Improving Probabilistic Lightning Forecasts through Ensemble Postprocessing with Mesoscale Information

Improving Probabilistic Lightning Forecasts through Ensemble Postprocessing with Mesoscale Information | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Short Report Improving Probabilistic Lightning Forecasts through Ensemble Postprocessing with Mesoscale Information Haoyue Li, Jieyu Chen, Ziqiang Huo, Jialing Wang, Yong Wang This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8550311/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Accurate short-term lightning forecasting requires reliable representations of both lightning occurrence and intensity, as well as the underlying convective processes. While ensemble prediction systems (EPS) provide valuable probabilistic information, their ability to resolve mesoscale and convective-scale variability remains limited. In this study, we assess the added value of mesoscale information for probabilistic lightning forecasting over eastern China. A mesoscale ensemble is constructed from deterministic forecasts of the China Meteorological Administration (CMA) Mesoscale Model (MESO) using spatiotemporal neighborhood and time-lagging techniques and is combined with predictors from the CMA regional ensemble prediction system (REPS). Lightning occurrence and counts are modeled within a Bayesian additive model for location, scale, and shape (BAMLSS) framework, using a hurdle-based count regression to account for excess zeros and overdispersion. Influential nonlinear predictors are selected via stability selection combined with gradient boosting. Forecast performance with and without MESO-derived predictors is systematically evaluated. The results show that incorporating mesoscale information consistently improves forecast skill for both lightning occurrence and intensity across all verification metrics. These improvements are primarily associated with MESO predictors related to convective available potential energy and convective precipitation, highlighting the importance of mesoscale processes for probabilistic lightning forecasting. Lightning forecast Ensemble postprocessing Gradient boosting Count data regression Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 1 Introduction Lightning, a long-distance electrical discharge phenomenon, is both a fundamental atmospheric process involving intense energy release and an important indicator of severe convective weather (Underwood, 2006; Schultz et al., 2009). Accurate lightning forecasts are therefore of significant scientific and societal importance, particularly for early-warning systems and the mitigation of weather-related hazards. Lightning forecasting approaches can generally be classified into three categories: (i) statistical extrapolation methods based on historical observations (Betz et al., 2008; Kohn et al., 2011), (ii) machine learning-based approaches (Lin et al., 2019; Geng et al., 2021, 2022), and (iii) methods that rely on numerical weather prediction (NWP) outputs. While extrapolation and machine learning approaches may perform well at very short lead times, their skill typically declines for forecasts extending beyond a few hours. This study focuses on short-term lightning prediction, for which high temporal and spatial resolution is particularly important. NWP-based methods explicitly account for atmospheric physical processes and their interactions, thereby offering clear advantages for short-term lightning forecasting. However, raw NWP outputs often exhibit systematic biases, motivating the application of statistical postprocessing techniques. For example, Bright et al. (2005) developed a multi-parameter lightning rate parameterization scheme incorporating the lifting condensation level, mid-level convective available potential energy (CAPE), and equilibrium layer temperature. Gijben et al. (2017) applied stepwise logit regression to lightning forecasting, effectively mitigating the tendency of conventional logistic regression to underestimate the occurrence of rare events. Using high-resolution NWP data from the European Centre for Medium-Range Weather Forecasts (ECMWF), Simon et al. (2018, 2019) proposed probabilistic lightning forecasting approaches for the Alpine region based on generalized additive models and count data regression. Their results demonstrate that incorporating ensemble forecast information substantially enhances lightning prediction skill. Despite these advances, many existing approaches lack explicit convective-scale information, limiting their ability to adequately represent the mesoscale processes governing lightning activity. The Regional Ensemble Prediction System (REPS) from the China Meteorological Administration (CMA) provides advanced, high-resolution ensemble forecasts in China. However, convection-scale ensemble forecasts have not yet been implemented operationally. To address this limitation, this study constructs a mesoscale ensemble from the deterministic Mesoscale Model (MESO) forecasts utilizing spatiotemporal neighborhood and time-lagging techniques, thereby incorporating additional convective-scale information into lightning prediction. The southeast coastal regions of China have experienced a notable increase in lightning frequency, partly associated with enhanced atmospheric instability under global warming (Williams, 2004). Lightning activity in Jiangsu Province exceeds the global average for regions at comparable latitudes (Ma et al., 2004; Christian et al., 2003), making it a suitable study area for lightning forecasting research. In this context, we apply statistical postprocessing techniques to ensemble forecasts from both REPS and the combined REPS + MESO configuration to generate probabilistic forecasts of lightning occurrence and intensity for the afternoon period of the following day. The performance of the two experimental configurations is then compared to assess the relative contribution of mesoscale information. The remainder of this paper is structured as follows. Section 2 describes the data and methodology, including the lightning observations, NWP datasets, and the postprocessing framework, which comprises a parametric distribution for count data and the ensemble construction approach. Section 3 presents the results, including the selection of influential predictors, a comparative evaluation of the REPS and REPS + MESO experiments, and illustrative case studies. Finally, Section 4 discusses the findings and summarizes the key contributions of this study. 2 Data and Method 2.1 Data 2.1.1 Lightning Detection Data Lightning occurrence is derived from lightning location data over Jiangsu Province for the period from May to August during 2021–2023. Jiangsu Province and its surrounding areas (116°~122° E, 30°~36° N) are discretized into a regular grid with a spatial resolution of 0.1°×0.1°, resulting in a total of 60×60 grid cells. For each grid cell, the total number of lightning flashes recorded between 12:00 and 18:00 Beijing Time (BJT) is aggregated and used as the target variable. Statistical analysis indicates that grid cells with at least one lightning flash account for only approximately 4.28% of all samples. Consequently, the binary variable of lightning occurrence exhibits pronounced class imbalance, and the corresponding count data are characterized by a large proportion of zero values and substantial overdispersion, a common feature in convective event datasets (Cameron et al., 2013). 2.1.2 Numerical Weather Prediction (NWP) Data The predictor variables are obtained from the Regional Ensemble Prediction System (REPS), which is initialized daily at 20:00 BJT. REPS consists of 14 perturbed ensemble members and one control member, yielding a total of 15 members, with forecast lead times ranging from 16 to 22 h and a horizontal resolution of 10 km. The dataset spans the period from 2 May to 31 August in 2021 and 2022. From this period, 36 days are randomly selected for model training, while the remaining days served as an independent test set. To enhance the representation of convective-scale variability, an additional mesoscale ensemble constructed from the Mesoscale Model (MESO) is incorporated. The list of weather variables and derived statistics for MESO is the same as for REPS, and is provided in supplementary material, primarily including 2-meter temperature, total cloud cover, as well as relative humidity, temperature, horizontal wind speed, and geopotential height at the 500 hPa and 700 hPa levels. For all non-accumulated variables, several temporal statistics are computed, including (i) the afternoon mean, (ii) the difference between values at 18:00 and 12:00, and (iii) anomalies of the three afternoon values (12:00, 15:00, and 18:00) relative to the afternoon mean. Subsequently, two ensemble-based summary statistics are derived: the ensemble median (med) representing central tendency, and the inter-quartile range (iqr) quantifying ensemble spread. For each predictor, the indicator of the temporal statistic is inserted after the variable name and followed by the indicator of the ensemble statistic, where the components are separated by underscores. 2.2 Method 2.2.1 Count Data Regression To explicitly account for the large proportion of zero counts and the pronounced overdispersion in lightning observations, a hurdle modeling framework (Mullahy et al., 1986) is adopted. The hurdle model decomposes the lightning distribution into two components: a binary hurdle component and a truncated count component. The binary hurdle component models lightning occurrence (0/1) using a binomial distribution, while the truncated count component models strictly positive lightning counts using a zero-truncated negative binomial (ZTNB) distribution. The binary hurdle component is specified as a logit-binomial model with occurrence probability π. The truncated count component is characterized by a location parameter µ > 0 and a dispersion parameter θ > 0. All three parameters (π, µ, θ) are modeled as functions of EPS predictors and are estimated independently within the hurdle framework. To avoid overfitting and ensure model parsimony, candidate predictors are selected using gradient boosting combined with stability selection. The resulting final model is estimated using Markov chain Monte Carlo (MCMC) simulation. All model fitting and inference are conducted using the bamlss package in R. 2.2.2 Ensemble Construction of MESO To increase ensemble size and better represent forecast uncertainty, ensemble members are constructed using a combination of spatiotemporal neighborhood and time-lagging methods. Forecast data initialized daily at 21:00 BJT and 00:00 BJT on the following day, with a horizontal resolution of 3 km, are utilized for this purpose. In the temporal dimension, the time-lagging and neighborhood approaches are applied to generate six sets of ensemble members, corresponding to different initialization times for the same valid time and different forecast lead times for the same initialization, as illustrated in Fig. 1 . In the spatial dimension, a neighborhood approach is employed in which each grid point is treated as the center and its eight surrounding grid points are included. As a result, a total of 54 ensemble members (6 × 9) is constructed for each grid point. 3 Results In this section, we first present the results of the boosting-based selection of nonlinear predictors for lightning occurrence probability and intensity. Next, we evaluate and compare the predictive performance of models based on REPS and REPS + MESO predictors. Finally, we analyze representative case studies to illustrate the impact of incorporating mesoscale information. 3.1 Feature Importance Selection The selected nonlinear predictors associated with the lightning occurrence probability π at a one-day lead time are summarized in Table 1 . In total, five predictors are retained in the REPS-based experiment, whereas seven are selected when MESO-derived predictors are included. The selected predictors are primarily related to convective precipitation (cp) and convective available potential energy (cape), which jointly characterize the potential for updraft development and the vertical growth of convective clouds. In contrast, most of the predictors selected exclusively from the MESO ensemble are associated with atmospheric circulation fields. This suggests that MESO-derived predictors provide a stronger capability for representing the physical mechanisms underlying lightning activity. Table 1 Selected REPS/REPS + MESO predictors for lightning probability forecasts REPS REPS + MESO v700_18.mei v500_15.mei sqrt_cp_1812.me sqrt_cp_1812.me sqrt_cp_1812.iq sqrt_cp_1812.iq sqrt_cape_12.me sqrt_cape_12.me doy3 mv700_18.mei mv500_1812.mei msqrt_cape_18.iq The selected predictors for lightning intensity are summarized in Table 2 . In the REPS-based experiment, four predictors are retained for the mean parameter µ (mu) of the truncated count distribution, whereas only a single predictor is selected for the dispersion parameter θ (theta). The predictors associated with µ are predominantly related to the development of strong convective systems, consistent with their role in regulating lightning flash rates. With the incorporation of MESO-derived predictors, an additional term associated with θ is selected, which is closely related to the geopotential height field at 700 hPa. This pressure level is dynamically relevant for mesoscale convection, as variations in mid-tropospheric geopotential height reflect synoptic- and mesoscale flow patterns that modulate vertical wind shear, moisture advection, and large-scale ascent. These processes play a critical role in the initiation, organization, and maintenance of convective systems, thereby influencing the variability of lightning activity. Table 2 Selected REPS/REPS + MESO predictors for lightning intensity forecasts REPS REPS + MESO t500_15.iq(mu) t500_15.iq(mu) layth_1812.mei(mu) gh700_1812.mei(mu) gh700_1812.mei(mu) doy4(mu) u700_1812.mei(mu) t500_1812.mei(mu) sqrt_cape_12.me(theta) sqrt_cape_12.me(theta) mgh700_1812.mei (theta) 3.2 Model Performance Evaluation In this section, we compare the predictive performance of Bayesian Additive Models for Location, Scale, and Shape (BAMLSS) utilizing predictors selected from REPS (BR) and from REPS + MESO (BRM), respectively. Both models are evaluated on the same independent test dataset to ensure a fair comparison. The predictive skill for lightning occurrence is assessed using the estimated occurrence probability π. Verification is performed using the Brier score (BS) and the area under curve (AUC) derived from receiver operating characteristics (ROC). Figure 2 shows the temporal evolution of daily mean scores over the test period, with mean values across all dates indicated in the bottom-left corner of each panel. The results show that the BRM model generally outperforms the BR model across both verification metrics. In particular, the inclusion of MESO-derived predictors leads to systematically higher AUC values, indicating an improved ability to discriminate between lightning and non-lightning events. This improvement is likely attributable to the additional mesoscale information, which provides a more refined representation of mesoscale convective processes. By better capturing key physical mechanisms such as boundary-layer convergence, local thermal inhomogeneities, and organized vertical motion, the MESO-derived predictors enable a more accurate characterization of local environmental conditions favorable for lightning initiation. Overall, the inclusion of MESO-derived predictors substantially enhances the skill of probabilistic lightning occurrence forecasts. In terms of the BS, the BRM model substantially outperforms the BR model on most test days, across both low and high lightning frequencies. An exception is observed on two days characterized by extremely high lightning events, for which BR exhibits slightly better performance. This behavior indicates a reduced model skill under extreme conditions, suggesting limitations in learning and representing rare, high-impact lightning events. Such behavior may be attributed to two primary factors. First, periods of intense lightning activity are often associated with complex interactions across mesoscale and microscale processes that are inherently challenging to represent. Second, such events constitute only a small proportion of the training dataset, resulting in limited information for accurately learning the tails of the lightning occurrence distribution. The predictive performance for lightning intensity is evaluated based on the full predictive distribution of lightning counts, characterized by the parameters µ and θ. For each day, a discrete probability mass function is obtained and evaluated using the ranked probability score (RPS; Epstein, 1969) and the logarithmic score (LogS; Wood, 2017). The temporal evolution of daily mean scores over the test period is shown in Fig. 3 . The results indicate a clear improvement in forecast skill when MESO-derived predictors are included. The BRM model consistently outperforms the BR model across all test dates in terms of LogS and shows superior performance on most dates with respect to the RPS. In particular, the substantial improvement in LogS suggests that the incorporation of mesoscale information enhances both the sharpness and calibration of the predicted lightning intensity distributions. This reflects an improved ability to represent not only typical lightning activity but also higher-intensity events, which are especially relevant for impact-oriented forecasting. 3.3 Case Study To further illustrate the added value of mesoscale predictors, a comparative case study is conducted using forecasts based on REPS and REPS + MESO predictors. For both models, we analyze observed lightning counts, forecasted lightning occurrence probabilities, predicted lightning counts, and the probability of lightning counts exceeding four flashes per grid point. A representative high-impact case featuring widespread and intense lightning activity on 23 August 2021 is selected. Forecasts based solely on REPS predictors successfully capture the large-scale spatial patterns of lightning activity. However, they tend to overestimate both the spatial extent and the occurrence probability of lightning in regions with high observed activity, resulting in an increased number of false alarms. While such behavior may be advantageous for operational early-warning purposes, it also reflects limited discrimination capability in regimes of intense convection. Furthermore, lightning intensity in the upper tail of the distribution (the 90th percentile) is generally underestimated, likely due to the coarse spatial resolution and limited representation of mesoscale processes in the REPS predictors. In contrast, forecasts incorporating MESO-derived predictors exhibit a more refined representation of convective-scale structures. This leads to improved spatial localization of lightning-prone areas and a noticeable reduction in false alarms. Moreover, the inclusion of mesoscale information enhances the depiction of high-intensity lightning events, particularly in regions of strong convective development. These results highlight the added value of mesoscale predictors for capturing the structural and dynamical characteristics of severe convective systems. 4 Discussion and Conclusion The systematic evaluation conducted in this study demonstrates that lightning forecasts incorporating MESO-derived predictors consistently outperform those based solely on REPS predictors across all verification metrics. The inclusion of mesoscale information leads to notable improvements in probabilistic lightning forecasts, lightning count forecast skill, and the spatial localization of lightning-prone regions. The most pronounced improvement is observed for lightning count forecasts, where the REPS + MESO configuration yields substantially better verification scores. This result highlights the added value of MESO-derived predictors in providing relevant information for estimating the magnitude and variability of lightning activity. From a physical perspective, these gains can be attributed to the more refined representation of the convective environment provided by MESO-derived variables, which better capture key processes associated with convective initiation and evolution, including boundary-layer convergence, localized thermal heterogeneity, and organized vertical motion. Consequently, the model is better equipped to represent the triggering and intensification of mesoscale and microscale convective systems. From a predictor selection standpoint, the inclusion of MESO-derived predictors not only enlarges the set of available explanatory variables but also strengthens the predictive signals of convection-related variables originally derived from REPS. The model identifies a larger number of physically interpretable and dynamically relevant features as influential predictors, particularly those directly linked to convective development. This suggests that mesoscale information enhances the effective signal-to-noise ratio of the predictor set, thereby facilitating more robust statistical learning. Overall, the improved forecast skill obtained through the incorporation of MESO-derived predictors underscores the dominant role of mesoscale and microscale processes in lightning generation. Compared with the large-scale environmental background fields provided by REPS, MESO-derived predictors offer a more direct representation of local conditions conducive to convective initiation and development, thereby yielding higher discriminative ability for lightning forecasting. Looking ahead, the continued integration of higher-resolution, more physically consistent, and more accurate mesoscale NWP predictors is expected to further improve probabilistic forecasts of lightning activity and other forms of severe convective weather. In addition, future work may explore the incorporation of machine learning–based feature selection and hybrid modeling approaches to more effectively identify key predictors and further enhance forecasting skill. Declarations This work supported by the Sichuan Science and Technology Program (NO.2025YFNH0006). The authors have no relevant financial or non-financial interests to disclose. All authors contributed to the study conception and design. Material preparation, data collection and analysis were performed by Haoyue Li. The first draft of the manuscript was written by Haoyue Li and all authors commented on previous versions of the manuscript. All authors read and approved the final manuscript. The datasets generated during the current study are not publicly available due to confidentiality agreements but are available from the corresponding author on reasonable request. Competing interests We declare that we have no competing interests. Author Contribution All authors contributed to the study conception and design. Material preparation, data collection and analysis were performed by Haoyue Li. The first draft of the manuscript was written by Haoyue Li and all authors commented on previous versions of the manuscript. All authors read and approved the final manuscript. Acknowledgement This work supported by the Sichuan Science and Technology Program (NO.2025YFNH0006). The authors thank the anonymous reviewers for their valuable comments and suggestions, which have greatly contributed to improving the quality of this paper. Data Availability The datasets generated during the current study are not publicly available due to confidentiality agreements but are available from the corresponding author on reasonable request. References Ma M, Tao SC, Zhu BY, et al. (2004) Climate distribution of lightning density in China and surrounding areas observed by satellites. Science China, 34(4): 298-306. Doi: 10.1360/zd2004-34-4-298 Bauer P, Quintino T, Wedi N, et al. (2020) The ECMWF scalability program: progress and plans. European Centre for Medium-range Weather Forecasts: 12-14. Betz HD, Schmidt K, Oettinger WP, et al. (2008) Cell-tracking with lightning data from LINET. Advances in Geosciences, 17: 55-61. Doi: 10.5194/adgeo-17-55-2008 Bright DR, Wandishin MS, Jewell RE, et al. (2005) A physically based parameter for lightning prediction and its calibration in ensemble forecasts. Conference on Meteorological Applications of Lightning Data. Christian HJ, Blakeslee RJ, Boccippio DJ, et al. (2003) Global frequency and distribution of lightning as observed from space by the optical transient detector. Journal of Geophysical Research: Atmospheres, 108(D1): 4005-4019. Doi: 10.1029/2002JD002347 Crane RK. (1979) Automatic cell detection and tracking. Geoscience Electronics, IEEE Transactions on, 17(4): 250-262. Doi: 10.1109/TGE.1979.294654 Dixon M, Wiener G. (1993) TITAN: Thunderstorm identification, tracking, analysis, and nowcasting-a radar-based methodology. Journal of Atmospheric and Oceanic Technology, 10(6): 785. Doi: 10.1175/1520-0426(1993)0102.0.CO;2 Epstein ES. (1969) A scoring system for probability forecasts of ranked categories. Journal of Applied Meteorology and Climatology, 8(6): 985–987. Doi: 10.1175/1520-0450(1969)0082.0.CO;2 Geng YA, Li QY, Lin TY, et al. (2021) A deep learning framework for lightning forecasting with multi-source spatiotemporal data. Quarterly Journal of the Royal Meteorological Society, 147(741): 4048-4062. Doi: 10.1002/qj.4167 Glahn HR, Lowry DA. (1972) The use of model output statistics (MOS) in objective weather forecasting. Journal of Applied Meteorology and Climatology, 11: 1203-1211. Doi: 10.1175/1520-0450(1972)0112.0.CO;2 Hering AM, S'en'esi S, Ambrosetti P, et al. (2004) Nowcasting thunderstorms in the Alpine region using a radar based adaptive thresholding scheme. Proceedings of the 3th European Conference on Radar in Meteorology and Hydrology: 206-211. Hering AM, S'en'esi S, Ambrosetti P, et al. (2005) Nowcasting thunderstorms in complex cases using radar data WSN05 conference, Toulouse 5-9 September 2005. Proceedings on CD, contribution 2, 14. Johnson JT, Mackeen PL, Witt A, et al. (1998) The storm cell identification and tracking algorithm: an enhanced WSR-88D algorithm. Weather and Forecast, 13: 263–276. Doi: 10.1175/1520-0434(1998)0132.0.CO;2 Klein WH. (1971) Computer prediction of precipitation probability in the United States. Journal of Applied Meteorology and Climatology, 10 (5): 903-915. Doi: 10.1175/1520-0450(1971)0102.0.CO;2 Kober K, Tafferner A. (2009) Tracking and nowcasting of convective cells using remote sensing data from radar and satellite. Meteorologische Zeitschrift, 18(1): 75–84. Doi: 10.1127/0941-2948/2009/359 Kohn M, Galanti E, Price C, et al. (2011) Nowcasting thunderstorms in the Mediterranean region using lightning data. Atmospheric Research, 100(4): 489-502. Doi: 10.1016/j.atmosres.2010.08.010 Lin TY, Li QY, Geng YA, et al. (2019) Attention-based dual-source spatiotemporal neural network for lightning forecast. IEEE Access, 7: 158296−158307. Doi: 10.1109/ACCESS.2019.2950328 Mayr A, Fenske N, Hofner B, et al. (2012) Generalized additive models for location, scale and shape for high dimensional data—a flexible approach based on boosting. Royal Statistical Society, 61(3), 403–427. Doi: 10.1111/j.1467-9876.2011.01033.x Michalon N, Nassif A, Saouri T, et al. (1999) Contribution to the climatological study of lightning. Geophysical Research Letters, 26(20): 3097-3100. Doi: 10.1029/1999GL010837 Mostajabi A, Finney DL, Rubinstein M, et al. (2019) Nowcasting lightning occurrence from commonly available meteorological parameters using machine learning techniques. Climate and Atmospheric Science, 2(1): 41. Doi: 10.1038/s41612-019-0098-0 Mueller C, Saxen T, Roberts R, et al. (2003) NCAR auto-nowcast system. American Meteorological Society, 18: 545–561. Doi: 10.1175/1520-0434(2003)0182.0.CO;2 Mullahy J. (1986) Specification and testing of some modified count data models. Journal of econometrics, 33(3): 341–365. Doi: 10.1016/0304-4076(86)90002-3 Price C, Rind D. (1992) A simple lightning parameterization for calculating global lightning distributions. Journal of Geophysical Research Atmospheres, 1992, 97(D9): 9919-9933. Doi: 10.1029/92JD00719 Rajeevan M, Madhulatha A, Rajasekhar M, et al. (2012) Development of a perfect prognosis probabilistic model for prediction of lightning over south-east India. Earth System Science Data, 121: 355–371. Doi: 10.1007/s12040-012-0173-y Rinehart RE, Garvey ET. (1978) Three-dimensional storm motion detection by conventional weather radar. Nature, 273: 287-289. Doi: 10.1038/273287a0 Schön, C, Dittrich J, Müller R. (2019) The error is the feature: how to forecast lightning using a model prediction error. In Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining: 2979-2988. Schultz CJ, Petersen WA, Carey LD. (2009) Preliminary development and evaluation of lightning jump algorithms for the real-time detection of severe weather. Journal of Applied Meteorology and Climatology, 48(12): 2543-2563. Doi: 10.1175/2009JAMC2237.1 Seed AW. (2003) A dynamic and spatial scaling approach to advection forecasting. Journal of Applied Meteorology and Climatology, 42(3): 381-388. Doi: 10.1175/1520-0450(2003)0422.0.CO;2 Simon T, Fabsic P, Mayr GJ, et al. (2018) Probability forecasting of thunderstorms in the Eastern Alps. Monthly Weather Review, 146(9): 2999-3009. Doi: 10.1175/MWR-D-17-0366.1 Simon T, Mayr GJ, Umlauf N, et al. (2019) NWP-based lightning prediction using flexible count data regression. Advances in Statistical Climatology, Meteorology and Oceanography, 2019. DOI:10.5194/ASCMO-5-1-2019 Underwood SJ. (2006) Cloud-to-ground lightning flash parameters associated with heavy rainfall alarms in the Denver, Colorado, urban drainage and flood control district ALERT network. Monthly Weather Review, 134(9): 2566-2580. Doi: Williams ER. (2005) Lightning and climate: A review. Atmospheric Research, 76(1): 272-287. Doi: 10.1029/2004JD005332 Wood SN. (2017) Generalized Additive Models: An Introduction with R. 2nd ed. Texts in Statistical Science Series, Chapman & Hall/CRC: 476. Xu MY, Qie XS, Pang WJ, et al. (2022) Lightning climatology across the Chinese continent from 2010 to 2020. Atmospheric Research, 275: 106251. Doi: Yair Y, Lynn B, Price C, et al. (2010) Predicting the potential for lightning activity in Mediterranean storms based on the Weather Research and Forecasting (WRF) model dynamic and microphysical fields. Journal of Geophysical Research: Atmospheres, 115(D4): D4205. Doi: 10.1016/j.atmosres.2022.106251 Zhou XY, Geng YA, Yu HM, et al. (2022) LightNet+: A dual-source lightning forecasting network with bi-direction spatiotemporal transformation. Applied Intelligence, 52: 11147-11159. Doi: 10.1007/s10489-021-03089-5 Additional Declarations No competing interests reported. Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-8550311","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Short Report","associatedPublications":[],"authors":[{"id":576609215,"identity":"301d3ed2-760c-41af-b03f-12daabe24a45","order_by":0,"name":"Haoyue Li","email":"","orcid":"","institution":"Nanjing University of Information Science and Technology","correspondingAuthor":false,"prefix":"","firstName":"Haoyue","middleName":"","lastName":"Li","suffix":""},{"id":576609219,"identity":"fd921521-0988-4d05-b0dc-244da2e2446c","order_by":1,"name":"Jieyu Chen","email":"","orcid":"","institution":"Nanjing University of Information Science and Technology","correspondingAuthor":false,"prefix":"","firstName":"Jieyu","middleName":"","lastName":"Chen","suffix":""},{"id":576609223,"identity":"51c63a3c-74bc-4551-9310-242ce688e197","order_by":2,"name":"Ziqiang Huo","email":"","orcid":"","institution":"Nanjing University of Information Science and Technology","correspondingAuthor":false,"prefix":"","firstName":"Ziqiang","middleName":"","lastName":"Huo","suffix":""},{"id":576609228,"identity":"f1e5edfd-5d95-4fc6-ad67-1e14a4ea7af5","order_by":3,"name":"Jialing Wang","email":"","orcid":"","institution":"Nanjing University of Information Science and Technology","correspondingAuthor":false,"prefix":"","firstName":"Jialing","middleName":"","lastName":"Wang","suffix":""},{"id":576609230,"identity":"6a7691b1-0cdd-4524-9b9f-18fbf42ba599","order_by":4,"name":"Yong Wang","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA8UlEQVRIiWNgGAWjYFAD9gYGhgQwK4FYLTwHQIoNSNEiAVZJhBaD472HXzC22eXJR749JvHgzx8GfvYcA4afO/BoOXMuzYKxLbnY8HZemkRimwGDZM8bA8beM3i03MgxM2Dcxpy4cXaOmURigwFIxICZsY2glvrEjTPPmEkk/DFgsCdCi/EDxm2HE+dL8AC1sAFtkSCgRfLMGTOGxH/HEzfw5BhbJLYZ80iceVZwsBePFr7jPcYfPpypTpzffsbw5o8/cnL87ckbH/zEo0XhAAMbOEYMDkAEeEDEAdwaGBjkGxiYP0AZo2AUjIJRMAqwAwC9J1H73YkULAAAAABJRU5ErkJggg==","orcid":"","institution":"Nanjing University of Information Science and Technology","correspondingAuthor":true,"prefix":"","firstName":"Yong","middleName":"","lastName":"Wang","suffix":""}],"badges":[],"createdAt":"2026-01-08 10:38:15","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-8550311/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-8550311/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":100799469,"identity":"0aa332b9-603b-40c8-8a35-30f669c1d5bb","added_by":"auto","created_at":"2026-01-21 13:59:15","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":879228,"visible":true,"origin":"","legend":"","description":"","filename":"Lietal.submitted8Jan2026.docx","url":"https://assets-eu.researchsquare.com/files/rs-8550311/v1/fddf638b7de5689ac4814995.docx"},{"id":100798910,"identity":"849bb25a-c8bf-41e0-be5a-c70d54cb4992","added_by":"auto","created_at":"2026-01-21 13:56:33","extension":"json","order_by":1,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":6946,"visible":true,"origin":"","legend":"","description":"","filename":"8a302ed6ce754801813ad55f6059864b.json","url":"https://assets-eu.researchsquare.com/files/rs-8550311/v1/9e8c98a5d0fc7e9bfc1f825c.json"},{"id":100797992,"identity":"5bc1dd35-40c7-42a8-892f-fff61a34e608","added_by":"auto","created_at":"2026-01-21 13:52:09","extension":"xml","order_by":2,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":57802,"visible":true,"origin":"","legend":"","description":"","filename":"8a302ed6ce754801813ad55f6059864b1enriched.xml","url":"https://assets-eu.researchsquare.com/files/rs-8550311/v1/23ec52070fcda420a7553e36.xml"},{"id":100798858,"identity":"31ea3932-d0de-44e5-ace0-c13ac284c19c","added_by":"auto","created_at":"2026-01-21 13:56:19","extension":"png","order_by":8,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":6469,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-8550311/v1/42d979b6c1b1e8f8dea2710c.png"},{"id":100798810,"identity":"e636c088-e1bf-4e88-970a-79a138a9cd39","added_by":"auto","created_at":"2026-01-21 13:56:03","extension":"png","order_by":9,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":62620,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-8550311/v1/7a8145b6e9627aae17625e4f.png"},{"id":100797948,"identity":"62c9b3f7-74b0-4944-b1ca-63a9cb446975","added_by":"auto","created_at":"2026-01-21 13:51:49","extension":"png","order_by":10,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":63097,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-8550311/v1/150003d8c73e22a955862c26.png"},{"id":100798811,"identity":"496f1d96-a4cd-4fa2-82bd-d2a2047378ba","added_by":"auto","created_at":"2026-01-21 13:56:07","extension":"png","order_by":11,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":64992,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-8550311/v1/fa0a3611e51988ae2545db93.png"},{"id":100798944,"identity":"eb7ab6ad-57a5-4c31-ae91-6b15ec32edd0","added_by":"auto","created_at":"2026-01-21 13:56:58","extension":"png","order_by":12,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":70912,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-8550311/v1/5f5bdd8ab76e177f6ef45715.png"},{"id":100799360,"identity":"d179f7a1-1b37-4486-9dfa-9f371b7cdcbe","added_by":"auto","created_at":"2026-01-21 13:58:55","extension":"xml","order_by":13,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":56312,"visible":true,"origin":"","legend":"","description":"","filename":"8a302ed6ce754801813ad55f6059864b1structuring.xml","url":"https://assets-eu.researchsquare.com/files/rs-8550311/v1/b180796418f1af58e5b9df39.xml"},{"id":100798847,"identity":"684c7246-1ea0-421b-8477-3407eabfa68a","added_by":"auto","created_at":"2026-01-21 13:56:14","extension":"html","order_by":14,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":62266,"visible":true,"origin":"","legend":"","description":"","filename":"earlyproof.html","url":"https://assets-eu.researchsquare.com/files/rs-8550311/v1/ae574713440f33dbf1ef331e.html"},{"id":100798047,"identity":"0d63b965-e89b-4d77-9074-c13b99b1e234","added_by":"auto","created_at":"2026-01-21 13:52:24","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":16895,"visible":true,"origin":"","legend":"\u003cp\u003eSchematic diagram of ensemble construction at 13:00 using the time-lagging method (red) and time-neighborhood method (green)\u003c/p\u003e","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-8550311/v1/5c2b7d016628c5e6d25c070d.png"},{"id":100798813,"identity":"0453ee62-4e13-4654-9fe9-baa8a32ecbc2","added_by":"auto","created_at":"2026-01-21 13:56:08","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":176316,"visible":true,"origin":"","legend":"\u003cp\u003eLine plots of temporal evolution of daily mean (a) AUC and (b) BS for the binary hurdle part of REPS and REPS+MESO models\u003c/p\u003e","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-8550311/v1/25471ccf1da9ab4343cdfa6e.png"},{"id":100797956,"identity":"b229cd7a-ae5e-4e78-a244-c06204b383dc","added_by":"auto","created_at":"2026-01-21 13:51:53","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":173786,"visible":true,"origin":"","legend":"\u003cp\u003eLine plots of temporal evolution of daily mean (a) LogS and (b) RPS for the truncated count part of REPS and REPS+MESO models\u003c/p\u003e","description":"","filename":"floatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-8550311/v1/cff9c950fa029c336573b349.png"},{"id":100797929,"identity":"bf5db7d2-ea79-4509-b02e-89ce1db4e1b1","added_by":"auto","created_at":"2026-01-21 13:51:38","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":283268,"visible":true,"origin":"","legend":"\u003cp\u003e(a) Observed lightning counts, (b) Predicted lightning occurrence probability, (c) Predicted lightning counts at 90th-percentile, (d) Predicted probability of lightning counts exceeding 4 at each grid point, on 23 August 2021, derived from the REPS experiment\u003c/p\u003e","description":"","filename":"floatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-8550311/v1/8bda79eea12a965d612fea26.png"},{"id":100798730,"identity":"3f6d60b1-0987-41a6-a340-0acfd649257f","added_by":"auto","created_at":"2026-01-21 13:55:43","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":237772,"visible":true,"origin":"","legend":"\u003cp\u003e(a) Observed lightning counts, (b) Predicted lightning occurrence probability, (c) Predicted lightning counts at 90th-percentile, (d) Predicted probability of lightning counts exceeding 4 at each grid point, on 23 August 2021, derived from the REPS+MESO experiment\u003c/p\u003e","description":"","filename":"floatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-8550311/v1/54fb07ab33600f0fce5282d1.png"},{"id":101721752,"identity":"9bf79e3f-6898-41c8-9759-98496feab680","added_by":"auto","created_at":"2026-02-03 03:09:59","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1270326,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8550311/v1/fecffaa4-0d08-427f-bc3a-eefc4ce7a351.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Improving Probabilistic Lightning Forecasts through Ensemble Postprocessing with Mesoscale Information","fulltext":[{"header":"1 Introduction","content":"\u003cp\u003eLightning, a long-distance electrical discharge phenomenon, is both a fundamental atmospheric process involving intense energy release and an important indicator of severe convective weather (Underwood, 2006; Schultz et al., 2009). Accurate lightning forecasts are therefore of significant scientific and societal importance, particularly for early-warning systems and the mitigation of weather-related hazards.\u003c/p\u003e \u003cp\u003eLightning forecasting approaches can generally be classified into three categories: (i) statistical extrapolation methods based on historical observations (Betz et al., 2008; Kohn et al., 2011), (ii) machine learning-based approaches (Lin et al., 2019; Geng et al., 2021, 2022), and (iii) methods that rely on numerical weather prediction (NWP) outputs. While extrapolation and machine learning approaches may perform well at very short lead times, their skill typically declines for forecasts extending beyond a few hours. This study focuses on short-term lightning prediction, for which high temporal and spatial resolution is particularly important.\u003c/p\u003e \u003cp\u003eNWP-based methods explicitly account for atmospheric physical processes and their interactions, thereby offering clear advantages for short-term lightning forecasting. However, raw NWP outputs often exhibit systematic biases, motivating the application of statistical postprocessing techniques. For example, Bright et al. (2005) developed a multi-parameter lightning rate parameterization scheme incorporating the lifting condensation level, mid-level convective available potential energy (CAPE), and equilibrium layer temperature. Gijben et al. (2017) applied stepwise logit regression to lightning forecasting, effectively mitigating the tendency of conventional logistic regression to underestimate the occurrence of rare events. Using high-resolution NWP data from the European Centre for Medium-Range Weather Forecasts (ECMWF), Simon et al. (2018, 2019) proposed probabilistic lightning forecasting approaches for the Alpine region based on generalized additive models and count data regression. Their results demonstrate that incorporating ensemble forecast information substantially enhances lightning prediction skill.\u003c/p\u003e \u003cp\u003eDespite these advances, many existing approaches lack explicit convective-scale information, limiting their ability to adequately represent the mesoscale processes governing lightning activity. The Regional Ensemble Prediction System (REPS) from the China Meteorological Administration (CMA) provides advanced, high-resolution ensemble forecasts in China. However, convection-scale ensemble forecasts have not yet been implemented operationally. To address this limitation, this study constructs a mesoscale ensemble from the deterministic Mesoscale Model (MESO) forecasts utilizing spatiotemporal neighborhood and time-lagging techniques, thereby incorporating additional convective-scale information into lightning prediction.\u003c/p\u003e \u003cp\u003eThe southeast coastal regions of China have experienced a notable increase in lightning frequency, partly associated with enhanced atmospheric instability under global warming (Williams, 2004). Lightning activity in Jiangsu Province exceeds the global average for regions at comparable latitudes (Ma et al., 2004; Christian et al., 2003), making it a suitable study area for lightning forecasting research. In this context, we apply statistical postprocessing techniques to ensemble forecasts from both REPS and the combined REPS\u0026thinsp;+\u0026thinsp;MESO configuration to generate probabilistic forecasts of lightning occurrence and intensity for the afternoon period of the following day. The performance of the two experimental configurations is then compared to assess the relative contribution of mesoscale information.\u003c/p\u003e \u003cp\u003eThe remainder of this paper is structured as follows. Section 2 describes the data and methodology, including the lightning observations, NWP datasets, and the postprocessing framework, which comprises a parametric distribution for count data and the ensemble construction approach. Section 3 presents the results, including the selection of influential predictors, a comparative evaluation of the REPS and REPS\u0026thinsp;+\u0026thinsp;MESO experiments, and illustrative case studies. Finally, Section 4 discusses the findings and summarizes the key contributions of this study.\u003c/p\u003e"},{"header":"2 Data and Method","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003e2.1 Data\u003c/h2\u003e \u003cdiv id=\"Sec4\" class=\"Section3\"\u003e \u003ch2\u003e2.1.1 Lightning Detection Data\u003c/h2\u003e \u003cp\u003eLightning occurrence is derived from lightning location data over Jiangsu Province for the period from May to August during 2021\u0026ndash;2023. Jiangsu Province and its surrounding areas (116\u0026deg;~122\u0026deg; E, 30\u0026deg;~36\u0026deg; N) are discretized into a regular grid with a spatial resolution of 0.1\u0026deg;\u0026times;0.1\u0026deg;, resulting in a total of 60\u0026times;60 grid cells. For each grid cell, the total number of lightning flashes recorded between 12:00 and 18:00 Beijing Time (BJT) is aggregated and used as the target variable.\u003c/p\u003e \u003cp\u003eStatistical analysis indicates that grid cells with at least one lightning flash account for only approximately 4.28% of all samples. Consequently, the binary variable of lightning occurrence exhibits pronounced class imbalance, and the corresponding count data are characterized by a large proportion of zero values and substantial overdispersion, a common feature in convective event datasets (Cameron et al., 2013).\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec5\" class=\"Section3\"\u003e \u003ch2\u003e2.1.2 Numerical Weather Prediction (NWP) Data\u003c/h2\u003e \u003cp\u003eThe predictor variables are obtained from the Regional Ensemble Prediction System (REPS), which is initialized daily at 20:00 BJT. REPS consists of 14 perturbed ensemble members and one control member, yielding a total of 15 members, with forecast lead times ranging from 16 to 22 h and a horizontal resolution of 10 km. The dataset spans the period from 2 May to 31 August in 2021 and 2022. From this period, 36 days are randomly selected for model training, while the remaining days served as an independent test set. To enhance the representation of convective-scale variability, an additional mesoscale ensemble constructed from the Mesoscale Model (MESO) is incorporated. The list of weather variables and derived statistics for MESO is the same as for REPS, and is provided in supplementary material, primarily including 2-meter temperature, total cloud cover, as well as relative humidity, temperature, horizontal wind speed, and geopotential height at the 500 hPa and 700 hPa levels.\u003c/p\u003e \u003cp\u003eFor all non-accumulated variables, several temporal statistics are computed, including (i) the afternoon mean, (ii) the difference between values at 18:00 and 12:00, and (iii) anomalies of the three afternoon values (12:00, 15:00, and 18:00) relative to the afternoon mean. Subsequently, two ensemble-based summary statistics are derived: the ensemble median (med) representing central tendency, and the inter-quartile range (iqr) quantifying ensemble spread. For each predictor, the indicator of the temporal statistic is inserted after the variable name and followed by the indicator of the ensemble statistic, where the components are separated by underscores.\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv id=\"Sec6\" class=\"Section2\"\u003e \u003ch2\u003e2.2 Method\u003c/h2\u003e \u003cdiv id=\"Sec7\" class=\"Section3\"\u003e \u003ch2\u003e2.2.1 Count Data Regression\u003c/h2\u003e \u003cp\u003eTo explicitly account for the large proportion of zero counts and the pronounced overdispersion in lightning observations, a hurdle modeling framework (Mullahy et al., 1986) is adopted. The hurdle model decomposes the lightning distribution into two components: a binary hurdle component and a truncated count component. The binary hurdle component models lightning occurrence (0/1) using a binomial distribution, while the truncated count component models strictly positive lightning counts using a zero-truncated negative binomial (ZTNB) distribution.\u003c/p\u003e \u003cp\u003eThe binary hurdle component is specified as a logit-binomial model with occurrence probability π. The truncated count component is characterized by a location parameter \u0026micro;\u0026thinsp;\u0026gt;\u0026thinsp;0 and a dispersion parameter θ\u0026thinsp;\u0026gt;\u0026thinsp;0. All three parameters (π, \u0026micro;, θ) are modeled as functions of EPS predictors and are estimated independently within the hurdle framework.\u003c/p\u003e \u003cp\u003eTo avoid overfitting and ensure model parsimony, candidate predictors are selected using gradient boosting combined with stability selection. The resulting final model is estimated using Markov chain Monte Carlo (MCMC) simulation. All model fitting and inference are conducted using the \u003cem\u003ebamlss\u003c/em\u003e package in R.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec8\" class=\"Section3\"\u003e \u003ch2\u003e2.2.2 Ensemble Construction of MESO\u003c/h2\u003e \u003cp\u003eTo increase ensemble size and better represent forecast uncertainty, ensemble members are constructed using a combination of spatiotemporal neighborhood and time-lagging methods. Forecast data initialized daily at 21:00 BJT and 00:00 BJT on the following day, with a horizontal resolution of 3 km, are utilized for this purpose.\u003c/p\u003e \u003cp\u003eIn the temporal dimension, the time-lagging and neighborhood approaches are applied to generate six sets of ensemble members, corresponding to different initialization times for the same valid time and different forecast lead times for the same initialization, as illustrated in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e. In the spatial dimension, a neighborhood approach is employed in which each grid point is treated as the center and its eight surrounding grid points are included. As a result, a total of 54 ensemble members (6 \u0026times; 9) is constructed for each grid point.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003c/div\u003e"},{"header":"3 Results","content":"\u003cp\u003eIn this section, we first present the results of the boosting-based selection of nonlinear predictors for lightning occurrence probability and intensity. Next, we evaluate and compare the predictive performance of models based on REPS and REPS\u0026thinsp;+\u0026thinsp;MESO predictors. Finally, we analyze representative case studies to illustrate the impact of incorporating mesoscale information.\u003c/p\u003e \u003cdiv id=\"Sec10\" class=\"Section2\"\u003e \u003ch2\u003e3.1 Feature Importance Selection\u003c/h2\u003e \u003cp\u003eThe selected nonlinear predictors associated with the lightning occurrence probability π at a one-day lead time are summarized in Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e. In total, five predictors are retained in the REPS-based experiment, whereas seven are selected when MESO-derived predictors are included. The selected predictors are primarily related to convective precipitation (cp) and convective available potential energy (cape), which jointly characterize the potential for updraft development and the vertical growth of convective clouds. In contrast, most of the predictors selected exclusively from the MESO ensemble are associated with atmospheric circulation fields. This suggests that MESO-derived predictors provide a stronger capability for representing the physical mechanisms underlying lightning activity.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eSelected REPS/REPS\u0026thinsp;+\u0026thinsp;MESO predictors for lightning probability forecasts\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"2\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eREPS\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eREPS\u0026thinsp;+\u0026thinsp;MESO\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ev700_18.mei\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003ev500_15.mei\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003esqrt_cp_1812.me\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003esqrt_cp_1812.me\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003esqrt_cp_1812.iq\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003esqrt_cp_1812.iq\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003esqrt_cape_12.me\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003esqrt_cape_12.me\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003edoy3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003emv700_18.mei\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003emv500_1812.mei\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003emsqrt_cape_18.iq\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eThe selected predictors for lightning intensity are summarized in Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e. In the REPS-based experiment, four predictors are retained for the mean parameter \u0026micro; (mu) of the truncated count distribution, whereas only a single predictor is selected for the dispersion parameter θ (theta). The predictors associated with \u0026micro; are predominantly related to the development of strong convective systems, consistent with their role in regulating lightning flash rates. With the incorporation of MESO-derived predictors, an additional term associated with θ is selected, which is closely related to the geopotential height field at 700 hPa. This pressure level is dynamically relevant for mesoscale convection, as variations in mid-tropospheric geopotential height reflect synoptic- and mesoscale flow patterns that modulate vertical wind shear, moisture advection, and large-scale ascent. These processes play a critical role in the initiation, organization, and maintenance of convective systems, thereby influencing the variability of lightning activity.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eSelected REPS/REPS\u0026thinsp;+\u0026thinsp;MESO predictors for lightning intensity forecasts\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"2\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eREPS\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eREPS\u0026thinsp;+\u0026thinsp;MESO\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003et500_15.iq(mu)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003et500_15.iq(mu)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003elayth_1812.mei(mu)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003egh700_1812.mei(mu)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003egh700_1812.mei(mu)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003edoy4(mu)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eu700_1812.mei(mu)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003et500_1812.mei(mu)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003esqrt_cape_12.me(theta)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003esqrt_cape_12.me(theta)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003emgh700_1812.mei (theta)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec11\" class=\"Section2\"\u003e \u003ch2\u003e3.2 Model Performance Evaluation\u003c/h2\u003e \u003cp\u003eIn this section, we compare the predictive performance of Bayesian Additive Models for Location, Scale, and Shape (BAMLSS) utilizing predictors selected from REPS (BR) and from REPS\u0026thinsp;+\u0026thinsp;MESO (BRM), respectively. Both models are evaluated on the same independent test dataset to ensure a fair comparison.\u003c/p\u003e \u003cp\u003eThe predictive skill for lightning occurrence is assessed using the estimated occurrence probability π. Verification is performed using the Brier score (BS) and the area under curve (AUC) derived from receiver operating characteristics (ROC). Figure\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e shows the temporal evolution of daily mean scores over the test period, with mean values across all dates indicated in the bottom-left corner of each panel.\u003c/p\u003e \u003cp\u003eThe results show that the BRM model generally outperforms the BR model across both verification metrics. In particular, the inclusion of MESO-derived predictors leads to systematically higher AUC values, indicating an improved ability to discriminate between lightning and non-lightning events. This improvement is likely attributable to the additional mesoscale information, which provides a more refined representation of mesoscale convective processes. By better capturing key physical mechanisms such as boundary-layer convergence, local thermal inhomogeneities, and organized vertical motion, the MESO-derived predictors enable a more accurate characterization of local environmental conditions favorable for lightning initiation. Overall, the inclusion of MESO-derived predictors substantially enhances the skill of probabilistic lightning occurrence forecasts.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eIn terms of the BS, the BRM model substantially outperforms the BR model on most test days, across both low and high lightning frequencies. An exception is observed on two days characterized by extremely high lightning events, for which BR exhibits slightly better performance. This behavior indicates a reduced model skill under extreme conditions, suggesting limitations in learning and representing rare, high-impact lightning events. Such behavior may be attributed to two primary factors. First, periods of intense lightning activity are often associated with complex interactions across mesoscale and microscale processes that are inherently challenging to represent. Second, such events constitute only a small proportion of the training dataset, resulting in limited information for accurately learning the tails of the lightning occurrence distribution.\u003c/p\u003e \u003cp\u003eThe predictive performance for lightning intensity is evaluated based on the full predictive distribution of lightning counts, characterized by the parameters \u0026micro; and θ. For each day, a discrete probability mass function is obtained and evaluated using the ranked probability score (RPS; Epstein, 1969) and the logarithmic score (LogS; Wood, 2017). The temporal evolution of daily mean scores over the test period is shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e.\u003c/p\u003e \u003cp\u003eThe results indicate a clear improvement in forecast skill when MESO-derived predictors are included. The BRM model consistently outperforms the BR model across all test dates in terms of LogS and shows superior performance on most dates with respect to the RPS. In particular, the substantial improvement in LogS suggests that the incorporation of mesoscale information enhances both the sharpness and calibration of the predicted lightning intensity distributions. This reflects an improved ability to represent not only typical lightning activity but also higher-intensity events, which are especially relevant for impact-oriented forecasting.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec12\" class=\"Section2\"\u003e \u003ch2\u003e3.3 Case Study\u003c/h2\u003e \u003cp\u003eTo further illustrate the added value of mesoscale predictors, a comparative case study is conducted using forecasts based on REPS and REPS\u0026thinsp;+\u0026thinsp;MESO predictors. For both models, we analyze observed lightning counts, forecasted lightning occurrence probabilities, predicted lightning counts, and the probability of lightning counts exceeding four flashes per grid point. A representative high-impact case featuring widespread and intense lightning activity on 23 August 2021 is selected.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eForecasts based solely on REPS predictors successfully capture the large-scale spatial patterns of lightning activity. However, they tend to overestimate both the spatial extent and the occurrence probability of lightning in regions with high observed activity, resulting in an increased number of false alarms. While such behavior may be advantageous for operational early-warning purposes, it also reflects limited discrimination capability in regimes of intense convection. Furthermore, lightning intensity in the upper tail of the distribution (the 90th percentile) is generally underestimated, likely due to the coarse spatial resolution and limited representation of mesoscale processes in the REPS predictors.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eIn contrast, forecasts incorporating MESO-derived predictors exhibit a more refined representation of convective-scale structures. This leads to improved spatial localization of lightning-prone areas and a noticeable reduction in false alarms. Moreover, the inclusion of mesoscale information enhances the depiction of high-intensity lightning events, particularly in regions of strong convective development. These results highlight the added value of mesoscale predictors for capturing the structural and dynamical characteristics of severe convective systems.\u003c/p\u003e \u003c/div\u003e"},{"header":"4 Discussion and Conclusion","content":"\u003cp\u003eThe systematic evaluation conducted in this study demonstrates that lightning forecasts incorporating MESO-derived predictors consistently outperform those based solely on REPS predictors across all verification metrics. The inclusion of mesoscale information leads to notable improvements in probabilistic lightning forecasts, lightning count forecast skill, and the spatial localization of lightning-prone regions.\u003c/p\u003e \u003cp\u003eThe most pronounced improvement is observed for lightning count forecasts, where the REPS\u0026thinsp;+\u0026thinsp;MESO configuration yields substantially better verification scores. This result highlights the added value of MESO-derived predictors in providing relevant information for estimating the magnitude and variability of lightning activity. From a physical perspective, these gains can be attributed to the more refined representation of the convective environment provided by MESO-derived variables, which better capture key processes associated with convective initiation and evolution, including boundary-layer convergence, localized thermal heterogeneity, and organized vertical motion. Consequently, the model is better equipped to represent the triggering and intensification of mesoscale and microscale convective systems.\u003c/p\u003e \u003cp\u003eFrom a predictor selection standpoint, the inclusion of MESO-derived predictors not only enlarges the set of available explanatory variables but also strengthens the predictive signals of convection-related variables originally derived from REPS. The model identifies a larger number of physically interpretable and dynamically relevant features as influential predictors, particularly those directly linked to convective development. This suggests that mesoscale information enhances the effective signal-to-noise ratio of the predictor set, thereby facilitating more robust statistical learning.\u003c/p\u003e \u003cp\u003eOverall, the improved forecast skill obtained through the incorporation of MESO-derived predictors underscores the dominant role of mesoscale and microscale processes in lightning generation. Compared with the large-scale environmental background fields provided by REPS, MESO-derived predictors offer a more direct representation of local conditions conducive to convective initiation and development, thereby yielding higher discriminative ability for lightning forecasting. Looking ahead, the continued integration of higher-resolution, more physically consistent, and more accurate mesoscale NWP predictors is expected to further improve probabilistic forecasts of lightning activity and other forms of severe convective weather. In addition, future work may explore the incorporation of machine learning\u0026ndash;based feature selection and hybrid modeling approaches to more effectively identify key predictors and further enhance forecasting skill.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003eThis work supported by the Sichuan Science and Technology Program (NO.2025YFNH0006). The authors have no relevant financial or non-financial interests to disclose. All authors contributed to the study conception and design. Material preparation, data collection and analysis were performed by Haoyue Li. The first draft of the manuscript was written by Haoyue Li and all authors commented on previous versions of the manuscript. All authors read and approved the final manuscript. The datasets generated during the current study are not publicly available due to confidentiality agreements but are available from the corresponding author on reasonable request.\u003c/p\u003e\u003cp\u003e\u003cstrong\u003eCompeting interests\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eWe declare that we have no competing interests.\u003c/p\u003e\u003ch2\u003eAuthor Contribution\u003c/h2\u003e\u003cp\u003eAll authors contributed to the study conception and design. Material preparation, data collection and analysis were performed by Haoyue Li. The first draft of the manuscript was written by Haoyue Li and all authors commented on previous versions of the manuscript. All authors read and approved the final manuscript.\u003c/p\u003e\u003ch2\u003eAcknowledgement\u003c/h2\u003e\u003cp\u003eThis work supported by the Sichuan Science and Technology Program (NO.2025YFNH0006). The authors thank the anonymous reviewers for their valuable comments and suggestions, which have greatly contributed to improving the quality of this paper.\u003c/p\u003e\u003ch2\u003eData Availability\u003c/h2\u003e\u003cp\u003eThe datasets generated during the current study are not publicly available due to confidentiality agreements but are available from the corresponding author on reasonable request.\u003c/p\u003e"},{"header":"References","content":"\u003cp\u003eMa M, Tao SC, Zhu BY, et al. (2004) Climate distribution of lightning density in China and surrounding areas observed by satellites. Science China, 34(4): 298-306. Doi: 10.1360/zd2004-34-4-298\u003c/p\u003e\n\u003cp\u003eBauer P, Quintino T, Wedi N, et al. (2020) The ECMWF scalability program: progress and plans. European Centre for Medium-range Weather Forecasts: 12-14.\u003c/p\u003e\n\u003cp\u003eBetz HD, Schmidt K, Oettinger WP, et al. (2008) Cell-tracking with lightning data from LINET. Advances in Geosciences, 17: 55-61. Doi: 10.5194/adgeo-17-55-2008\u003c/p\u003e\n\u003cp\u003eBright DR, Wandishin MS, Jewell RE, et al. (2005) A physically based parameter for lightning prediction and its calibration in ensemble forecasts. Conference on Meteorological Applications of Lightning Data. \u003c/p\u003e\n\u003cp\u003eChristian HJ, Blakeslee RJ, Boccippio DJ, et al. (2003) Global frequency and distribution of lightning as observed from space by the optical transient detector. Journal of Geophysical Research: Atmospheres, 108(D1): 4005-4019. Doi: 10.1029/2002JD002347\u003c/p\u003e\n\u003cp\u003eCrane RK. (1979) Automatic cell detection and tracking. Geoscience Electronics, IEEE Transactions on, 17(4): 250-262. Doi: 10.1109/TGE.1979.294654\u003c/p\u003e\n\u003cp\u003eDixon M, Wiener G. (1993) TITAN: Thunderstorm identification, tracking, analysis, and nowcasting-a radar-based methodology. Journal of Atmospheric and Oceanic Technology, 10(6): 785. Doi: 10.1175/1520-0426(1993)0102.0.CO;2\u003c/p\u003e\n\u003cp\u003eEpstein ES. (1969) A scoring system for probability forecasts of ranked categories. Journal of Applied Meteorology and Climatology, 8(6): 985\u0026ndash;987. Doi: 10.1175/1520-0450(1969)008\u0026lt;0985:ASSFPF\u0026gt;2.0.CO;2\u003c/p\u003e\n\u003cp\u003eGeng YA, Li QY, Lin TY, et al. (2021) A deep learning framework for lightning forecasting with multi-source spatiotemporal data. Quarterly Journal of the Royal Meteorological Society, 147(741): 4048-4062. Doi: 10.1002/qj.4167\u003c/p\u003e\n\u003cp\u003eGlahn HR, Lowry DA. (1972) The use of model output statistics (MOS) in objective weather forecasting. Journal of Applied Meteorology and Climatology, 11: 1203-1211. Doi: 10.1175/1520-0450(1972)011\u0026lt;1203:TUOMOS\u0026gt;2.0.CO;2\u003c/p\u003e\n\u003cp\u003eHering AM, S\u0026apos;en\u0026apos;esi S, Ambrosetti P, et al. (2004) Nowcasting thunderstorms in the Alpine region using a radar based adaptive thresholding scheme. Proceedings of the 3th European Conference on Radar in Meteorology and Hydrology: 206-211.\u003c/p\u003e\n\u003cp\u003eHering AM, S\u0026apos;en\u0026apos;esi S, Ambrosetti P, et al. (2005) Nowcasting thunderstorms in complex cases using radar data WSN05 conference, Toulouse 5-9 September 2005. Proceedings on CD, contribution 2, 14.\u003c/p\u003e\n\u003cp\u003eJohnson JT, Mackeen PL, Witt A, et al. (1998) The storm cell identification and tracking algorithm: an enhanced WSR-88D algorithm. Weather and Forecast, 13: 263\u0026ndash;276. Doi: 10.1175/1520-0434(1998)013\u0026lt;0263:TSCIAT\u0026gt;2.0.CO;2\u003c/p\u003e\n\u003cp\u003eKlein WH. (1971) Computer prediction of precipitation probability in the United States. Journal of Applied Meteorology and Climatology, 10 (5): 903-915. Doi: 10.1175/1520-0450(1971)010\u0026lt;0903:CPOPPI\u0026gt;2.0.CO;2\u003c/p\u003e\n\u003cp\u003eKober K, Tafferner A. (2009) Tracking and nowcasting of convective cells using remote sensing data from radar and satellite. Meteorologische Zeitschrift, 18(1): 75\u0026ndash;84. Doi: 10.1127/0941-2948/2009/359\u003c/p\u003e\n\u003cp\u003eKohn M, Galanti E, Price C, et al. (2011) Nowcasting thunderstorms in the Mediterranean region using lightning data. Atmospheric Research, 100(4): 489-502. Doi: 10.1016/j.atmosres.2010.08.010\u003c/p\u003e\n\u003cp\u003eLin TY, Li QY, Geng YA, et al. (2019) Attention-based dual-source spatiotemporal neural network for lightning forecast. IEEE Access, 7: 158296\u0026minus;158307. Doi: 10.1109/ACCESS.2019.2950328\u003c/p\u003e\n\u003cp\u003eMayr A, Fenske N, Hofner B, et al. (2012) Generalized additive models for location, scale and shape for high dimensional data\u0026mdash;a flexible approach based on boosting. Royal Statistical Society, 61(3), 403\u0026ndash;427. Doi: 10.1111/j.1467-9876.2011.01033.x\u003c/p\u003e\n\u003cp\u003eMichalon N, Nassif A, Saouri T, et al. (1999) Contribution to the climatological study of lightning. Geophysical Research Letters, 26(20): 3097-3100. Doi: 10.1029/1999GL010837\u003c/p\u003e\n\u003cp\u003eMostajabi A, Finney DL, Rubinstein M, et al. (2019) Nowcasting lightning occurrence from commonly available meteorological parameters using machine learning techniques. Climate and Atmospheric Science, 2(1): 41. Doi: 10.1038/s41612-019-0098-0\u003c/p\u003e\n\u003cp\u003eMueller C, Saxen T, Roberts R, et al. (2003) NCAR auto-nowcast system. American Meteorological Society, 18: 545\u0026ndash;561. Doi: 10.1175/1520-0434(2003)018\u0026lt;0545:NAS\u0026gt;2.0.CO;2\u003c/p\u003e\n\u003cp\u003eMullahy J. (1986) Specification and testing of some modified count data models. Journal of econometrics, 33(3): 341\u0026ndash;365. Doi: 10.1016/0304-4076(86)90002-3\u003c/p\u003e\n\u003cp\u003ePrice C, Rind D. (1992) A simple lightning parameterization for calculating global lightning distributions. Journal of Geophysical Research Atmospheres, 1992, 97(D9): 9919-9933. Doi: 10.1029/92JD00719\u003c/p\u003e\n\u003cp\u003eRajeevan M, Madhulatha A, Rajasekhar M, et al. (2012) Development of a perfect prognosis probabilistic model for prediction of lightning over south-east India. Earth System Science Data, 121: 355\u0026ndash;371. Doi: 10.1007/s12040-012-0173-y\u003c/p\u003e\n\u003cp\u003eRinehart RE, Garvey ET. (1978) Three-dimensional storm motion detection by conventional weather radar. Nature, 273: 287-289. Doi: 10.1038/273287a0\u003c/p\u003e\n\u003cp\u003eSch\u0026ouml;n, C, Dittrich J, M\u0026uuml;ller R. (2019) The error is the feature: how to forecast lightning using a model prediction error. In Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining: 2979-2988.\u003c/p\u003e\n\u003cp\u003eSchultz CJ, Petersen WA, Carey LD. (2009) Preliminary development and evaluation of lightning jump algorithms for the real-time detection of severe weather. Journal of Applied Meteorology and Climatology, 48(12): 2543-2563. Doi: 10.1175/2009JAMC2237.1\u003c/p\u003e\n\u003cp\u003eSeed AW. (2003) A dynamic and spatial scaling approach to advection forecasting. Journal of Applied Meteorology and Climatology, 42(3): 381-388. Doi: 10.1175/1520-0450(2003)042\u0026lt;0381:ADASSA\u0026gt;2.0.CO;2\u003c/p\u003e\n\u003cp\u003eSimon T, Fabsic P, Mayr GJ, et al. (2018) Probability forecasting of thunderstorms in the Eastern Alps. Monthly Weather Review, 146(9): 2999-3009. Doi: 10.1175/MWR-D-17-0366.1\u003c/p\u003e\n\u003cp\u003eSimon T, Mayr GJ, Umlauf N, et al. (2019) NWP-based lightning prediction using flexible count data regression. Advances in Statistical Climatology, Meteorology and Oceanography, 2019. DOI:10.5194/ASCMO-5-1-2019\u003c/p\u003e\n\u003cp\u003eUnderwood SJ. (2006) Cloud-to-ground lightning flash parameters associated with heavy rainfall alarms in the Denver, Colorado, urban drainage and flood control district ALERT network. Monthly Weather Review, 134(9): 2566-2580. Doi: \u003c/p\u003e\n\u003cp\u003eWilliams ER. (2005) Lightning and climate: A review. Atmospheric Research, 76(1): 272-287. Doi: 10.1029/2004JD005332\u003c/p\u003e\n\u003cp\u003eWood SN. (2017) Generalized Additive Models: An Introduction with R. 2nd ed. Texts in Statistical Science Series, Chapman \u0026amp; Hall/CRC: 476.\u003c/p\u003e\n\u003cp\u003eXu MY, Qie XS, Pang WJ, et al. (2022) Lightning climatology across the Chinese continent from 2010 to 2020. Atmospheric Research, 275: 106251. Doi: \u003c/p\u003e\n\u003cp\u003eYair Y, Lynn B, Price C, et al. (2010) Predicting the potential for lightning activity in Mediterranean storms based on the Weather Research and Forecasting (WRF) model dynamic and microphysical fields. Journal of Geophysical Research: Atmospheres, 115(D4): D4205. Doi: 10.1016/j.atmosres.2022.106251\u003c/p\u003e\n\u003cp\u003eZhou XY, Geng YA, Yu HM, et al. (2022) LightNet+: A dual-source lightning forecasting network with bi-direction spatiotemporal transformation. Applied Intelligence, 52: 11147-11159. Doi: 10.1007/s10489-021-03089-5\u003c/p\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"Lightning forecast, Ensemble postprocessing, Gradient boosting, Count data regression","lastPublishedDoi":"10.21203/rs.3.rs-8550311/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-8550311/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eAccurate short-term lightning forecasting requires reliable representations of both lightning occurrence and intensity, as well as the underlying convective processes. While ensemble prediction systems (EPS) provide valuable probabilistic information, their ability to resolve mesoscale and convective-scale variability remains limited. In this study, we assess the added value of mesoscale information for probabilistic lightning forecasting over eastern China. A mesoscale ensemble is constructed from deterministic forecasts of the China Meteorological Administration (CMA) Mesoscale Model (MESO) using spatiotemporal neighborhood and time-lagging techniques and is combined with predictors from the CMA regional ensemble prediction system (REPS). Lightning occurrence and counts are modeled within a Bayesian additive model for location, scale, and shape (BAMLSS) framework, using a hurdle-based count regression to account for excess zeros and overdispersion. Influential nonlinear predictors are selected via stability selection combined with gradient boosting. Forecast performance with and without MESO-derived predictors is systematically evaluated. The results show that incorporating mesoscale information consistently improves forecast skill for both lightning occurrence and intensity across all verification metrics. These improvements are primarily associated with MESO predictors related to convective available potential energy and convective precipitation, highlighting the importance of mesoscale processes for probabilistic lightning forecasting.\u003c/p\u003e","manuscriptTitle":"Improving Probabilistic Lightning Forecasts through Ensemble Postprocessing with Mesoscale Information","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-01-20 11:30:57","doi":"10.21203/rs.3.rs-8550311/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"3e317d36-dc53-470d-a0ca-600ce9f4312d","owner":[],"postedDate":"January 20th, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2026-02-03T03:08:41+00:00","versionOfRecord":[],"versionCreatedAt":"2026-01-20 11:30:57","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-8550311","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-8550311","identity":"rs-8550311","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}