Introduction

Lightning is one of the most common natural disasters inflicting large number of casualties worldwide1,2. A Global study reported around 6000 to 24,000 lightning fatalities annually3. There are many studies describing the lightning fatality rate in different countries including India1,4,5,6,7,8,9,10,11,12. Based on the fatality rate by five major calamities (flood, landslide, cold stroke, heat stroke and lightning) during 1967–2012 in India, Illiyas et al.13 found that lightning caused maximum deaths of 39%. The maximum lightning fatality occurs in west-central, central northeast and peninsular Indian states mainly Maharashtra (29%), followed by West Bengal (12%) and Uttar-Pradesh (9%) respectively14. The dense population in these states likely exacerbates the high incidence of lightning-related fatalities. A lightning event in India on June 25, 2020, claimed over 100 lives in a single day, underscoring the catastrophic impact of such disasters. Whilst, Selvi and Rajapandian15 revealed that the deaths due to lightning in the rural areas of south India is mainly through ‘step voltage’ and ‘touch voltage’ mechanisms. Lightning injuries far exceed fatalities, causing long-term or temporary disabilities like memory loss, numbness, dizziness, etc., highlighting the need of public awareness alongside establishing an effective early warning system, as well as strategies for mitigation and preparedness.

Lightning prediction is one of the most challenging tasks in tropical regions such as India since the characteristics of the lightning activity are highly variable with various synoptic conditions across seasons over different parts of the country. The behaviour of lightning over the land, over the ocean, over different latitude belts, over different climatic situations, etc. are different16. According to the World Meteorological Organisation17, the maximum annual frequency of thunderstorm occurs in northeast India predominately Assam19. The combined lightning climatology from Lightning Imaging Sensor (LIS) on-board the Tropical Rainfall Measuring Mission (TRMM) satellite and Optical Transient Detector (OTD) on-board Orbview-1 during May 1995 - December 201419 shows that the maximum lightning activity occurs over northern (Jammu & Kashmir) and eastern (Assam, Meghalaya) states (Fig. 1). Study by Ray et al.18 under the scientific project between the South Asian Association for Regional Cooperation (SAARC STORM) on the convective activity over India during the pre-monsoon of the year 2013, most of the thunderstorms occur during the night over northeast India, evening hours over east India, and in the afternoon and evening hours over peninsular and northwest India. The contrast in convective activity between the land and the ocean are reported in many studies19,20,21,22,23,24,25,26,27,28,29. Also there is a variability in the lightning activity over Bay of Bengal (BoB) and Arabian Sea (AS). A statistical analysis of convective events (1979–2004) shows an increasing trend in central and north BoB during the pre-monsoon, in north BoB during the monsoon, and in south AS in post-monsoon season30. During 1992–2004, increased convective activity is noticed over the AS in June and July and over the BoB in July and August with a recent increase in BoB during June. The average lightning frequency during 1998–2014 obtained from TRMM satellite data is higher over the Indian landmass compared to BoB and the AS31, with hotspots along the Himalayan foothills, Indo-Gangetic Plains, coastal regions, BoB, and AS. However, the source of moisture supply and hence the characteristics and frequency of lightning are different in the aforementioned regions32. During monsoon, the main source of moisture over the southwest coast of India is from AS and mostly the rainfall occurs due to the orographic lifting of westerly winds. Mostly this is warm rain without lightning. However, during pre- and post-monsoon periods, occurrences of deep convective systems lead to lightning activity over this area. Indo-Gangetic plains experience monsoon lightning due to deep convection from moisture convergence between the BoB and the monsoon trough, enhanced by land heating. Whereas during winter, moisture supply from the western disturbances and orographic lifting leads to lightning activity33. This underscores the highly complex variability of convective activity over the Indian region, presenting a formidable challenge for lightning prediction using numerical weather prediction (NWP) models.

Fig. 1
figure 1

Climatology of total lightning flash density (flashes km− 2 year− 1) from LIS and OTD combined during May 1995 to December 2014. The states/union territories are marked on the map.

Full size image

The main components participating in the cloud electrification and charge generation are ice-phase hydrometeors and super-cooled liquid water in the presence of strong updrafts34,35,36,37,38,39,−40. There are many empirical relationships exist between the storm features and lightning. Price and Rind41 (PR92, hereafter) lightning parameterization scheme is one of the modest schemes to simulate global lightning distributions, which provides the relation between cloud top height (CTH) and lightning flash rates. Since the cloud dynamics over land and ocean are different, PR92 suggests separate empirical relationships over land and the ocean. Generally, PR92 underestimates lightning, especially over the Ocean42. Apart from PR92, there are a few more empirical relationships that connect the lightning flash counts/ flash rates with the number concentration of ice hydrometeors and ice water path43,44,45,46,47. Mohan et al.48 evaluated different lightning parameterization schemes for pre-monsoon lightning cases over Maharashtra, a state of India and found supreme accuracy for PR92. PR92 diagnoses the lightning activity from the CTH, the seasonal and regional variability of cloud structure reflects in the lightning activity. There are many studies focus on the cloud properties derived from satellite, radiosonde, and other ground-based observations6,49,50,51,52,53,54. Though the performance of PR92 is the best compared to most of the existing lightning schemes, the spatial standard deviation with observation is greater in this scheme42,55. In this scenario, the lightning schemes based on hydrometeors seem to be more convincing as lightning phenomena are microphysical in origin. Lopez lightning parameterization scheme in European Centre for Medium-Range Weather Forecasts (ECMWF) Integrated Forecasting System accounts for the existence of hydrometeors (graupel, cloud liquid water, and snow), convective available potential energy (CAPE), and convective cloud base for lightning diagnosis56 and Sarkar et al.57 reported its performance over India during pre-monsoon.

Effective lightning forecasting and warning requires balancing lead time with precision in location and magnitude. While nowcasting and limited area models with advanced microphysics provide accurate location-specific lightning forecast, they are limited to short lead times. Medium-range lightning forecasting with global models offers great value in identifying lightning-prone areas in advance, aiding preparedness and economic planning, though with less location-specific precision. Despite the availability of various lightning parameterization schemes, forecasting lightning activity over the Indian subcontinent remains a challenging task due to the diverse geographical characteristics as well as the seasonality of convective activity18,58,59. Lightning in these regions is often associated with localized deep convection or forms part of organized mesoscale convection. Hence, a parameterization approach integrating the macro and microphysical aspects from global models offers a promising solution for improvements in lightning prediction. Combining all these aspects, the present work proposes a methodology to estimate lightning threats using cloud-to-ground (CG) flashes from a global atmospheric model at a medium-range scale over the Indian domain for all seasons. This is achieved by developing a blended lightning parameterization scheme to estimate the total lightning flash counts based on revised PR92 and Lopez schemes and segregate the CG flashes from the total lightning flash counts. The PR92 scheme depends on the macro-physical properties of clouds, while the Lopez scheme contains their dynamical and the microphysical characteristics. The CG flash is most lethal, which affects the life and property directly, and are segregated from the total lightning flash counts. Color-coded threat levels are generated for CG flashes to facilitate quicker and more effective mitigation measures.

Results and discussion

The most challenging aspect of lightning prediction is forecasting the exact locations of lightning activity within the broad convective area, which is highly sensitive to the storm parameters chosen for estimation. Therefore, identifying the storm parameters that are closely associated with lightning is crucial. The Supplementary Table 1 lists 31 lightning cases from 2023 along with the affected areas and associated synoptic features. A threshold-based technique is used to modify the criteria for identifying the grid cell susceptible to electrification (referred to as storm point) based on macrophysical factors. An examination of 31 lightning cases serves as the basis for the development of storm point identification criteria and weight calculations for the blending method. All the qualitative and quantitative analysis are carried out over Indian domain extending 60.03°E − 99.93°E and 0.058°N − 39.9°N. Figure 1 shows the regions discussed in the following sections.

Requirement for a revised storm point detection criteria

The requirement of a better storm point detection criterion is described in this section through the analysis of four selected cases from different seasons- (i) winter: 24th Jan 2023, (ii) pre-monsoon: 07th Apr 2023, (iii) monsoon: 28th Jul 2023 and (iv) post-monsoon: 22nd Nov 2023. The observed lightning activity during these days are depicted in Fig. 2a,d,g,j from lightning location network (LLN). Under the influence of the synoptic features mentioned in Supplementary Table 1, a widespread lightning during the winter case is observed over peninsular India (Fig. 2a) and northern parts of India; isolated few lightning activities over west-central India. The lightning occurred in this case over peninsular India found to be more active than over northern India. The pre-monsoon case (Fig. 2d) developed due to trough in easterlies over lower troposphere from peninsular India to west-central India. During the monsoon case (Fig. 2g), thunderstorms occurred in isolated places in southern India and hailstorms occurred over west-central India. The western end of the monsoon trough was in its climatological normal position whereas the eastern end was displaced southward, resulting widespread lightning over central and north-west. In the post-monsoon case (Fig. 2j), widespread lightning was associated with a trough in easterlies extending from Maldives to south India. This period is influenced mostly by tropical disturbances and residual convection from retreating monsoon. There was a widespread lightning on 28th Jul 2023 and localized and intense on 7th Apr 2023. These two cases represent the typical characteristics of these seasons. The relative humidity (RH) peaks during monsoon and is lowest in pre-monsoon, with a large spatial variability60. Khan et al.60 confirms the region-wise variability of CAPE follows that of RH. Increased RH reduces the lifting condensation level and level of free convection, increasing CAPE and driving strong moist convection31. Murugavel et al.61 reported that over India CAPE and lightning activity are more correlated during the monsoon season than pre-monsoon.

Fig. 2
figure 2

Spatial distribution of 24 h accumulated lightning flash counts from (a, d, g, j) LLN, and estimated lightning using (b, e, h, k) PR92 scheme and Rev_PR92 scheme (c, f, i, l) for the lightning cases from different seasons – (a, b, c) winter, (d, e, f) pre-monsoon, (g, h, i) monsoon, and (j, k, l) post-monsoon.

Full size image

The flash counts of these cases are estimated from the Global NCMRWF Unified Model (NCUM-G) following PR9241 and Lopez56 lightning parameterization schemes and they are further modified with a revised storm point detection criteria (Figs. 2 and 3). The existing storm identification criterion in the PR92 scheme of NCUM-G is based on cloud thickness42. According to this criterion, the lightning estimation will be done only where the cloud thickness is > 5 km. The electrification in thunderstorms occurs within the mixed-phase region, typically situated between the altitudes corresponding to temperatures of -10 °C and − 40 °C. A cloud depth of at least 5 km ensures the presence of this region, facilitating interactions among supercooled water droplets, ice crystals, and graupel particles, which are essential for charge separation leading to lightning42. However, it may not be a robust criterion for the Indian region where the nature of the convective activity is highly variable with the seasons, which is revealed from the observed flash count distribution of these four selected cases (Fig. 2a,d,g,j). These cases showcase a widespread lightning during monsoon, whereas comparably isolated activity during pre-monsoon and weak and less spatially spread activity during winter and post-monsoon. The existing PR92 scheme in NCUM-G shows large false alarms over the land as well as over the Ocean (Fig. 2b,e,h,k) with overestimated flash counts except in the pre-monsoon case, indicating the storm points are not identified effectively with this criterion. To address the limitations of this criteria, the simulated cloud thickness at peak lightning activity (Supplementary Fig. 1) is compared with the observed activity. Evidently, the spatial spread of the cloud depth is contributing to the false alarms in lightning estimation both over land and Ocean, indicating the need for further refinement. Based on Stolz et al.62 and Lynn and Yair63 vertical velocity is incorporated as a tuning parameter, alongside the cold cloud thickness threshold suggested by Yoshida et al.64 (Rev_PR92, hereafter). The modified algorithm, incorporating the revised storm point identification criteria, is shown in Fig. 4. Rev_PR92 improved the spatial distribution of the flash counts across all seasons. Remarkably, it enhanced the flash counts and reduced the spatial spread for the pre-monsoon case (Fig. 2d,e,f), indicating reduced the missed events and false alarms.

Fig. 3
figure 3

Same as Fig. 2 but for Lopez scheme.

Full size image
Fig. 4
figure 4

Schematic of revised schemes (a) PR92 and (b) Lopez.

Full size image

Similarly, the estimation of lightning flash counts with the Lopez scheme without any additional storm identification criterion produced a large spatial extent of the lightning activity (Fig. 3). Compared to PR92 and Rev_PR92, Lopez shows a larger flash extent density. Same as in Rev_PR92, the spatial spread and overestimation can be reduced further by adapting an additional storm identification criterion. Thus, it is further tuned with convective vertical velocity – a property that is closely related to the dynamics of the deep convective systems. Deierling and Petersen65 reported the strong association between this parameter and lightning flashes. The highlight of this revised scheme (Rev_Lopez, hereafter) is that over land and Ocean the thresholds for the storm identification are different (Fig. 4b). With this revised approach, the lightning estimations are improved considerably in terms of both spatial spread and magnitude (Fig. 3c,f,i,l). The section Data and Methods details the modifications made in Rev_PR92 and Rev_Lopez schemes. The idea of blending of these schemes (RPLB, hereafter) came across from the understanding that these two schemes include the micro and macro cloud characteristics, dynamics and thermodynamics of the convective system, and blending may enable them to balance their shortcomings with each other.

Qualitative assessment of RPLB scheme

The selected lightning cases are associated with - cyclonic circulations during monsoon, induced cyclonic circulations during western disturbances (WDs), troughs/ wind discontinuity, etc., indicating the test cases associated with a wide variety of synoptic conditions occurred over different topographic regions. This makes this study more robust in assessing the model performance.

Spatial distribution of lightning flash counts

Both schemes effectively captured the 24 h accumulated lightning patterns (Supplementary Fig. 2). However, during the winter case, Rev_PR92 represented the lightning events in the northern regions well, while Rev_Lopez overestimated them (Supplementary Fig. 2b,c); vice versa in the post-monsoon case (Supplementary Fig. 2n,o) relative to LLN (Supplementary Fig. 2a). In pre-monsoon case, both schemes underestimated the flash counts (Supplementary Fig. 2f,g), although the events were captured except along the west coast with respect to the observed flash counts (Supplementary Fig. 2e). A similar pattern is noticed for other pre-monsoon cases also. The flash counts on 28th Jul 2023 are overestimated in Rev_PR92 than Rev_Lopez (Supplementary Fig. 2j,k) and LLN (Supplementary Fig. 2i). A large spatial spread is observed along the Himalayan ranges, particularly in Rev_Lopez, with consistent false alarms over the elevated terrain of northeast India and found to be a systematic bias of the model.

The analysis of spatial features shows Rev_PR92 overestimates the flash counts over southern India during monsoon and post-monsoon than LLN (Supplementary Fig. 2m). On the contrary, Rev_Lopez mostly overestimates the events over north India and along the Himalayan region. Even though these schemes are good at capturing the spatial patterns, the limitations of these schemes arise due to the nature of the predictors used such as Rev_PR92 which has more contributions from cloud macro-physical property, and Rev_Lopez scheme which has more contributions from microphysics and CAPE. So the limitations of either Rev_PR92 or Rev_Lopez can be minimized by blending them with season-specific weightage. The blending methodology is illustrated in the Data and Methods section. Supplementary Fig. 2d,h,l,p show the spatial distribution of flash counts from RPLB, with distinct differences in spatial spread and magnitudes from revised schemes. The events over northern India is overestimated by Rev_Lopez (Supplementary Fig. 2c) in winter case, while isolated events over west-central India are missed by Rev_PR92 (Supplementary Fig. 2b). Additionally, lightning events over the southern peninsula are overestimated by Rev_PR92. These deviations in contributing schemes of RPLB is well balanced, effectively marking the active regions with relatively higher flash counts (Supplementary Fig. 2d). Similarly, the missed events in Rev_Lopez over west-central India in pre-monsoon case (Supplementary Fig. 2g) are well represented in RPLB (Supplementary Fig. 2h) due to the weighted contributions from Rev_PR92 (Supplementary Fig. 2f). The overestimated flash counts in Rev_PR92 in monsoon case (Supplementary Fig. 2j) and post-monsoon case (Supplementary Fig. 2n) are reduced significantly in RPLB (Supplementary Fig. 2l,p) through the appropriate selection of percentage contributions from Rev_PR92. The spatial biases in the estimated lightning events are primarily derived from the global forecast data used. Overall, the estimated flash counts in RPLB are more comparable with the LLN. The overestimations and misses are considerably reduced however false alarms are still there. The blended scheme shows significant improvement in lightning forecasts compared to the individual revised schemes.

Quantitative assessment of RPLB scheme

Frequency distribution of lightning flash counts

Understanding the distribution of flash counts in different flash count ranges (bins) is an important aspect. This reveals the capability of this estimation method to represent the severity of the lightning event. Figure 5 shows the average number frequency of estimated flash counts from Rev_PR92, Rev_Lopez, and RPLB over the active regions from the selected lightning events along with the distribution from LLN. The average number frequency of flash counts on the y-axis is shown on a logarithmic scale, ranging up to 105, indicating a wide range of lightning activity.

Fig. 5
figure 5

Frequency distribution of lightning flash counts across the seasons observed from LLN (bars), and estimated from Rev_PR92 (blue circles), Rev_Lopez (red circles), and RPLB (green circles) schemes averaged over the cases selected from different seasons. The different colors of bars represent different bins.

Full size image

Considering the frequency of the flashes shown, pre-monsoon, and monsoon are the most active seasons for lightning. Post-monsoon and winter are comparably less active. All seasons have a large frequency of occurrences of lightning in the range 0–1. The revised schemes and RPLB could represent elevated number frequency in the lower bins as well as for the bins ranges 200–300. In the winter cases, Rev_PR92 follows the observation well till the bin 20–30 and overestimates largely thereafter. However, Rev_Lopez overestimates throughout all the bins. The blended scheme effectively maintains the distribution, though there is an overestimation in all bins. A similar nature is noticed in the post-monsoon season. In pre-monsoon and monsoon seasons Rev_Lopez is overestimating the number frequency, whereas Rev_PR92 is comparably good up to the range 100–200 and underestimates thereafter. In these two seasons, Rev_Lopez is closer to the observed distribution corresponding to the flash counts > 200 and blending shows considerable underestimation from the range 200–300 to higher. The frequency of occurrences of lightning from the blended scheme is compared with the distribution of LLN (Supplementary Fig. 4) and found that almost all the lightning bins except 0–1 overestimated. The maximum overestimation found in the bin 1–10 with nearly ~ 9%, whereas overestimation in all other bins (10–20 to 500–1000) is limited to 0-2.5%.

The performance of the RPLB scheme across all 31 cases is evaluated using grid-to-grid metrics from contingency table66 (Supplementary Fig. 3), including the probability of detection (POD), miss rate (MR), false alarm ratio (FAR), and equitable threat score (ETS). The median POD of 0.6 reflects the ability of RPLB to capture the hit events, though the POD of individual cases ranges from 0.2 to ~ 0.95. The interquartile range of POD (~ 0.45- ~0.69) indicates a good agreement with the observed events. The miss rate of RPLB estimated lightning flash counts varies between 0.3 and 0.52 corresponding to lower and upper quartiles respectively. A median of 0.39 indicates a reduced miss rate. FAR values indicate a narrow distribution ranging from 0.58 to 0.7 within the lower and higher quartiles respectively with the median of 0.62. This infers a moderate and consistent false detection. Similarly, ETS values also spread across a short range (0.21–0.31) with a median of 0.29, quantifying this scheme’s ability to capture hits considering the random chances.

Fractions skill score analysis

The fractions skill score67 (FSS) analysis quantifies the model performance with different thresholds, by calculating the scores using various neighbourhood lengths. Figure 6 displays the FSS calculated over each active region along with the mean of all those cases in each season for different flash count thresholds with different neighbourhood lengths from Rev-PR92, Rev_Lopez and RPLB. The error bars show the spread of FSS from individual active regions and the mean as a line with markers. Across all seasons and thresholds, FSS improves as the neighbourhood lengths increase; indicating larger neighbourhoods allow for a better spatial match between the modelled and observed data. RPLB demonstrates higher FSS values for thresholds exceeding 1 and 10 during pre-monsoon (Fig. 6a,b), while Rev_Lopez shows the lowest for all thresholds (Fig. 6a–e). Both Rev_Lopez and RPLB exhibit a sharp decline in FSS beyond the threshold 50. During the monsoon (Fig. 6f–j), Rev_Lopez performs better, achieving higher FSS for both lower thresholds (1) and thresholds beyond 100. RPLB closely aligns with Rev_PR92, highlighting the substantial contribution of Rev_PR92 to RPLB for this season. In post-monsoon, the revised schemes show almost the same mean FSS for threshold 1 and 200 (Fig. 6k,o). In winter (Fig. 6p–t), RPLB shows larger FSS for the threshold till 50 and Rev_Lopez thereafter specifically for larger neighbourhood. RPLB and Rev_Lopez present almost same FSS values for the thresholds 100 and 200 until spatial resolution of 144 km and diverges thereafter. This analysis highlights that mostly Rev_PR92 experiment performs better than Rev_Lopez across all seasons and thresholds with larger mean FSS values. Interestingly, the broader ranges of FSS values of Rev_PR92 from individual lightning cases leads to a higher mean FSS. However, the increasing steepness of FSS with neighbourhood length suggests a larger spatial shift or less accurate in finer resolutions. On the contrary, a smaller slope in the FSS curve infers a consistent error across the scales. The steepness increases with an increase in neighbourhood rapidly for thresholds 1 and 50 and gradual increase thereafter indicating the difficulty in estimating the larger lightning thresholds. This can be either due to the lower spatial and temporal resolution of the model fields chosen for the lightning flash count estimation or the spatial shift of the simulated weather event. RPLB shows higher FSS for the thresholds 1 and 10 for all seasons. It produces either highest or intermediate mean FSS compared to Rev_PR92 and Rev_Lopez. Post-monsoon and monsoon seasons demonstrate higher mean FSS by all schemes with threshold 1, suggesting the storm identification criteria perform well during this period. The frequency distribution aligns with this as well. The better performance of RPLB ensures the percentage contributions from revised schemes holds well for all seasons.

Fig. 6
figure 6

Fractions skill score calculated for pre-monsoon (a–e), monsoon (f–j), post-monsoon (k–o) and winter (p–t) over the active lightning regions from Rev_PR92 (green), Rev_Lopez (blue) and RPLB (red) schemes for the thresholds: 1 (a, f, k, p), 10 (b, g, l, q), 50 (c, h, m, r),100 (d, i, n, s), and 200 (e, j, o, t). The error bars show the extent of FSS values for all the cases and the mean shown with a solid line with markers.

Full size image

Segregation of cloud-to-ground flashes

The previous sections discussed the performance of RPLB for different seasons considering the total lightning flashes which comprises of intra/inter cloud (IC) and CG. As the hazards arise particularly from CG, segregating them is crucial for warning purposes. The segregated CG flashes from RPLB are validated with the LLN observations (Fig. 7). In winter, sparse CG lightning activity mostly confined over southern India and scattered activity over central and northern India (Fig. 7a). Though small, these events are detrimental. RPLB captures the general pattern well. Winter lightning (total) over northern parts are often linked to localized convections due to WDs and the underestimation suggests difficulties in capturing these localized convections by the global model. In the pre-monsoon case, much more intense CG lightning is observed (Fig. 7b), particularly in west-central India and along the Western Ghats. This season is characterized by strong localized convection due to surface heating. RPLB captures the broad spatial pattern but significantly underestimates the CG flash counts, a consistent issue in this season. During the monsoon, widespread CG discharges occur along the monsoon trough from north BoB to northwest India, with high counts in central to eastern regions. The model captures the overall spatial extent of lightning activity with significant underestimation. Taori et al.68 observed that CG lightning occurrences peak during the August–September months across six homogeneous zones in India, with a secondary smaller peak during March–April. The minimum occurrences were recorded in December–January. The post-monsoon events are influenced by tropical disturbances and residual convection from the retreating monsoon. In this case, the CG lightning activity is observed mainly over southern India. RPLB captures this event well with some underestimations.

Fig. 7
figure 7

Spatial distribution of 24-h accumulated CG flash counts from LLN (a–d), RPLB scheme (e–h) and the corresponding threat levels (i–l) for the cases from different seasons viz., winter: 24 Jan 2023 (a, e, i), pre-monsoon: 07 Apr 2023 (b, f, j), monsoon: 28 Jul 2023 (c, g, k) and post-monsoon: 22 Nov 2023 (d, h, l).

Full size image

Overall, RPLB shows an underestimation of CG flash counts, especially in the seasons with strong convective activity such as the pre-monsoon and monsoon periods. It is possibly due to the missed convective activities occurring at frequencies higher than the lightning estimation time interval.

RPLB-CG threat product

Lightning products with varying threat levels are essential for effective forecast warnings and mitigation, enabling timely and appropriate action. Figure 7i–l illustrate the distribution of CG from RPLB across four distinct seasons, categorized into different threat levels based on percentile ranges of flash counts. The different categories are “No Threat, Very Low, Low, Moderate, High, and Extreme”. In winter, there are a few lightning events in the southern part of India, where the model categorizes the lightning threat as Moderate to High (Fig. 7i) and some small pockets in northern India with Low to High threat. RPLB marked the active regions of pre-monsoon case (Fig. 7j) as High threat. In the monsoon case (Fig. 7k), most of the activity is confined to central and northern India and it is categorized as Low to Moderate, with some pockets of High threat along the monsoon trough. The post-monsoon season (Fig. 7l) shows localized lightning, particularly over southern India and it is marked with Low to High over southernmost states. Overall, RPLB captures the seasonal pattern well.

In general, the threat categorization offers a valuable perspective on the forecasting/warning of lightning. It effectively captures the seasonal patterns, reflecting atmospheric processes driving the convection. The pre-monsoon and monsoon seasons present higher threats with broader activity, while winter and post-monsoon periods exhibit more localized events.

Verification of RPLB scheme with lead time

This scheme is further verified for the pre-monsoon case with a lead time upto 5 days. The observed total and CG (Fig. 8a,e) flashes are compared with total and CG (Fig. 8b–d,f–h) threats from RPLB. The main lightning activity on this day occurred over west-central and southern parts. Model estimation with 1-day lead time (Fig. 8b) shows more lightning activity over central India, signifying a large spatial shift and considerable underestimation. There are some false alarms over western and eastern India. The Day3 (Fig. 8c) is similar to Day1, but with a larger magnitude and fewer false alarms. On the contrary, Day5 shows weaker lightning activity with more missed events over southern India. The verification of CG threat levels for this case (Fig. 8f–h) with 1 and 3-day lead time indicate a High category of lightning threat over west-central India (Fig. 8f,g), whereas with 5-day lead time the signals are weakened (Fig. 8h). Day5 forecast still shows some indication of this activity with Moderate to High threat over isolated regions.

Fig. 8
figure 8

Spatial distribution of 24-h accumulated lightning flash counts valid on 07th Apr 2023 from (a) LLN, and forecast from RPLB for lead time (b) 1-day, (c) 3-day, and (d) 5-day. The bottom panel shows the accumulated CG flash counts from LLN (e) along with the estimated threat levels from RPLB with lead time (f) 1-day, (g) 3-day and (h) 5-day.

Full size image

Summary

The lightning flash count estimations from NCUM-G model using existing PR92 and Lopez lightning parameterization schemes show a large overestimation and false alarms. There is a large spatial variation in the characteristics of the convective activity across India due to diversity in the geography. Thus, the same criterion for the detection of storm points may not work for all seasons. This highlights the mandate of having different storm point detection criteria for each season. Here, we developed a lightning estimation system applicable for all seasons using the diagnostic parameters from the NCUM-G model.

Towards this aim, the storm point detection criteria for the existing PR92 scheme is revised by pre-conditioning the CTH data with a dynamic parameter, say, vertical velocity. Since the lightning flash rate estimation in PR92 is a function of cloud top height alone, more robust estimations can be gained if microphysical parameters are used. The Lopez scheme estimates flash density as a function of different microphysical components, CAPE which is a thermodynamical parameter, and convective cloud base height. As in the existing PR92 scheme, Lopez also showed broad lightning activity. The convective vertical velocity is used to tune the storm points in the Lopez scheme. The revised schemes applied to the selected lightning cases from the year 2023 and significantly improved the spatial distribution of lightning.

Considering the limitation of PR92 which displays the same lightning flash rates for the same cloud top height values irrespective of the distribution of the hydrometeors within the convective column, blending with the Lopez scheme may correct the estimation to a larger extent. The seasonal contributions corresponding to each scheme are quantified and applied to blend these schemes. The qualitative and quantitative analysis provides the usefulness of the RPLB scheme across all seasons. The bin-wise frequency distribution suggests that the RPLB is good in capturing the lower flash count bins compared to higher bins, especially in monsoon and pre-monsoon though there are some over-estimation. The higher lightning flash count bins are well captured in monsoon and pre-monsoon, with slight under-estimations. Though the spatial representation of the lightning activity in RPLB is considerably good, the underestimation of the flash counts impacted on the drastic reduction in the FSS values after threshold 10, except pre-monsoon where RPLB performs good till threshold 50. The CG flashes are segregated from the total lightning produced by RPLB and found that the spatial distribution of the CG activity is well captured. Although the CG flash counts are underestimated, the threat levels defined could highlight the areas of possible severity, serving the purpose of effective warning and mitigation. All these analyses highlight the effectiveness of the storm point identification criteria with separate treatment over land and ocean. The primary limitation of the current method is the occurrence of false alarms over highly elevated areas, which requires further investigation. Additionally, the overestimation of lightning events during the post-monsoon and monsoon seasons needs detailed analysis, particularly in relation to fundamental meteorological variables derived from model forecasts.

Some errors in this diagnostic method rooted from the limited temporal resolution of the model data, currently taken hourly for lightning flash count estimation. This may lead to potential loss of critical convection details. The accuracy may improve, if the diagnosis done at every model time step, providing finer temporal resolution and reducing potential data loss by implementing this scheme within the model. In view of the advancements in the development of the medium-range weather forecasting models with advanced physics parameterization schemes may improve the representation of convective activity and thereby enhance the performance of the current approach. Additionally, improved computational resources allow high-resolution models, which further improve the convective scale simulations and thereby the lightning estimations. In short, the current approach of estimating lightning threat is in line with the developmental activities of the model and in future it may lead to the development of a new lightning parameterization scheme. This framework is highly applicable to tropical regions and holds immense potential to transform lightning risk management in areas frequently threatened by severe lightning activity.

Data and methods

Lightning climatology from LIS OTD

The gridded lightning climatology is analysed using the combined data from LIS and OTD on-board TRMM satellite during the period May 1995 to December 2014 with a resolution of 0.5° × 0.5° to elucidate the most lightning prone area over India. This data set is available online (https://ghrc.nsstc.nasa.gov/pub/lis/climatology/) from the NASA EOSDIS Global Hydrology Resource Center Distributed Active Archive Center.

Lightning location network (LLN) data

To compare the simulated lightning flashes, data from the LLN which is installed and maintained by IITM, is used. LLN network data provides both IC and CG. It comprises nearly 85 Earth Networks Lightning Sensors (ENLS) all over India with a detection efficiency of 90–95% for CG and ~ 50% for IC69. Different types of lightning are recorded while operating in different frequency ranges viz., low frequency (1 kHz) is used for the detection of CG, middle range frequencies (1 kHz to 1 MHz) for locating return strokes, and the highest frequencies (1 MHZ to 12 MHz) used for detecting in-cloud pulses. The electromagnetic energy emitted from the lightning is recorded by ENLS and sends the waveforms to the central lightning detection server. By correlating the waveforms from the sensors that detected the lightning strokes, the time of arrival is estimated. Further, the peak current and location of the activity (latitude, longitude, and altitude) are determined by using the signal amplitude and the arrival time. The strokes are considered as a flash when they are within 700 milliseconds and 10 km. A flash is considered as CG when at least one return stroke is included. These lightning instances are then gridded into equivalent model grids, which are used for the validation of the lightning flash counts that are diagnosed from the model.

Global NCMRWF unified model (NCUM-G): brief description

NCMRWF operational global NWP system, NCUM-G, is adapted from UK-Met Office, consists of ENDGame (Even Newer Dynamics for General atmospheric modelling of the environment70 dynamical core which uses semi-implicit semi-Lagrangian formulation for solving non-hydrostatic atmospheric dynamics. The science configuration used here is of standard Global Atmosphere 7.0 (GA7.0) and Global Land (GL7.0) which is given detailed in Walters et al.71. The revised mass flux convection scheme of Gregory and Rowntree72 and a single-moment scheme based on Wilson and Ballard73 with extensive modifications for cloud microphysics are the important physics component considered in the GA7, sensitive to the scheme used for the estimation of lightning. The horizontal resolution of the model is ~ 12 km and has 70 vertical levels up to 80 km. The initialization of model is through a continuous 6 hourly-cycling incremental hybrid-4D-Var data assimilation system using the available observation distributed over that time window74. Model time step of the NCUM-G is 5 min and it delivered forecast for the next 240 h.

Lightning parameterization schemes

Given the large differences in the characteristics of the convection as well as cloud properties across the seasons over India, the forecast of lightning which is microphysical in origin, is a tedious task. One of the limitations of the existing scheme in NCUMG is that it produces the same lightning flash counts for the same CTH, irrespective of the underlying hydrometeor distribution/interaction. Thus it fails to represent the realistic lightning activity over the Indian region over different seasons. The present study involves interlacing two independent lightning parameterization schemes viz., PR92 and Lopez scheme to optimize a lightning scheme for India for all seasons.

Price and Rind (1992) lightning parameterization scheme

The lightning flash rates are calculated using CTH. The formulation is different for land and ocean points41.

$$\:{f}_{l}=3.44\times\:{10}^{-5}\times\:{CTH}^{4.9}$$
(1)
$$\:{f}_{o}=6.2\times\:{10}^{-4}\times\:{CTH}^{1.73}$$
(2)

where f represents the flash rates, and the subscripts l,and o denote the land and ocean points. CTH represents the CTH in kilometres.

Lopez lightning parameterization scheme

Lopez scheme56 is the lightning parameterization scheme implemented in the ECMWF forecasting system. It defines the lightning flash density fT as a function of charging rate (QR), CAPE, and convective cloud base height (zbase). It is calculated as,

$$\:{f}_{T}=\propto\:{Q}_{R}\sqrt{CAPE}{\left({z}_{base},1800\right)}^{2}$$
(3)

Where α is a constant and it is given as 32.4. The QR is defined as,

$$\:{Q}_{R}={\int\:}_{{z}_{0}}^{{z}_{-25}}{q}_{graup}\left({q}_{cond}+{q}_{snow}\right)\underset{\_}{\rho\:}dz\:$$
(4)

Where z0 and z− 25 are the heights in meters for the 0 °C and 25 °C isotherms. qgraup, qcond, and qsnow are the mixing ratios of graupel, cloud liquid water, and snow respectively (kg/kg). The qgraup and qsnow are calculated from the frozen precipitation convective flux (Pf in kg m− 2 s− 1) using the following Eq. 

$$\:{q}_{graup}=\beta\:\frac{{P}_{f}}{\underset{\_}{\rho\:}{V}_{graup}}$$
(5)
$$\:{q}_{snow}=\left(1-\beta\:\right)\frac{{P}_{f}}{\underset{\_}{\rho\:}{V}_{snow}}$$
(6)

Revised PR-Lopez blended (RPLB) scheme and CG threat product

These schemes were tested as they are and resulted in significant spatial overestimations. To refine them for the Indian region, dynamical variables such as vertical winds (for PR92) and convective vertical winds (for Lopez) were incorporated. Given the seasonal variations in convective activity, different thresholds were applied to further refine lightning estimations. The schematic diagram (Fig. 4a) illustrates the flow of the lightning estimation algorithm for the PR92 and Lopez lightning parameterization schemes. The storm points are identified in the revised PR92 based on the column maximum vertical velocity and cold cloud thickness (CCD-cloud thickness corresponding to temperatures between 0 0C to -40 0C). The condition for column maximum vertical velocity is set to > 0.25 ms− 1 during October-April and > 0.5 ms− 1 during May-September. The CCD is limited to 5.5 to 12 km. Further, the lightning flash rates are estimated by following Luhar et al.42. Similarly, the Lopez scheme also revised by modifying the storm detection criteria. The algorithm is shown in Fig. 4b. Here, the storm points are identified based on column maximum convective vertical velocity (Wconv) and have different thresholds over land and ocean following Deierling and Petersen65: (i) during October to May: Wconv>0 ms− 1 over land and < 100 ms− 1 over the Ocean, (ii) during June to September: Wconv>5ms− 1 over land and < 40 ms− 1 over Ocean. The flash densities are calculated over these points. The flash counts are estimated from both of these schemes and blended. The RPLB is a linear combination of the flash counts from Rev_PR92 and Rev_Lopez scheme.

$$\:Total\:FC=\left({w}_{pr}\times\:{FC}_{PR}\right)+\left({w}_{lopez}\times\:{FC}_{Lopez}\right)$$
(7)

Where, wpr and wlopez are the weightages from Rev_PR92 and Rev_Lopez respectively. These weights are assigned differently for each season using a regression-based approach. The regression analysis, applied between LLN observations and the estimated lightning from the revised schemes, was involved in optimizing appropriate weights for each scheme across the seasons. On any given day, lightning events can occur over different parts of the country, driven by distinct large-scale or synoptic weather systems. Each active region is considered separately for the regression analysis to account for variations in cloud dynamics. Hence, approximately 64 lightning samples are analysed to obtain the weights for the schemes. Moreover, the lightning events over land and ocean are examined separately to incorporate the effect of their differing dynamics. Table 1 shows that the contribution of Rev_PR92 is more for all seasons except pre-monsoon as compared to Rev_Lopez for land points and vice versa for ocean points. FCPR and FCLopez are the total flash counts from Rev_PR92 and Rev_Lopez respectively. From this, IC and CG flashes are segregated based on CCD75.

Table 1 Percentage contribution towards the total lightning flash counts from PR92 and Lopez schemes over land and ocean across different seasons.
Full size table
$$\:P={\left(\left(0.021\times\:{CCD}^{4}\right)-\left(0.648\times\:{CCD}^{3}\right)+\left(07.49\times\:{CCD}^{2}\right)-(36.54\times\:CCD)+64.09\right)}^{-1}$$
(8)

where, P is the proportion of CG flashes, CCD is the cold cloud depth, which is assumed to occur between the levels of temperatures between 0 °C to -40 °C and this is limited between 5.5 and 14 km. The CG flash counts are estimated from total lightning flash counts (total) as follows,

$$\:CG\:flash\:counts=\frac{total}{P+1}$$
(9)

The final product is a CG lightning threat product that follows a color-coded map, which displays the CG forecast in different threat categories: “No Threat, Very Low, Low, Moderate, High, and Extreme”. This map is prepared by categorizing the CG flash counts in different ranges of percentile values (Table 2). This map projects the forecast in terms of the necessity for quick action.

Table 2 Percentile ranges applied to categorise the CG flashes to threat levels.
Full size table