This session invites contributions from developments in
• Convective-scale data assimilation techniques
• Use of machine learning in convective scale data assimilation
• Applications of machine learning to forecasting on convective scales
• Convective-scale model and observation uncertainty representation
• Ensembles and uncertainty quantification using machine learning
• Advances in convective-scale modeling and parameter estimation
• Assimilation of ground and space-based radar data
• Active and passive satellite data assimilation
• Assessment of the impact of convective-scale data assimilation on global and regional prediction
• Observation operators for remote sensing and data assimilation
• Observations at convective scales: new observing technologies and strategies
Orals: Mon, 4 May, 14:00–15:45 | Room 1.61/62
A convective-scale ensemble data assimilation (EDA) system has been developed in Taiwan to improve very short-term heavy rainfall prediction. The ground-based GNSS Zenith Total Delay (ZTD) data provides fast moisture information, which captures the precursor of convection initialization over complex terrain. Focusing on thunderstorm prediction in the Taipei Basin, previous studies have shown that assimilating ZTD data from the Central Weather Administration (CWA) operated stations provides effective moisture adjustment. Incorporating the surface 10-meter wind further exploits the benefit of ZTD assimilation in very short-term precipitation prediction. Including non-CWA-operated stations, there are more than 400 GNSS stations in Taiwan, forming a uniquely dense GNSS observation network. In addition to ZTD observation, the tropospheric gradient (TG) measurement provides spatial moisture variations in the low troposphere. Based on a severe afternoon thunderstorm on 24 June 2022 in the Taipei Basin, we conducted rapid-update data assimilation experiments to investigate the impact of the ground-based GNSS data. Data assimilation was performed over a three-hour period at 30-minute interval to predict a heavy rainfall event lasting two hours.
Compared to standard ZTD assimilation using CWA-operated stations, the assimilation of dense ZTD observations improves the moisture representation near the Taipei Basin, which is critical for the timing of convection initialization. For this case, TG observation reveals a strong moisture gradient into an inland river valley upstream of the Basin. Additional TG assimilation enhances moisture, facilitating the rapid convection development and the merging of the convection cells. Consequently, assimilating both dense ZTD and TG leads to significant improvements in the forecasted intensity and location of heavy rain, as well as the forecast performance at a longer lead time. Notably, the impact of TG assimilation is more pronounced when combined with dense ZTD data.
How to cite: Yang, S.-C., Chang, Y.-P., Yeh, T.-K., Zus, F., Thundathil, R., and Wickert, J.: Improving Afternoon Thunderstorm Prediction in Taiwan: Insights from Dense Ground-based GNSS Assimilation, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-5447, https://doi.org/10.5194/egusphere-egu26-5447, 2026.
Ground-based profilers provide continuous information on atmospheric boundary-layer (ABL) temperature and humidity, but zenith-only observations suffer from large representation errors in heterogeneous environments. This contribution explores the potential of scanning Microwave Radiometer (MWR) brightness temperatures (TBs) to better constrain ABL water vapor and to reduce representation error relevant for convection-permitting data assimilation. It especially aims to eventually evaluate the synergy of scanning MWR humidity observations with the already planned Differential Absorption Lidar (DIAL) network LIDIA by DWD.
As a proof of concept, radiosonde profiles are combined with co-located ground-based HATPRO MWR observations from recent field campaigns in Germany, including FESSTVaL (2021), Socles (2021–2022), and Vital I (2024). For each radiosonde launch, temporally matched MWR measurements are extracted for several viewing geometries. The evaluation by TB forward modeled from radiosondes gives first promising results.
The presentation highlights how low elevation azimuth scan TB information, especially combined with the upcoming LIDIA network can provide additional constraints on horizontal gradients and boundary layer humidity. The next steps are: assimilation experiments with data from the upcoming Vital II campaign (summer 2026), where combined zenith-pointing DIAL and scanning MWR observations will be assimilated into ICON-D2 to quantify their impact on short-range forecasts of humidity and convective initiation on convection-permitting resolution.
How to cite: Pschera, A., Toporov, M., Löhnert, U., and Schomburg, A.: A Proof of Concept for Boundary-Layer Moisture Data Assimilation Using Scanning Microwave Radiometer Observations, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-5510, https://doi.org/10.5194/egusphere-egu26-5510, 2026.
The INvestigation of Convective UpdraftS (INCUS), a NASA Earth Ventures Mission scheduled for launch in early 2027, will use three small satellites to deliver the first estimates of convective mass flux and its evolution within tropical and subtropical clouds. To assist the retrieval algorithm development, the INCUS team is producing a database of updrafts and their environments using high-resolution simulations of convective clouds conducted with the Weather Research and Forecasting Model (WRF) and the Regional Atmospheric Modeling System (RAMS). Specifically, case studies are selected in 15+ regions across the (sub)tropics. Each case study is simulated three times: once with RAMS (2-moment, bin-emulating microphysics) and twice with WRF (Morrison and Thompson aerosol-aware microphysics). Simulations utilize 3 nested grids, with the outermost grids having 1.6 km grid spacing and typically spanning well over 1,000 km in zonal and meridional directions. The innermost grids also have large areas (~230 km by ~230 km), with 100 m horizontal grid spacing, ~100 m vertical grid spacing, and 30 second output.
Using this output, the Tracking and Object-Based Analysis of Clouds (tobac) algorithm is run offline to quantify the 3D evolution of storm updrafts and link them to their associated anvils and environments. The tracked updrafts and their properties are directly used in INCUS algorithm development. The model output is also run through the Community Radiative Transfer Model (CRTM) and INCUS Passive Active Microwave Simulator (PAMS) to forward model geostationary IR brightness temperatures and INCUS-observed radar reflectivity, respectively. The simulations are systematically evaluated with available observations to ensure that the output is realistic and to identify gaps in the model database in terms of observed convective environments, updraft profiles, storm morphologies, and reflectivity profiles.
The goals of this talk are to: (1) give an overview of the INCUS LES database and its role in the INCUS mission; (2) provide statistics about the global nature of tropical and subtropical convection obtained using high-resolution models; and (3) showcase results from evaluation of the database against different observational datasets. The multi-model, high-resolution INCUS simulation database continues to grow as more simulations are completed and will be a useful resource for the community for understanding tropical and subtropical convective clouds.
How to cite: Singh, I., Bukowski, J., Marinescu, P., Dolan, B., Posselt, D., Schulte, R., Grant, L., Leung, G., Lewis, J., Prasanth, S., Rasmussen, K., Schumacher, C., Storer, R., and van den Heever, S. C.: High-resolution Numerical Simulations of Tropical and Subtropical Convection for the NASA INCUS Mission, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-5970, https://doi.org/10.5194/egusphere-egu26-5970, 2026.
The demand for accurate weather forecasts is increasing because of, for example, high-impact events becoming more frequent and more extreme. However, convection-permitting numerical weather prediction systems still lack precise initial conditions due to observational gaps. In order to improve the predictions, filling those gaps is of great importance.
Ground-based water vapor profiling networks could contribute to that using broadband differential absorption light detection and ranging systems (DIALs). In this work, the impact of different hypothetical configurations of such networks on the convection-permitting Icosahedral Non-hydrostatic (ICON) D2 model operationally used by the German Weather Service is assessed. For that, an ensemble sensitivity analysis (ESA) is employed. Using an ensemble-based sensitivity, the ESA method quantifies the change of variance in a forecast metric ensemble due to the assimilation of additional data. Compared to other observational impact assessment (OIA) methods, it benefits from not requiring forecast validation, actual measurements, or additional model runs.
First results indicate that for the default network configuration and two summer afternoon cases, the additional observations decrease the spread of an ensemble of the model on average by 5.5 to 7 percent (%) depending on the specific water vapor forecast metric used. The forecast metrics respond as expected to changes in the network parameters, that is, the spread decreases further for smaller instrumental errors and larger vertical ranges of the DIALs. Assimilating point measurements leads to a 70% smaller spread reduction relative to the one obtained for water vapor profiles up to a height of 1200 meters.
On the one hand, the results of these case studies confirm the expected benefit of a DIAL network, while also indicating what to focus on for further improvements. On the other hand, they demonstrate the applicability of the flexible and computationally cheap ESA method to this kind of evaluation. With the results being comparable to those obtained by other OIA methods, an assessment of how realistic the ESA results here are could be conducted. That could help to judge whether the method could and should be applied in OIA more broadly.
How to cite: Raabe, N., Toporov, M., Griewank, P., Weissmann, M., Matsunobu, T., and Löhnert, U.: Observational impact of ground-based water vapor profiling networks on convection-permitting numerical weather prediction – an ensemble sensitivity analysis case study, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-7077, https://doi.org/10.5194/egusphere-egu26-7077, 2026.
The transition from Meteosat Second Generation (MSG) to Meteosat Third Generation (MTG) represents a major step forward in European geostationary satellite observing capability. This work presents preparatory and preliminary results from extending the established all-sky radiance assimilation framework, originally developed for MSG SEVIRI to MTG FCI. The study emphasizes methodological continuity to ensure a seamless evolution of satellite data assimilation practices within convective-scale numerical weather prediction.
Earlier all-sky assimilation experiments with SEVIRI water vapour channels established a robust framework for handling cloud-affected radiances through dynamically adaptive, cloud-dependent observation error modelling. Here, the same principles are applied to FCI water vapour channels, enabling a consistent assessment of how these observations interact with existing assimilation systems. FCI’s enhanced spatial resolution, improved radiometric accuracy, and expanded spectral sampling promise greater information content, but also introduce increased representativeness and cloud-related uncertainties, reinforcing the value of adaptive error characterisation.
Preliminary 3D-EnVar experiments within the pre-operational regional high resolution ensemble prediction system of GeoSphere Austria compare the all-sky assimilation approaches for both SEVIRI and FCI radiances. Results show that the cloud-dependent observation error model previously validated with SEVIRI yields a neutral to positive impact on the short-range forecast. This demonstrates a structured pathway toward the operational exploitation of FCI observations, supporting improved short-range forecasts of high-impact weather. Progress on assimilation of visible channels from SEVIRI and FCI are also foreseen.
How to cite: Neduncheran, A., Meier, F., Scheffknecht, P., Wittmann, C., Weissmann, M., and Griewank, P.: From MSG SEVIRI to MTG FCI: Extending All-Sky IR radiance assimilation in the regional high-resolution ensemble prediction system of GeoSphere Austria , EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-9329, https://doi.org/10.5194/egusphere-egu26-9329, 2026.
In recent years, with the development of computing power, kilometer-scale resolution has become possible in regional numerical weather prediction (NWP) and climate simulation. While the refined numerical mesh allows for a more detailed representation of weather and climate, it also moves atmospheric modeling into gray zones, where the parameterization of turbulence and convection becomes a challenge in NWP. The newly developed machine learning (ML) methods would be a better choice to address this challenge. Previous ML weather prediction models primarily focus on global mesoscale forecasting. This study develops a purely data-driven three-dimensional Fourier neural operator (FNO) model for simulating the idealized convective boundary layer (CBL) at 800-m grid spacing, which is a resolution in the gray zone. The filtered large-eddy simulation (LES) data of the CBL are used for training the FNO models. The FNO models can accurately predict the instantaneous spatial structures and flow statistics of the boundary layer. The structures of vertical velocity near the surface predicted by the FNO models are consistent with those of the filtered LES, overcoming the issue of overly large cell structures predicted by traditional numerical simulations. The FNO models perform better than the gray-zone CM1 simulations in predicting profiles of flow statistics. Furthermore, the temperature and velocity spectra predicted by the FNO models are close to the results of filtered LES. The FNO models, trained using data for a few surface heat flux values Qs from 0.14 to 0.26 K m s-1, demonstrate certain generalization capabilities for other Qs within and out of this range. Overall, the FNO model is a promising ML method for fast and accurate weather prediction in the gray zone.
How to cite: Luo, T., Shi, X., Li, Z., Wang, J., Chan, P. W., and Wu, N.: Prediction of Convective Boundary Layer in the Gray Zone Using Fourier Neural Operator, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-10136, https://doi.org/10.5194/egusphere-egu26-10136, 2026.
To improve the forecast quality of numerical weather prediction (NWP), the German Meteorological Service (Deutscher Wetterdienst, DWD) has initiated a project aimed at assessing data quality and assimilation of observations from ground-based remote sensing instruments that have not yet been exploited operationally.
The objective of this initiative is to fill the observational gap in the atmospheric boundary layer, especially with respect to short time scales, by providing continuous, high-temporal-resolution profiles of thermodynamic variables, wind, and cloud properties. These observations are expected to be especially beneficial for weather forecasting applications. The DWD is evaluating various remote sensing systems with regard to the continuous data supply, their operational use, and their impact on NWP.
In this contribution, we present first results of the assimilation of two ground-based remote sensing instruments into the kilometer-scale ensemble data assimilation system (KENDA): water vapour mixing ratio from a Differential Absorption Lidar (DIAL) and radar reflectivity from a cloud radar. For the integration of the DIAL observations into the data assimilation code environment, only small adjustments were necessary. In contrast, the cloud radar data required an adaptation of the complex forward operator EMVORADO (Efficient Modular Volume scan Radar Operator), which was originally developed and previously used only for precipitation radars.
In an initial step, single observation data assimilation experiments and their observation minus first guess statistics have been shown to produce promising results. To assess the impact in an operational setting, dedicated data assimilation experiments were conducted and compared to reference experiments without these additional observations. Based on the successful data assimilation cycling experiments, first forecast experiments including the DIAL have been performed. Current results indicate a neutral to positive impact on humidity, temperature, and wind forecasts. The impact of cloud radar data in such experiments is currently under investigation by testing different settings.
Our findings suggest that ground-based remote sensing data can provide valuable additional information for convective-scale data assimilation, and justify more extensive impact studies in the context of NWP.
How to cite: Pruschke, J., Schomburg, A., Mendrok, J., Stephan, K., Görsdorf, U., Löffler, M., Knist, C., and Schraff, C.: First Steps Towards Data Assimilation of Differential Absorption Lidar and Cloud Radar Data, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-11124, https://doi.org/10.5194/egusphere-egu26-11124, 2026.
While understanding convective processes is critical for prediction of severe weather and cloud properties, observing convective mass flux at convective-scales is difficult using traditional techniques. The NASA INvestigation of Convective UpdraftS (INCUS) mission will quantify convective mass flux from a convoy of three Ka-band radars using a unique time-differencing approach. In order to develop the necessary observational approaches to calibrate and validate the INCUS products, preliminary testing was performed using data collected under the umbrella of the Testing of INCUS Measurements Experiment - Suborbital preLaunch Investigation of Convective Evolution (TIME-SLICE) campaigns. The focus of TIME-SLICE is to leverage existing and low-cost instruments to experiment with rapid sampling of ground assets to derive vertical motions and reflectivity time differences.
TIME-SLICE Colorado (TIME-SLICE-CO) was conducted in the northern Front Range of Colorado during the summer of 2024. During the two months of intensive operations, the Colorado State University CHIVO C-band radar scanned convection using the Multisensor Agile Adaptive Sampling (MAAS) framework, which uses ancillary information to select and follow targets of interest and then scan with RHIs with 30 s frequency. Additionally, a site with multiple frequencies of vertically pointing radars and ground instruments collected continuous data. Lessons learned from this preliminary testing highlighted the need for both priority sampling over the ground site by MAAS as well as larger coverage of convection by multiple scanning radars to broaden the coverage of vertical velocity retrievals and account for horizontal advection.
Building on the lessons from TIME-SLICE-CO, a follow-on campaign was held in North Alabama during summer 2025. TIME-SLICE Alabama (TIME-SLICE-AL) extends the objectives of TIME-SLICE to the variety of convection in the Alabama region while employing novel phased-array radar (PAR) sampling in concert with a high concentration of pre-existing instruments. In TIME-SLICE-AL, two X-band PARs, the Stony Brook University (SBU) SKYLER-2 radar and the SBU ROARS radar, were positioned alongside The University of Alabama in Huntsville (UAH) C-band ARMOR radar, operating in a rapid (30-second) RHI mode, all coordinated within the MAAS to autonomously sample a large number of convective clouds. Additionally, novel 3D observations of updrafts were collected by tilting the SKYLER PAR vertically. These innovative new sampling techniques and combined observations allow for rapid quantification of radar reflectivity time differences coincident with Doppler-derived vertical motion estimates, and the development of a large, rich case database. Further, existing assets deployed at the UAH Severe Weather Institute and Radar & Lightning Laboratories site, including a disdrometer and the new Ka Profiling Radar (KaPR), and at the nearby US Department of Energy ARM Mobile Facility 3 in the Bankhead National Forest, provide additional context to the primary radar observations. In this presentation, we will provide an overview of the TIME-SLICE-CO and -AL campaigns, present several of the cases captured, highlight some of the early novel science results, and apply the knowledge gained to the validation efforts for the INCUS mission.
How to cite: Dolan, B., Freeman, S., Kollias, P., Rasmussen, K., Gatlin, P., Luke, E., Chandrasekar, V., Amiot, C., Ebbert, E., Knupp, K., Kwinta, C., Pangle, P., Petersen, W., Schumacher, C., Tanelli, S., Treserras, B., van den Heever, S., Williams, C., and Wolff, D.: TIME-SLICE: Developing observation techniques for estimating convective mass flux through rapid, adaptive sampling , EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-14526, https://doi.org/10.5194/egusphere-egu26-14526, 2026.
NASA's upcoming INCUS mission will observe tropical deep convection at unprecedented spatiotemporal resolution (∼3 km horizontal sampling at intervals of 30 s, 90 s, and 120 s). INCUS will observe a wide variety of deep convective environments and tropical storms at different stages of their life cycle, producing many distinct convective outcomes. How do we distinguish these outcomes? Which differences are associated with genuinely different local environmental states, and which aspects vary even when the local environment is effectively the same?
We use Kernel Flows, a nonlinear machine learning estimator, to separate aspects of local deep convection that are constrained by their environmental state from those that are not. We first define the environmental state as a vector X formed from variables always present in the troposphere describing background thermodynamic and kinematic conditions, independent of whether convection is active (temperature, pressure, water vapor, horizontal wind fields). We define convective variables as vertical velocity, condensed water mass, and convective mass flux, which arise only where convection is present and correspond to quantities INCUS algorithms are designed to retrieve.
The analysis uses a database of high-resolution simulations (100 m horizontal grid spacing) across tropical and subtropical regions in anticipation of INCUS. From these simulations, we extract 25 km × 25 km neighborhoods containing a mixture of deep moist convective updrafts and surrounding non-updrafts. Within each neighborhood, we coarsen the environmental and convective variables to approximately 3 km resolution to match anticipated INCUS radar resolution. We represent the environmental state using principal components and define 19 scalar descriptors (Yi) to characterize various aspects of convection.
Using Kernel Flows with no assumptions on Gaussian conditional distributions, we learn the functional relationship between each convective descriptor (Yi) and environmental state (X) independently. We quantify how much variability in each descriptor is associated with environmental state differences by comparing the estimator's error variance to the prior variance of Yi.
We find that aggregate convective descriptors exhibit variability almost entirely associated with environmental state differences. Total vertical velocity and total convective mass flux over the neighborhood, along with neighborhood-mean column maxima of these quantities, achieve R² ≥ 0.97 with at least 84% of standard deviation explained by the environmental state. These convective descriptors act as invariants of the local environment: differences in these metrics reflect environmental state (X) differences, rather than natural variability within an equivalent environmental state. Conversely, convective descriptors emphasizing horizontal and vertical organization of updrafts, such as how total updraft area is divided among horizontally contiguous clusters and maximum heights of vertical velocity and convective mass flux, have R² values typically below 0.5–0.55 and retain substantial variability even when X does not vary significantly. For these descriptors, only a modest fraction of variance is associated with environmental state differences, indicating that most observed differences reflect variability not captured by the local environmental state, potentially from smaller-scale dynamics or stochastic processes.
Our findings identify which aspects of deep convection are almost entirely tied to their local environmental state at spatiotemporal scales commensurate with INCUS observations.
How to cite: Prasanth, S., Haddad, Z., Susiluoto, J., Marinescu, P., Bukowski, J., Singh, I., Grant, L., and van den Heever, S.: Using machine learning to unearth the aspects of deep convection that are robustly predictable from the local environmental state, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-15811, https://doi.org/10.5194/egusphere-egu26-15811, 2026.
The NASA TROPICS (Time-Resolved Observations of Precipitation structure and storm Intensity with a Constellation of Smallsats) Earth Venture (EVI-3) mission, was successfully launched into orbit on May 8 and May 25, 2023 (two CubeSats in each of the two launches into 550-km orbits with approximately 33-degree inclination). Over the course of the mission, TROPICS provided more spaceborne microwave soundings than any operational program, and the combined forecast impact was larger and more spatially coherent than that of any individual passive microwave platform, illustrating the benefit of constellation-based temporal sampling for constraining rapidly evolving tropical convection. Prior to the deorbit of the last TROPICS spacecraft in December 2025, observations of 3-D temperature and humidity, as well as cloud ice and precipitation horizontal structure, at high temporal resolution were used to conduct high-value science investigations of tropical cyclones. TROPICS has provided rapid-refresh microwave measurements (median refresh rate of better than 60 minutes early in the mission with four functional CubeSats) in twelve channels spanning 92 to 205 GHz over the tropics that can be used to observe the thermodynamics of the troposphere and precipitation structure for storm systems at the mesoscale and synoptic scale over the entire storm lifecycle. Thousands of high-resolution images of tropical cyclones have been captured by the TROPICS mission, revealing detailed structure of the eyewall and surrounding rain bands. The new 205-GHz channel in particular (together with a traditional channel near 92 GHz) has provided new information on the inner storm structure, and, coupled with the relatively frequent revisit and low downlink latency, has informed tropical cyclone analysis at operational centers. The suite of TROPICS products is publicly available with much improved median revisit rates and were provided with data latencies that are sufficient to enable their use in operational tropical cyclone forecasting applications. In this presentation, we highlight the use of these high-revisit thermodynamic data from TROPICS to better characterize storm structure and environmental conditions over a variety of cases over the 30-month mission lifetime.
How to cite: Blackwell, W. and the TROPICS Science Team: A Summary of the New Capabilities for Observing Tropical Cyclone Thermodynamic Structure Provided by the NASA TROPICS Mission, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-12774, https://doi.org/10.5194/egusphere-egu26-12774, 2026.
A multiradar mosaic is a key solution to the insufficient detection range of a single weather radar. In the traditional grid-preprocessed mosaicking method (GPM), radar polar coordinate data are interpolated into Cartesian grids to compensate for vertically undersampled regions in radar volume scans. However, such interpolation fails to accurately reconstruct the polarization parameters in these regions. Therefore, this study presentss a polar coordinate direct-mosaicking method (PDM) for the high-density radar network in South China, which directly operates on polar coordinate data and avoids initial interpolation. Based on typical precipitation cases from May to August 2021, three key issues in the PDM are addressed: First, horizontal reflectivity (ZH) biases and differential reflectivity (ZDR) offsets are corrected; second, the number of radars in the mosaicking process is evaluated, with five radars determined to be optimal; and third, the weights of different radar data are optimized by considering vertical and horizontal distances, along with the melting layer position. Compared with the GPM, the PDM yields a more accurate representation of the melting layer, with a smaller mean height error (192 m compared with 470 m) and a more realistic estimation of thickness (661 m compared with 1507 m). It also improves the continuity of polarimetric parameters within convective core regions. The case studies indicate that the PDM enables earlier identification of ZDR columns and more accurate estimation of their heights. These advancements provide high-quality observational constraints for cloud microphysical research and offer potential for improving convective-scale data assimilation.
How to cite: Li, Z., Wu, C., Liu, L., and Zhang, Y.: Enhancing Convective-scale Polarimetric Signatures through a Polar Coordinate Direct-Mosaicking Method for High-density Radar Networks, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-2320, https://doi.org/10.5194/egusphere-egu26-2320, 2026.
The indirect radar reflectivity assimilation method, which assimilates retrieved hydrometeors from radar reflectivity data, is simple and efficient in severe weather forecasting applications. However, it suffers from retrieval errors due to the uncertainties in discerning multiple hydrometeor types based solely on reflectivity observations. To mitigate these inaccuracies, dual‐polarization radar data are incorporated into the background‐dependent indirect reflectivity assimilation method in this study. First, the contribution of multiple hydrometeor species to the whole reflectivity is estimated using the observed reflectivity and background microphysical information; then, the hydrometeor classification algorithm (HCA) product from dual‐ polarization radar observations is introduced to correct the dominant hydrometeor type if in error; and finally, the contribution factors are adjusted and used to retrieve multiple hydrometeor species from reflectivity data. Through a single squall line case, it is demonstrated that the incorporation of the HCA product from dual‐ polarization radar data leads to more reasonable hydrometeor identification, with more supercooled rainwater above the melting layer and more graupel at low levels, thereby refining the hydrometeor analysis. With the 15‐ min rapid update cycling configuration, the changes in the analysis field enable more cold rain processes, resulting in more intense latent heat release at higher levels and stronger cooling near the surface in the forecast. This in turn strengthens updraft motion and cold pools in the convective regions, thereby improving the reflectivity and precipitation forecasts. Four cases' quantitative evaluations of the 0–3‐hr reflectivity and precipitation forecasts further validate the effectiveness of incorporating dual‐polarization radar data in the assimilation process.
How to cite: Chen, H. and Zhao, K.: Assimilation of Dual-Polarization Radar Data Based on Hydrometeor Classification for Improving the Short-Term Prediction of Convective Storms, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-17383, https://doi.org/10.5194/egusphere-egu26-17383, 2026.
How to cite: Nemani, R. S., Bannister, R. N., Lawless, A. S., Wei, H., and Thomas, C.: Machine Learning-Driven Background Error Covariances for High-Resolution Data Assimilation, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-1685, https://doi.org/10.5194/egusphere-egu26-1685, 2026.
Assimilation of column-integrated observations such as precipitable water vapor (PWV) remains challenging when vertical moisture profile constraints (e.g., radiosondes) are unavailable. Ensemble data assimilation systems typically employ Gaspari–Cohn (GC) localization (e.g., in DART) to limit the spatial influence of observations and reduce spurious correlations arising from finite ensemble size. However, GC assumes homogeneous and isotropic correlations and does not represent physically driven vertical inhomogeneity, such as transient moisture structures associated with convection or moisture transport. Consequently, vertically displaced but dynamically sensitive layers may be underrepresented during PWV assimilation. The generalized Gaspari–Cohn (GenGC) localization function introduced by Gilpin et al. (2023) relaxes these assumptions by allowing localization parameters to vary spatially, enabling inhomogeneous and anisotropic correlation structures. This flexibility is particularly relevant for PWV assimilation, where the vertical distribution of moisture sensitivity can vary substantially with atmospheric state.
Sensitivity of water vapor mixing ratio (qvapor) to PWV was analyzed for Tucson, Flagstaff, Albuquerque, and Santa Teresa at 12 UTC during July–September 2021. These sensitivities were used to estimate the appropriate vertical influence of PWV assimilation and to construct a vertically varying GenGC localization. The performance of GenGC was evaluated relative to a standard GC localization with a fixed vertical radius of 3.5 km. These four locations in the Southwest US are chosen since they are impacted by the North American monsoon. To date, forecasting the monsoon precipitation is still challenging even with convective-permitting models coupled with data assimilation (Risanto et al. 2026 – in review).
Global Positioning System PWV (GPS-PWV) observations from Tucson and Flagstaff for three summer days in 2021 were assimilated into a convective-permitting (1.8 km) WRF ensemble with 40 members. HRRRv4 provided initial and boundary conditions. PWV was assimilated at 12 UTC using both GC and GenGC vertical localization, and radiosonde observations are used for independent verification.
Initial results indicate that qvapor exhibits strongest sensitivity to PWV from the surface up to approximately 6 km MSL, motivating a GenGC localization that extends vertical influence on this level. PWV analyses produced using GenGC were generally closer to observations than those using GC, reflecting the inclusion of moisture adjustments above 3.5 km MSL. In some cases, GenGC also improved near-surface qvapor relative to GC; however, larger positive qvapor biases were found above ~5 km MSL. Further investigation is required to assess the robustness of these results.
Future work will expand the number of cases and implement GenGC directly within the DART framework to evaluate its broader applicability for assimilating PWV and related datasets.
How to cite: Risanto, C. B., Gilpin, S., and Arellano, A.: Investigating the Application of Generalized Gaspari-Cohn Correlation Function in Vertical Localization, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-2699, https://doi.org/10.5194/egusphere-egu26-2699, 2026.
Bayesian model calibration with data assimilation methods receives continued interest for climate models, where turbulence-resolving large eddy simulations (LES) often serve as the ground truth for cloud processes. This study extends the calibration hierarchy towards the LES themselves, which are in turn validated against measurement data. Ensemble data assimilation for turbulent simulations is approached from a smoother perspective by calibrating a semi-idealistic LES simulation against averaged measurements with an ensemble Kalman smoother (EnKS).
This work re-visits the well-known test case simulating marine stratocumulus clouds (DYCOMS-II), which has been used extensively for forward model validation. Sub-grid scale (SGS) turbulence parameters are calibrated alongside the parameterized initial condition and forcing, aiming at a wholistic uncertainty representation. For the PyCLES model (dx=35 m), the calibrated setup achieves an improvement over the default model configuration used in previous studies. Experiments with different advection schemes reveal how the calibration result varies for implicit and explicit SGS modeling. For some of the tested schemes, consistent model errors on some observations require manual specification of larger data uncertainties in order to stabilize the calibration.
Through the analysis of partial increments, the EnKS methodology provides insights to parametric model sensitivities, and a means to explore a large parameter space. However, the performance is hindered by weak nonlinearities in the parameter-to-data map, which includes a nonlinear normalizing parameter transform. The practical application of EnKS-based calibration for LES models is facilitated by relying directly on a perturbed physics ensemble, which is commonly used for sensitivity studies. The results also invite to re-interpret biases found in previous model intercomparison studies on the DYCOMS-II case, as well as to consider the influence of uncertainties in initial condition and forcing when assessing parametric sensitivities in LES simulations.
How to cite: Grund, D., Mishra, S., and Schemm, S.: Bayesian Calibration of a Large-Eddy Resolving Model towards Campaign Measurements with an Ensemble Kalman Smoother, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-5745, https://doi.org/10.5194/egusphere-egu26-5745, 2026.
Quantifying the evolution of uncertainty is critical to both probabilistic forecasting and data assimilation (DA) in numerical weather prediction (NWP). In this study, we investigate the applicability of conformal prediction (CP), a recent machine learning (ML) method, to quantify uncertainty in a controlled, idealized setting. We use a toy model, designed to mimic convective process. The CP provides a set of possible outcomes with a chosen confidence level. Here, we compare and evaluate the average empirical coverage, the average interval length and average interval score loss (AISL) for the three variants of CP i.e., a) Standard CP, b) Normalized CP and c) Conformalized Quantile Regression. We also combine DA with the CP estimates of uncertainty and quantify the significance of improvement. Our results highlight the strengths and limitations of each approach, providing insights into the effectiveness of CP to complement ensemble-based uncertainty quantification in simplified atmospheric models.
How to cite: George, C., Janjic, T., Javanmardi, A., and Hüllermeier, E.: Uncertainty quantification via conformal prediction in data assimilation, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-7371, https://doi.org/10.5194/egusphere-egu26-7371, 2026.
An accurate forecast of T2m in complex terrain is essential for a wide range of economic and societal applications. However, improper assimilation of T2m observations can degrade forecast quality.
The 3DVar scheme relies on a climatological error covariance matrix, which can produce unrealistic analysis increments and consequently inaccurate forecasts, particularly over sloped terrain. For example, an observation from a valley station may still generate increments at the mountaintop, even though the valley observation does not adequately represent the atmospheric conditions at the mountaintop. In contrast, 3DEnVar utilizes flow-dependent error covariances that are representative of the recent atmospheric flow regime. Hybrid-3DEnVar, on the other hand, employs an error covariance with 50\% weight assigned to the climatological covariance matrix and 50\% to the flow-dependent covariance matrix. Hybrid-3DEnVar has been tested in AROME-Austria and improved the initial conditions of AROME-Austria compared to 3DVar.
Our study compares the T2m forecast of Hybrid-3DEnVar against the operational 3DVar scheme of AROME-Austria in complex terrain.
For this purpose, we utilize the convective-scale limited-area ensemble forecast system (C-LAEF) Alpe Adria based on the numerical weather prediction model AROME, which operates at a horizontal grid space of 1 km. The 32-member ensemble of C-LAEF Alpe Adria (with a 1 km horizontal resolution) provides the flow-dependent error covariances.
We conduct three sets of data denial experiments in which we run a 3-hourly assimilation cycle over 10 days, once with all observations and once with the T2m measurements removed. The first set of experiments uses 3DVar, the second 3DEnVar, and the third Hybrid-3DEnVar. We present first-guess departure statistics and forecast verification for T2m over the Alpine terrain.
How to cite: Jyoti, K., Griewank, P., Meier, F., and Weissmann, M.: Can AROME-Austria Hybrid-3DEnVar improve the forecast of 2-meter temperature (T2m) in complex Alpine terrain?, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-9782, https://doi.org/10.5194/egusphere-egu26-9782, 2026.
Radar data assimilation (DA) is critical for convective-scale forecasting, as it provides real-time, high-resolution information on precipitation, wind, and convective system dynamics that is not captured by surface observations or satellite data. Polarimetric radar observations complement conventional reflectivity (Zh) and radial velocity (Vr) measurements by enabling the determination of hydrometeor types and particle size characteristics. Differential reflectivity (ZDR) is one of the key polarimetric radar variables, defined as the ratio between horizontal and vertical reflectivity, that provides direct information on hydrometeor shape and size. Despite its strong potential to better constrain storm microphysics and improve convective-scale forecasts, the assimilation of ZDR remains challenging. Challenges associated with observation operators, error characterisation, and data quality underscore the need for further research in this area.
The current study investigates the direct assimilation of ZDR in an observing system simulation experiment (OSSE) of a long-lived supercell. The OSSE is conducted using the ICOsahedral Nonhydrostatic (ICON) model with a two-moment microphysics scheme and the Local Ensemble Transform Kalman Filter (LETKF), employing both hydrometeor mixing ratios and number concentrations as analysis variables. Synthetic observations are generated using the polarimetric radar forward operator EMVORADO developed at the Deutscher Wetterdienst. The synthetic ZDR observations are assimilated in addition to the non-polarimetric variables, namely Zh and Vr, while a reference experiment assimilating only non-polarimetric synthetic observations was conducted for comparison. A series of sensitivity experiments are performed to assess the impact of DA settings on assimilation performance, for varying observation error, localisation radius, and ensemble size. In addition, appropriate thresholds and no-reflectivity (clear air) equivalents for ZDR observations are examined.
How to cite: Bardachova, T., Janjic, T., Carrassi, A., de Lozar, A., and Mendrok, J.: Direct assimilation of differential reflectivity in an idealised setup , EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-10222, https://doi.org/10.5194/egusphere-egu26-10222, 2026.
Partial analysis increments (PAI) are an efficient diagnostic to determine the influence individual observations or groups of observations on the analysis (Diefenbach et al.). Addittionally, the diagnostic can be used to approximate the influence that observations would have had for different assimilation settings. We use PAI to investigate whether observations could be more beneficial if they were assimilated using different localization settings. Localization is an essential component of any ensemble-based data-assimilation system, necessary to mitigate the effects of a limited ensemble size and to reduce the computational cost. Localization for satellite observations, which lack a constant or well-defined observation location, is particularly challenging, and numerous approaches have been proposed. Using PAI, we can estimate the performance of many different localization approaches without needing to rerun experiments and determine settings that lead to an optimal analsis using independent observations for verification. In this poster, we present results from a one-month-long cycled forecast of the regional modelling system of Deutscher Wetterdienst, in which the PAIs are compared against non-assimilated radiosondes. PAI is used to optimise and understand localisation by (1) considering a range of localisation functions over a normalised metric; (2) studying different combinations of parameters for a Gaspari-Cohn localization function and (3) optimising localization over a set of idealised functions. The current settings of DWD for vertical localisation of satellite radiances in the visible 0.6µm and infrared 6.2µm 7.3µm channels - which were originally designed to improve estimates of cloud-related variables - perform well against our metrics but could be improved upon by using a multi-peaked localisation function.
How to cite: Griewank, P., Diefenback, T., Necker, T., Parker, M., Schomburg, A., and Weissmann, M.: Optimizing localization for ensemble data assimilation using partial analysis increments, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-13080, https://doi.org/10.5194/egusphere-egu26-13080, 2026.
The atmospheric coherence length is a key parameter for quantitatively describing the strength of optical turbulence and is crucial for laser propagation, astronomical observation, and turbulence-degraded image restoration. Although existing studies have widely used convolutional neural networks (CNNs) to retrieve this parameter from single-frame distorted images, they fail to fully utilize the dynamic characteristics of turbulence. To address this, this study proposes a CNN-based inversion model incorporating dynamic optical flow features. By integrating the optical flow method with CNNs, the model captures and utilizes the turbulent motion information between consecutive image frames. The input to the model is the optical flow displacement field calculated from two successive image frames, and the angle-of-arrival fluctuation variance derived from the optical flow is incorporated into the loss function as a physical constraint. This design significantly enhances sensitivity to subtle image distortions induced by weak turbulence, effectively overcoming the performance bottleneck of traditional static single-frame input models under weak turbulence conditions.
The model was trained on a simulated dataset and validated using measured data obtained on June 13-14, 2022, at Science Island in Hefei, Anhui Province, China. The measured data were collected synchronously by an atmospheric coherence length monitor and an imaging device over a 500-meter horizontal near-ground propagation path. The study systematically compared the performance of four classic CNN architectures (AlexNet, VGGNet, EfficientNet, CVTStegoNet) with and without the incorporation of TV-L1 optical flow features. The results show that the introduction of optical flow features universally and significantly improves the inversion accuracy and robustness of the models under different turbulence strengths. Specifically, the method achieved faster training convergence and superior generalization performance across all tested models. On a test set comprising 4,500 training samples and 500 independent validation samples, the model with integrated optical flow features reduced the root mean square error (RMSE) and mean relative error (MRE) by an average of approximately 40%, while the coefficient of determination (R²) was generally above 0.99. Among the four models, the fused model based on AlexNet achieved the best overall performance.
This work demonstrates the critical role of utilizing turbulent dynamic features in enhancing the accuracy of inversion models, offering a novel and practical deep learning solution for high-precision, real-time detection of the atmospheric coherence length.
How to cite: Wang, Y., Wang, P., Huang, Y., and Mei, H.: Inversion of Atmospheric Coherence Length Using Optical Flow-based Convolutional Neural Networks, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-16922, https://doi.org/10.5194/egusphere-egu26-16922, 2026.
