations. Model evaluation is essential to draw development perspectives. Existing methods mostly consist of direct comparison between observed and modelled fluxes, blending other sources of errors such as incoherent grid-scale representation (soil and vegetation types) and inaccurate environmental forcing (radiative fluxes, temperature, moisture and wind speed). Thus, quantifying errors solely due to fluxes formulation is still challenging. This study, within the framework of the French project MOSAI (Model and Observation for Surface-Atmosphere Interactions), aims at proposing a novel evaluation approach to better identify the weakness of numerical models in surface turbulent fluxes formulation. The concept is to freeze the errors due to other sources by using a machine-learning model, notably a multi-layer perceptron, trained to estimate the fluxes from variables describing the conditions in the surface layer. Hourly data collected over several years at three operational instrumented sites of ACTRIS-France research infrastructure (SIRTA in Paris, Météo-pole in Toulouse and P2OA in Lannemezan), are used. Then, after adaptation to the outputs of numerical models involved in the MOSAI project (RegIPSL, LMDZ, AROME and ARPEGE), the trained-perception will be applied to assess their surface turbulent fluxes.

Raw data issued from meteorological radars are often corrupted by unwanted signals generically called clutter. Hills, tall buildings, atmospheric turbulence, birds, and insects yield patterns that complicate the interpretation of radar images and might add bias in the quantitative precipitation estimates (QPE). Clutter differs from precipitating echoes by both their polarimetric signatures and their particular shapes. This work deals with the removal of clutter. The core idea is to use a fully convolutional network (FCN) to take clutter shapes into account. For a straightforward approach by supervised learning, one would need radar images and their cleaned counterpart, which are not available. We developed a weakly supervised learning method that allows circumventing this issue. This method only requires auxiliary data from rain gauges. The additivity of the reflectivity allows presenting the learning problem in the form of a supervised restoration task with noisy targets. This problem is solved by successive training of a standard FCN (U-net). As ground truth is missing, standard metrics cannot be employed to make a global evaluation. Nevertheless, our method is quantitatively assessed on two clutter classes: ground clutter and interferences. A case study completes the evaluation. A qualitative comparison with the Météo-France algorithm is also performed on a couple of difficult cases.

Artificial intelligence has provided many breakthroughs in the field of computer vision. The fully convolutional networks U-Net in particular have provided very promising results in the problem of retrieving rain rates from space-borne observations, a challenge that has persisted over the past few decades. The rain intensity is estimated from the measurement of the brightness temperatures on different microwave channels. However, these channels are slightly different depending on the satellite. In the case where a retrieval model has been developed from a single satellite, it may be advantageous to use domain adaptation methods in order to make this model compatible with all the satellites of the constellation. In this proposed feasibility study, a Cycle Generative Adversarial Nets model is used for adapting one set of brightness temperature channels to another set. Results of a toy experiment show that this method is able to provide qualitatively good precipitation structure but still could be improved in terms of precision.

The road traffic is highly sensitive to weather conditions. Accumulation of snow on the road can cause important safety problems. But road conditions monitoring is as hard as critical: in mid-latitude countries, on the one hand, the spatial variability of snowfall is high and on the other hand, accurate characterization of snow accumulation mainly relies on costly sensors. In recent decades, webcams have become ubiquitous along the road network. The quality of these webcams is variable but even low-resolution images capture information about the extent and the thickness of the snow layer. Their images are also currently used by forecasters to refine their analysis. The automatic extraction of relevant meteorological information is hence very useful. Recently, generic and efficient computer vision methods have emerged. Their application to image-based weather estimation has become an attractive field of research. However, the scope of existing work is generally limited to high-resolution images from one or a few cameras. In this study, we show that for a moderate effort of labelling, recent Machine Learning approaches allow us to predict quantitative indices of the snow depth for a large variety of webcam settings and illumination. Our approach is based on two datasets. The smallest one contains about 2.000 images coming from ten webcams that were set up near sensors devoted to snow depth measurements. The largest one contains 20,000 images coming from 200 cameras of the AMOS dataset. Meteorological standard rules of human observation and the specifics of the webcams have been taken into account to manually label each image. These labels are not only about the thickness and the extent of the snow layer but also describe the precipitation (rain or snow, presence of streaks), the optical range and the foreground noise. Both datasets contain night images (45%) and at least 15% of images corrupted by foreground noise (filth, droplets, and snowflakes on the lens). The labels of the AMOS subset allowed us to train ranking models for snow depth and visibility using a multi-task setting. The models are then calibrated on the smallest dataset. We tested several versions, built from pre-trained CNNs (ResNet152, DenseNet161, and VGG16). Results are promising with up to 85% accuracy for comparison tasks, but a 10% decrease can be observed when the test webcams have not been used during the training phase. A case study based on a widespread snow event over the French territory will be presented. We will show the potential of our method through a comparison with operational model forecasts.

The road traffic is highly sensitive to weather conditions. Accumulation of snow on the road can cause important safety problems. But road conditions monitoring is as hard as critical: in mid-latitude countries, on the one hand, the spatial variability of snowfall is high and on the other hand, accurate characterization of snow accumulation mainly relies on costly sensors. In recent decades, webcams have become ubiquitous along the road network. The quality of these webcams is variable but even low-resolution images capture information about the extent and the thickness of the snow layer. Their images are also currently used by forecasters to refine their analysis. The automatic extraction of relevant meteorological information is hence very useful. Recently, generic and efficient computer vision methods have emerged. Their application to image-based weather estimation has become an attractive field of research. However, the scope of existing work is generally limited to high-resolution images from one or a few cameras. In this study, we show that for a moderate effort of labelling, recent Machine Learning approaches allow us to predict quantitative indices of the snow depth for a large variety of webcam settings and illumination. Our approach is based on two datasets. The smallest one contains about 2.000 images coming from ten webcams that were set up near sensors devoted to snow depth measurements. The largest one contains 20,000 images coming from 200 cameras of the AMOS dataset. Meteorological standard rules of human observation and the specifics of the webcams have been taken into account to manually label each image. These labels are not only about the thickness and the extent of the snow layer but also describe the precipitation (rain or snow, presence of streaks), the optical range and the foreground noise. Both datasets contain night images (45%) and at least 15% of images corrupted by foreground noise (filth, droplets, and snowflakes on the lens). The labels of the AMOS subset allowed us to train ranking models for snow depth and visibility using a multi-task setting. The models are then calibrated on the smallest dataset. We tested several versions, built from pre-trained CNNs (ResNet152, DenseNet161, and VGG16). Results are promising with up to 85% accuracy for comparison tasks, but a 10% decrease can be observed when the test webcams have not been used during the training phase. A case study based on a widespread snow event over the French territory will be presented. We will show the potential of our method through a comparison with operational model forecasts.

Since the 2000’s, cameras are considered as an interesting source of opportunistic meteorological data. This short study deals with the comparison of visibility, in a meteorological sense, between images. A new dataset has been built from publicly available webcam sequences. An original labeling process, based on a mergesort algorithm, allowed us to sort more than 400 webcam sequences with respect to visibility. Standard Convolutional Neural Networks have been trained to predict pairwise comparisons and tested on independent webcams that are colocalized with visibilimeters. Results on the comparison task are promising.

Despite a lot of progress over the last decades, rain retrieval from spaceborne measurement has been a challenge since the first launch of a passive microwave radiometers on one of the NOAA Defense Meteorological satellites in the 70s. Deep-learning and convolutional U-Nets might be able to offer a breakthrough on the topic because they do take into account the topology of both the rain field and the measured brightness temperatures. The present paper offers the very first results on the application of such artificial neural networks on the rain retrieval problem.