Abstract The differential reflectivity (Z DR ) column is a notable radar signature in thunderstorms, serving as a proxy for updrafts and an early indicator of storm intensity. However, a quantitative relationship among the Z DR column morphology, related microphysical characteristics, and dynamic structures based on sufficient field observations has not yet been reported. To explore these quantitative relationships, we develop a three-dimensional Z DR column identification method based on 100 Z DR columns within severe thunderstorms (primarily supercells) in the Pearl River Delta region of South China from April to July in 2021 and 2024. 3D-connected component labeling with horizontal and vertical tolerance mechanisms is integrated into the method to ensure the structural independence and integrity of multiple Z DR columns coexisting within the same thunderstorm. This method shows that Z DR columns in severe thunderstorms exhibit mean depths of approximately 2 km, mean widths of approximately 6 km, mean volumes of approximately 30 km 3 , 30–70 min lifecycles, and high liquid water contents (mean =2.23 g m −3 ). Quantitative relationships between three morphological parameters (volume, width, and depth) of Z DR columns and updraft volumes retrieved from multiradar wind fields using 1–11 m s −1 thresholds were also determined. All three morphological parameters of the Z DR column outperform the traditional 35 dBZ echo top height in indicating updraft intensity. In addition, the distribution of hydrometeors was explored. Our results, derived from extensive observations, quantitatively highlight the strong relationship between Z DR column morphology and thunderstorm dynamic structure, providing a theoretical foundation for improving severe weather warnings and forecasts.
Abstract Accurately forecasting winter precipitation type and its transitions is critical for high-impact decision making. However, existing methods struggle in thermodynamically ambiguous regimes, and most do not quantify forecast uncertainty from a single model run. We developed an evidential neural network that predicts calibrated probabilities for four winter precipitation types (rain, snow, freezing rain, and ice pellets) along with epistemic uncertainty estimates at the computational cost of a standard neural network. The model was trained on quality-controlled and curated observations from the crowd-sourced mPING dataset paired with vertical thermodynamic profiles from the NOAA Rapid Refresh model analyses. Rigorous physical quality control removed thermodynamically implausible reports. Bulk evaluation against held-out mPING observations from June 2020 through June 2022 shows the ML model outperforms area-based deterministic methods in success ratio for freezing rain and ice pellets while maintaining comparable or better performance for rain and snow. A reduced freezing rain probability of detection reflects genuinely ambiguous thermodynamic environments rather than a uniform model deficiency and is more robustly represented through the full probability distribution than through the dominant predicted class alone. Thermodynamic regime analysis demonstrates that model prediction errors are physically structured and concentrated in interpretable regions of diagnostic space consistent with the known difficulty of freezing rain and ice pellet discrimination. We further demonstrate the model’s physical consistency and operational utility through two contrasting mid-western U.S. winter storm case studies and an interactive visualization tool that enables dynamic interrogation of model predictions and uncertainty in real time.
Abstract Gray-zone tropical cyclone (TC) simulations, in which convection is partially resolved and partially parameterized at grid spacings of approximately 3 to 15 km, are sensitive to how grid spacing changes across adaptive meshes. This study evaluates three scale-aware convective parameterization schemes (CPS), the Grell-Freitas (GF), Multi-Scale Kain-Fritsch (MSKF), and New Tiedtke (NTDK), across regional (WRF) and global (Model for Prediction Across Scales, MPAS; ClusterTech Platform for Atmospheric Simulation, CPAS) models for TC simulation over the South China Sea. Twenty-seven simulations spanning three TC cases (Category 2 to 5 intensity, including single-TC and binary interaction scenarios) are verified against observations. While domain-averaged precipitation differs by only 10 to 20% across models, the Convective Rain Ratio (CRR, the fraction of total precipitation produced by the CPS) shows that models achieve similar rainfall through different convective partitioning. Across the 27 simulations, CPS choice accounts for 44% of CRR variance, compared with 29% for the model framework and 16% for the selected TC case. CRR differences are largest in the gray zone, ranging from 30% to over 90% across schemes. Both global models exhibit mid-level drying relative to WRF regardless of CPS choice. This deficit may partly reflect differences in large-scale moisture constraints between regional and global model frameworks. CPAS’s adaptive mesh partially restores outer rainband precipitation relative to MPAS but introduces discontinuities where grid spacing changes sharply. Identical CPS produce different CRR patterns across models, differing by more than 20% for individual TC cases. Current scale-aware CPS adjust their behavior based on local grid spacing, but do not account for how rapidly that spacing changes across the mesh. Future schemes should explicitly account for local gradients in grid spacing.
Abstract The present study investigates potential orographic modifications of mesoscale environments by analyzing prestorm conditions associated with 62 supercells initiated near the Carpathian Mountains in Central Europe during April–September 2015–2019. Environmental diagnostics are derived from three data sources: high-resolution simulations of the COSMO model, ERA5 reanalysis, and upper-air soundings from an intermountain station. From these, multiple convection-related diagnostic variables are calculated, while storms are grouped into three categories based on their mean lifetime translation speed. A broader spatiotemporal context of key parameters, including mixed-layer CAPE, 0-6-km wind shear, 0-3-km storm-relative helicity, and the Supercell Composite Parameter, reveals substantial variability between datasets and across storm speed categories. Several prestorm environments exhibit shear values below commonly expected levels for supercell formation, particularly in slower-moving storms. Spatial analyses show that magnitudes and distributions of parameters differ over complex terrain. While southeastern slopes are often associated with enhanced buoyancy, lee sides of mountains more often exhibit environments favorable for mesocyclone development. These patterns may increase the occurrence of supercell-supportive conditions in specific locations. Among all evaluated variables, precipitable water is the strongest predictor of estimated precipitation intensity during the first 20 minutes, while the Downburst Environment Index best predicts maximum precipitation intensity over the full storm lifetime. Wind shear between the surface and the –10°C isotherm is the strongest predictor of mean storm speed. Spatial distributions of these predictors are also terrain-influenced, but in distinct ways. Overall, terrain can locally enhance both thermodynamic and dynamic components of supercell environments.
Abstract This paper contributes to ongoing efforts to enhance forecasting and understanding of bow echoes by examining their formation features in a sample of 290 cases studied in two prior works. The sample included events with non-severe, marginal severe, and high-end severe winds, separated into nocturnal and diurnal periods. The bow echo formation features explored included intersections of storm lines with frontal and outflow boundaries, interactions with individual cells ahead of the line, strong 0–6 km wind shear, and strong low-level line-normal storm-relative inflow into the main line. Apex curvature was also calculated to assess its relationship with wind severity and the possible influence of bowing features. A statistical analysis of this dataset revealed a strong relationship between bowing curvature and wind severity, consistent with past work linking greater curvature to stronger surface winds. The present analysis identifies the intersection of the convective line with a preexisting boundary, strong 0-6 km shear, and strong line-normal storm-relative inflow as features associated with bow echo formation and stronger bow echo winds, while individual cell interactions played a lesser role. However, in just over a tenth of events, none of the four features studied were present. This suggests that more complex interactions or storm-scale processes can also lead to bow echo formation and may not be fully captured by the feature set used here.
Abstract Accurate observation of cloud parameters is critical for weather forecasting, climate model improvement, and extreme-weather warning, yet the reconstruction of three-dimensional cloud structures remains a major challenge. In this study, geostationary infrared observations from FY-4B AGRI are combined with CMA-MESO background fields within a Bayesian framework to reconstruct the model-predicted hydrometeor field and subsequently derive the complete three-dimensional cloud structure, characterized by cloud liquid water content, cloud ice water content, and cloud fraction. The incorporation of prior information enhances the reliability and physical interpretability of the reconstructed cloud structure. For a squall line that affected South China on 30 April 2024, the reconstructed cloud field shows closer agreement with observations in both brightness temperature and radar reflectivity than the pre-reconstruction model simulation, yielding a more accurate representation of cloud-field characteristics. A further assimilation experiment demonstrates that introducing the reconstructed hydrometeors into the CMA-MESO initial field brings the spatial distribution of precipitation into closer agreement with observations and improves the threat score for light-to-moderate precipitation. Finally, combining the AGRI-reconstructed cloud field with GIIRS hyperspectral data to extract sub-grid-scale information highlights the value of multi-sensor synergy and establishes a foundation for hyperspectral all-sky data assimilation.
Abstract Forecasting rotation within thunderstorms requires high-resolution models which coarse convection-allowing models (CAMs) (e.g., 3-km horizontal grid spacing) may lack the ability to sufficiently resolve. Thunderstorm rotation forecasts from experimental versions of the Warn-on-Forecast System with 1- and 3-km horizontal grid spacing (WoFS-1 km and WoFS-3 km, respectively) are compared using 23 high-impact convective case studies. Both mid- and low-level rotation are identified using an object-based method as an attribute of parent composite reflectivity objects. Rotation verification using Multi-Radar Multi-Sensor (MRMS) system data and NWS tornado warnings is performed only when parent thunderstorms are accurately predicted by both WoFS-1 km and WoFS-3 km. Midlevel rotation probability of detection (POD) is higher in WoFS-1 km for all parent thunderstorm size bins, while the false alarm ratio (FAR) differences are statistically insignificant. This yields a statistically significant improvement in the equitable threat score (ETS) in WoFS-1 km. WoFS-1 km has a significantly higher ETS for parent thunderstorms with a minor axis length < 40 km for forecast lead times under 60 min. The low-level rotation statistics are similar to those of midlevel rotation although with fewer significant differences between WoFS-1 km and WoFS-3 km. The object-based method reveals that WoFS-3 km has too few rotating storms with rotation objects that are too large compared to WoFS-1 km. WoFS-1 km has smaller displacement errors for rotation than WoFS-3 km due to both smaller parent storm displacement errors and reduced intrastorm rotation location errors. Significance Statement This study shows that an ensemble data assimilation and short-term forecasting system with a 1-km horizontal grid spacing that better depicts rotation within thunderstorms than a system with 3-km horizontal grid spacing. This improvement includes more successful detection as well as better representation of thunderstorm rotation in terms of size and location. A 1-km system would provide improved model guidance of severe convection and the potential for associated hazards such as large hail, damaging winds, and tornadoes.
Abstract For decades, hurricane storm surge forecasting has been based on driving the hydrodynamics with reduced-order wind models, leveraging synthetic vortices that are generated from bulk storm parameters and distributed according to historical forecast errors. This strategy has the downside that the intricacies of the hurricane vortex and its far-field winds are lost. Here, we introduce a new methodology that leverages the development of high-fidelity meteorological products and high-resolution water level models, such as NOAA’s Surge and Tide Operational Forecast System 2D Global (STOFS-2D-Global), to generate probabilistic storm surge guidance (PSSG) at unprecedentedly high resolution. This is achieved by operating a global framework in a stochastic setting, driven by ensemble meteorological products. We explore this approach for Hurricane Ian (2022) in tandem with a nudging scheme that modifies core winds to be consistent with the National Hurricane Center’s peak wind intensity hindcast/forecast. We evaluate a range of meteorological hindcast and forecast products for Ian with and without nudging, all driving STOFS-2D-Global. It is shown that the accuracy of the projected storm surge in the Fort Myers region is highly sensitive to the forced winds, but that nudging allows the inherently muted (i.e., low bias) core wind products to perform well. Global Ensemble Forecast System (GEFS)-driven Ian forecasts with nudging are then implemented and statistical water levels are computed, with the standard deviation serving as a measure of confidence. Significance Statement Forecasting storm surge from hurricanes is a task complicated by uncertainty. This uncertainty stems primarily from meteorological forecasts. Producing probabilistic storm surge guidance (PSSG) rather than providing a single projection is a way to convey this uncertainty. Traditionally, idealized wind models of hurricanes, historical forecast errors, and relatively fast hydrodynamic solvers have been used to provide this guidance. Herein, we propose a new methodology that couples ensemble meteorological products with a high-fidelity finite-element-based hydrodynamics solver to produce detailed PSSG. We demonstrate this methodology for Hurricane Ian (2022). We show that this approach produces PSSG that reflects even small-scale coastal features.
Abstract We present a system for medium-range forecasting of large (2–5-cm diameter) and very large (≥5-cm diameter) hail using ensemble numerical weather prediction. Probabilistic models based on additive logistic regression and predictors identified in previous research were applied to 10-day forecasts from the 51-member ensemble of the European Centre for Medium-Range Weather Forecasts (ECMWF). We present verification during two warm seasons (May–September in 2024 and 2025). The models demonstrated notable skill at lead times beyond 7 days and area under the ROC curve (AUC) scores of 0.830/0.885 on day 7 and 0.770/0.834 on day 9 for large and very large hail, respectively. Bias correction was applied to the forecasts as a strong positive bias was noted, most pronounced at night. A process-based verification identified two main sources of the remaining forecast error: the evolution of convection into linear modes unlikely to produce hail and missed storm initiation. The product was used by participants of the European Severe Storms Laboratory (ESSL) Testbed in 2024 without any calibration. Although they noted its overforecasting tendency, they found it useful for the purpose of getting a first guess of large hail potential.
Abstract This study analyzes storm-total precipitation forecasts from the National Centers for Environmental Prediction Global Forecast System (GFS) and European Centre for Medium-Range Weather Forecasts (ECMWF) for the contiguous United States (CONUS) landfalling tropical cyclones (TCs) from 2018 to 2024 using an object-based methodology. Forecast errors are quantified in terms of location, precipitation intensity, and size errors for 1- and 5-in. threshold precipitation clusters for 40 cases from 36 landfalling TCs. While there are few signs of forecast error improvement as forecast lead time decreases, the GFS consistently underforecasts precipitation intensities within 1-in. clusters when compared to the ECMWF. In addition, cluster forecast errors are stratified by different TC-related characteristics. Overall, there is no systematic relationship between any cluster forecast error and TC track forecast errors; however, high track error cases have centroid location errors 75–100 km greater than low track error cases for 5-in. threshold clusters in both models. The most robust difference in cluster forecast errors is by landfall intensity, where 5-in. clusters disproportionally occur in hurricanes. Moreover, when compared to hurricanes, tropical depressions and storms exhibit higher 5-in. cluster location errors in the GFS and underforecast all amounts of precipitation in 5-in. clusters in both models. In addition, a GFS forecast for Hurricane Henri (2021) is examined to investigate potential factors leading to a poor location forecast. The GFS forecast appears to more slowly advance a nearby cutoff low, leading to a weaker interaction between the low and Henri and resulting in precipitation being forecast for the wrong location. Significance Statement Tropical cyclone (TC) precipitation forecast verification has traditionally been done using methods that can struggle to provide information on error sources that could be used to improve forecasts. Our study verifies TC precipitation forecasts from 2018 to 2024 using object-based methods that independently assess spatial and intensity errors. We found that the GFS underforecasts the amount of precipitation within 1-in. objects compared to the ECMWF. Additionally, the GFS and ECMWF tend to underforecast all amounts of precipitation for TCs that make landfall as a tropical depression or storm compared to TCs that make landfall as a hurricane. We also looked closer at a forecast from Hurricane Henri to understand what led to a poor precipitation location forecast.
Abstract The Bayesian model averaging (BMA) method has gained considerable attention and application in the realm of precipitation prediction, particularly in ensemble and probabilistic predictions of precipitation. This study applies the BMA method to generate an ensemble prediction of summer precipitation over China, using forecasts from four dynamical climate models including ECMWF_SEAS51, JMA_CPS3, NCC_CSM11, and NCEP_CFS2. We analyzed and evaluated the performance of the BMA ensemble prediction during 1991–2020, equal-weighted (EW) ensemble prediction, and individual model predictions based on the variables in terms of the spatial anomaly correlation coefficient (ACC), the prediction score (PS), and the root-mean-square errors (RMSEs). Additionally, we conducted an in-depth examination of the predictive uncertainty associated with each prediction. Primary findings are as follows: 1) Within the training period (1991–2014), the BMA ensemble prediction exhibited an average ACC of 0.12, surpassing both EW ensemble prediction and predictions of four individual models. The BMA ensemble prediction also achieved the highest PS score among all those approaches. 2) For the test period (2015–20), the BMA ensemble prediction consistently outperformed others with an average ACC and PS score of 0.21 and 78.3, respectively. Moreover, its RMSE value was consistently lower than that of competing predictions, further indicating the improved performance of BMA ensemble prediction. 3) BMA ensemble prediction result most closely aligned with the actual precipitation, as evidenced by the probability distribution of precipitation anomaly percentages. Notably, this outcome of BMA ensemble exhibited a superior signal-to-noise ratio compared to the other five predictions and demonstrated the least coefficient of variation among the precipitation predictions. Significance Statement The purpose of this study is to conduct multimodel ensemble precipitation forecasting through Bayesian model averaging (BMA) to improve forecast accuracy, as accurate precipitation prediction is essential for mitigating precipitation-related natural disasters. Our results show that the ensemble of model predictions by Bayesian averaging can further improve the accuracy of precipitation prediction and provide a feasible method for summer precipitation prediction in China.
Abstract Sea fog poses significant hazards to marine activities, yet its accurate numerical prediction remains challenging, largely due to the sensitive dependence on the initial moisture and thermal structure in the marine atmospheric boundary layer (MABL). This study proposes and evaluates an accelerated observational nudging scheme designed for assimilating satellite-derived humidity soundings into sea fog forecasts over the Yellow Sea. The new scheme is compared against the extended cycling three-dimensional variational data assimilation (3DVAR) approach, using two representative sea fog cases and 10 additional cases for validation. Results demonstrate that the accelerated nudging scheme successfully constrains analysis increments to the observed fog area, minimizing the spurious fog generation prevalent in the data assimilation experiments. This localization capability yields a more accurate representation of the initial MABL thermodynamic structure and significantly reduces forecast bias. Quantitative verification shows that the accelerated nudging scheme improves the equitable threat score (ETS) by an average of 36% relative to the extended cycling 3DVAR baseline and by 44% when combined with clear-sky processing. These improvements translate into more reliable fog warnings with reduced false alarms. The findings suggest that the proposed accelerated observational nudging scheme offers a computationally efficient and functionally superior alternative to the extended cycling 3DVAR for operational sea fog forecasting. Significance Statement Sea fog is a serious threat to marine safety by drastically reducing visibility, yet accurately forecasting it has remained challenging due to a lack of moisture information over the ocean. This study introduces a simple and efficient method to incorporate satellite-derived humidity information into weather models, helping to better predict where and when sea fog will form. Our approach reduces false alarms and improves forecast accuracy, offering a practical tool for marine operations and coastal management.
Abstract This paper assesses the full value chain resulting in the warnings issued in advance of Storm Eunice (18 February 2022) and Storm Ciarán (2 November 2023), focusing on impacts in southern England. Storm Eunice brought wind hazards over a large area, whereas Storm Ciarán’s hazard footprint covered a smaller area, but was more of a multi-hazard event, with wind, rainfall, and landslide hazards. This paper examines the successes, challenges, and lessons learned from these warnings within the warning value chain framework – a concept illustrating the process of warning production as a sequence of expertise sources (components) connected by bridges facilitating bidirectional information exchange. Using a questionnaire from the World Meteorological Organization’s High Impact Weather (HIWeather) Warning Value Chain Flagship Project, this paper offers a detailed assessment of the warning process, covering the full flow of information and decision-making across the value chain. The post-event analysis for these two storms involved diverse participants, including operational meteorologists, hydrometeorologists, civil contingency advisors, and scientists from the Met Office, Flood Forecasting Centre (FFC), and British Geological Survey. The process culminated in a group scoring exercise for each storm and value chain component. This paper presents the results of the two value chain assessments separately, before drawing comparisons. While both storms impacted southern UK, Storm Eunice posed different forecasting and warning challenges. The findings offer valuable insights for improving future warning systems, focusing on three key themes: (1) streamlining information flow, (2) improving warning communication, and (3) enhancing cross-organizational collaboration.
Abstract We investigated rainfall patterns in tropical South America (SA) using a multi-scale data approach that includes intra-seasonal, synoptic, and diurnal cycles. Rainfall patterns in SA are significantly influenced by El Niño-Southern Oscillation (ENSO), so we divided our analysis into the two opposite phases of ENSO. Although there is an advanced understanding of the role of the intra-seasonal mode on a global scale, the role of a potential 12–24-day climate mode over tropical South America is undocumented. In this study, we linked rainfall patterns to the 12–24-day mode, which appears to be the dominant intra-seasonal pattern but is also influenced by the 30–60-day mode. Therefore, rainfall variability over tropical South America is modulated by eastward-propagating 30–60-day intraseasonal disturbances. During cold ENSO years, these disturbances couple with a shortened 12–24-day mode, intensifying precipitation; during warm years, this coupling is necessary for inland propagation. Additionally, weather systems at the synoptic and diurnal scales are affected differently by the warm and cold phases of ENSO. First, wave patterns at each time scale propagate in opposite directions: synoptic waves propagate eastward, while diurnal waves move westward. Second, cold ENSO enhances synoptic wave disturbances, e.g., 2000; we find that the constructive interference of these synoptic and diurnal wave patterns appear to be a key mechanism contributing to the positive anomalous rainfall during cold ENSO phases. Conversely, warm ENSO suppresses synoptic wave patterns, corresponding to a drier rainfall regime, e.g., 1998.
During the spring and summer of 2023, wildfires in Canada produced large plumes of smoke, which signifi-cantly impacted air quality and poor visibility across the United States. In this study, the operational performance of the High-Resolution Rapid Refresh (HRRR) short-range forecast model was evaluated using quantitative comparisons between ground-based [Aerosol Robotic Network aerosol optical depth (AERONET-AOD), AirNow}PM2.5] and satellite [Moderate Resolution Imaging Spectroradiometer (MODIS) AOD] observations. A 3-month period (June-August 2023) was selected for these comparisons as these months capture some of the most widespread and persistent smoke-related air quality impacts across the United States during the 2023 wildfire season. Results showed that HRRR modeled surface pollutants reasonably well, with approximately 60% of days with an air quality index (AQI) > 100 (code orange and up) accurately predicted in the Northwest. Satellite-based AOD comparisons showed that the model struggled with lofted smoke. Additionally, both temporal and spatial proximities to the smoke source were evaluated. Shorter lead times as well as shorter smoke transport distances showed positive trends with measured and modeled AQI, with the Northwest and Northeast regions of the United States showing the highest accuracies. Code orange (and up) days were predicted accurately ' 45% of the time at a 24-h lead time and '25% of the time at a 48-h lead time in the Northwest. These comparisons provide opportunities to understand where improvements in the smoke modules of short-range guidance tools such as HRRR and Rapid Refresh Forecast System (RRFS) might be had.
This study tests the sensitivity of the National Centers for Environmental Prediction Global Forecast System (NCEP-GFS) forecasts to the presence of optically thick ice clouds with the Advanced Technology Microwave Sounder (ATMS) radiance observations. The tests use cloud phase/type (optically thick ice cloud amount) information available from the Visible Infrared Imaging Radiometer Suite (VIIRS) cloud product and include GFS seasonal experiments with a modified thinning criterion to select ATMS observations with minimum optically thick ice cloud content. Results indicate that the data assimilation and forecasting performance of the GFS were sensitive to the thinning criteria and, thus, to the presence of optically thick ice cloud amounts in the ATMS observations. Data assimilation statistics such as the root-mean-square errors (RMSEs) for the observation minus background (O-B) and observation minus analysis (O-A) (ATMS brightness temperatures) showed significant changes, while the GFS forecasts of several important variables (geopotential heights, temperature) showed improvements. These findings show that the assimilation of ATMS observations in the GFS is sensitive to the presence of optically thick ice cloud amounts. The minimization of optically thick ice cloud amounts in ATMS observations could have a positive impact on the assimilation and forecasting performance of the GFS. They also present a case for future work focusing on inclusion/testing of supplemental cloud phase/type-specific tests to the existing thinning criteria in the GFS to eliminate these observations at the earliest stage of the data assimilation process. SIGNIFICANCE STATEMENT: This study focuses on demonstrating the importance of removing microwave radiance observations impacted by the presence of optically thick ice clouds in the data assimilation system of an operational weather forecasting model [National Centers for Environmental Prediction Global Forecast System (NCEP-GFS)]. These observations are identified in the Advanced Technology Microwave Sounder (ATMS) dataset by using collocated Visible Infrared Imaging Radiometer Suite (VIIRS) information. Results show significant improvements in the forecasting performance of the GFS. Observations impacted by thick ice clouds can have adverse effects on the data assimilation and weather forecasting performance of NCEP-GFS. Findings from this study present the case to further refine existing microwave quality control to reduce the use of these observations.
Forecasting of extreme precipitation in areas of complex topography remains a challenge for the numerical weather prediction community. California receives 25%-50% of its annual rainfall from just a few storms, referred to as atmospheric rivers. Accurate and timely forecasts of atmospheric rivers are critical for managing water resources and flood risk. An important consideration is whether limited computational resources are best spent on larger ensembles or on finer horizontal grid spacing. In this study, we compare two experimental versions of the operational West-WRF ensemble run at 9-and 3-km grid spacing, each with 72 members and 60 vertical levels. Decreasing the horizontal grid spacing from 9 to 3 km adds skill to the forecast but increases the computational expense by a factor of-27. We then evaluate how the skill of the 3-km ensemble degrades as we reduce the number of members to 45, 30, and 15. We find that the 45-member version of the 3-km ensemble still outperforms the 72-member 9-km ensemble; however, a 45-member 3-km ensemble is too computationally expensive for most centers to run operationally. As we move to smaller ensembles, the skill degrades signifi-cantly, especially for extreme events. The 15-member version of the 3-km ensemble, while still being almost 6 times as computationally expensive, performs worse than the 9-km ensemble in almost every metric considered. Therefore, we conclude that while there is clear value in decreasing the horizontal grid spacing in an ensemble forecast, ensemble size is also important, particularly if the focus of the prediction is on extreme weather.
Abstract Global tropical cyclone warning centers construct representations of the vortex with estimates of intensity, central pressure, and position. This work analyzes differences between the real-time and postseason intensity estimates analyses. Using tropical cyclone data for 2004–23 seasons from the North Atlantic basin; eastern, central, and western North Pacific basins; north Indian Ocean Basin; and Southern Hemisphere basins, this work shows consistency between most estimates. But further examination illustrates that only 29%–38% of cases have consistent intensity estimates during extreme intensity change (i.e., rapid intensification and weakening). For the inconsistent cases, this work evaluates how the real-time estimates differ from observations and explores the impacts on future intensity forecast skill. During rapid intensification, forecasters can increase 12- and 24-h intensity forecast skill with improved initial intensities roughly 70% of the time with over a 10% skill improvement relative to a baseline forecast. This work shows similar results for 12-h rapid weakening intensity forecasts and also indicates that adjusting the initial intensity during rapid weakening provides trivial skill improvement relative to a baseline forecast at 24 h. The results suggest that improving the current intensity estimate through more accurate, more frequent, and timely low-latency detection will yield better tropical cyclone intensity forecasts. Significance Statement Tropical cyclone warning centers use the current maximum sustained near-surface wind speed (i.e., intensity) to convey and predict impacts and hazards. This work quantifies differences in the real-time, operational intensity estimates from the poststorm assessment of position, wind field, and intensity. Where the values differ, this work compares all analyses to analyses during extreme intensity changes (i.e., rapid intensification and weakening). This evaluation highlights that during rapid events, forecasters deviate from the available current intensity estimate information, potentially indicating a detection problem, large uncertainty in intensity estimates, or influence of other outside information like storm environment. Exploring these differences further, forecasters could improve 12- and 24-h intensity forecasts with better current intensity estimates.
Reliable predictions of tropical cyclone (TC) rapid intensification (RI) have long tantalized the research and forecasting communities. In recent years, considerable advances have been achieved in the probabilistic prediction of RI from the current time through the subsequent 24, 48, or even 72 h. It is thus natural to ask whether probabilistic methods may prove reliable in indicating RI episodes starting at more distant lead times. This question is particularly important in the case of a TC still in its formative stages. The present study addresses the prediction of RI at extended lead times using a logistic regression model that incorporates both environmental and inner-storm structure characteristics as prognostic indicators. These predictors are derived from the Global Ensemble Forecast System (GEFS) of the National Oceanic and Atmospheric Administration and a limited set of observations. In particular, the control member of the GEFS is used to derive and evaluate the model against observations. Drawing on data spanning 2019-24, the resulting lead-time-dependent logistic regression models derived for the Atlantic and eastern Pacific basins demonstrate skillful predictions of 24-h RI periods out to 120 h. Moreover, the regression scheme derived from the control member may be extended to the entire GEFS ensemble, yielding probabilistic estimates of RI along each forecasted track out to the same temporal horizon. Plots of the evolving values of relevant predictors along these trajectories further confer a measure of physical insight into the probabilities obtained, thereby linking statistical inference with dynamical interpretation. SIGNIFICANCE STATEMENT: This paper describes a simple probabilistic tool for anticipating tropical cyclone rapid intensification (RI) at extended forecast intervals. The prediction of RI episodes starting at such later leads has not hitherto been a capability in operational forecasting. In spite of its simplicity, this probabilistic model demonstrates skillful RI prediction at extended forecast horizons.
The National Oceanic and Atmospheric Administration/National Weather Service's Storm Prediction Center issues daily convective outlooks outlining spatial risk levels for severe thunderstorms with up to 8 days of lead time. Since 2002, categorical outlooks are a function of the hazard-specific (hail, wind, and tornado) probabilistic outlooks, which forecast the probability that each of these hazards will occur at a severe level within 25 mi of a given point. We develop two methods for evaluating gridded probabilistic day 1 outlook fields against their associated local storm reports (LSRs) and practically perfect hindcasts (PPHs; smoothed LSRs): probabilistic contingency tables and optical flow displacement. Probabilistic contingency tables reveal a trend toward underforecasting over the study period (2002-23), strongest for wind. These results are based on storm coverage (using an imperfect storm report dataset) and do not reflect forecast intensity or user perception of forecasts. This finding highlights the nuanced considerations involved in issuing and verifying convective outlooks. Next, an optical flow displacement method demonstrates utility in depicting how an outlook could be deformed to better match PPH. The analysis across this study period reveals small spatial biases in outlooks on moderate-and high-risk days, most notably that hail outlooks tend to be centered too far east while wind outlooks tend to be centered too far west relative to PPH. These findings may help guide forecasters in issuing convective outlooks and act as a basis for future research on convective outlooks and thunderstorm dynamics. SIGNIFICANCE STATEMENT: Probabilistic convective outlooks (COs) are daily forecasts for the probability of severe thunderstorm hazards (wind, hail, and tornadoes) across the United States. We develop two methods to compare COs to observed storm locations. First, probabilistic contingency tables reveal dataset-level trends, such as a trend toward underforecasting coverage over time. Second, optical flow displacement captures how a CO must be reshaped in space to be "correct." This analysis reveals that COs do not account enough for the common evolution from hail-dominant to wind-dominant storm modes as storms progress from west to east. We plan to use these verification methods in training a machine learning model that will assist in making COs and identify additional underlying patterns in severe thunderstorm environments.
