Seasonal forecasting of lightning and thunderstorm activity in tropical and temperate regions of the world

Thunderstorms are convective systems characterised by the occurrence of lightning. Lightning and thunderstorm activity has been increasingly studied in recent years in relation to the El Niño/Southern Oscillation (ENSO) and various other large-scale modes of atmospheric and oceanic variability. Large-scale modes of variability can sometimes be predictable several months in advance, suggesting potential for seasonal forecasting of lightning and thunderstorm activity in various regions throughout the world. To investigate this possibility, seasonal lightning activity in the world’s tropical and temperate regions is examined here in relation to numerous different large-scale modes of variability. Of the seven modes of variability examined, ENSO has the strongest relationship with lightning activity during each individual season, with relatively little relationship for the other modes of variability. A measure of ENSO variability (the NINO3.4 index) is significantly correlated to local lightning activity at 53% of locations for one or more seasons throughout the year. Variations in atmospheric parameters commonly associated with thunderstorm activity are found to provide a plausible physical explanation for the variations in lightning activity associated with ENSO. It is demonstrated that there is potential for accurately predicting lightning and thunderstorm activity several months in advance in various regions throughout the world.

here for the study region from 35°N to 35°S within the field of view of both the satellite sensors. Table 1 shows the percentage of the study region that has a significant correlation (at the 95% confidence level) between lightning flash density and seven indices representing the different large-scale modes of variability: NINO3. 4 In every one of the four seasons throughout the year, NINO3.4 is significantly related to lightning flash density at more locations than any of the other six indices (Table 1). Significant correlations for NINO3.4 occur at 15-24% of the total study region, depending on the season, with the largest value occurring during the SON season (24%). For DMI, significant correlations occur at 5-19% of the study region, with the largest value occurring during the SON season (19%). Other examples include significant correlations at 12% of the study region for PNA during the DJF season and at 11% of the study region for SAM during the SON season. Although there are some locations where NAM, NAO and QBO are significantly correlated with lightning flash density, ranging from 3-8% of the study region, it is noted that these values are relatively small given that 5% of the study region on average could be expected to have a significant correlation at the 95% confidence level based on random chance.
There are relatively few locations where lightning flash density is not significantly related to NINO3.4 but is significantly related to one of the other indices, with this occurring for 9% or less of the study region in all cases shown in Table 1. For the example of the DMI index during the SON season, 19% of the study region has a significant correlation, but only 5% of the study region has a significant correlation at locations that are different to the case for NINO3.4. This could be expected to some degree due to the strong relationship between DMI and NINO3.4 during the SON season (r = 0.68), as shown by the seasonal correlations between the seven different modes of variability listed in Supplementary Table 1. Many of the other indices show a considerably higher degree of correlation with each other: NAM and NAO is the pair of indices that are most closely related to each other, with r ≥ 0.65 in every individual season, noting that previous studies have also shown strong similarities between these two modes of variability 44 . Spatial and seasonal variability of the relationships. ENSO has the strongest and most widespread relationship to lightning activity of the seven modes of variability examined here, hence hereafter the focus of this study is primarily on this relationship. Large positive anomalies of the NINO3.4 index are typically associated with El Niño, the warm phase of ENSO, characterised by above average sea-surface temperatures in the eastern Pacific and higher air pressures in the western Pacific than the eastern Pacific 43 anomalies of the NINO3.4 index are typically associated with La Niña, the cold phase of ENSO, characterised by below average sea-surface temperatures in the eastern Pacific and higher air pressures in the eastern Pacific than the western Pacific. Spatial maps of the correlation between NINO3.4 and lightning flash densities throughout the study region are shown individually for each of the four seasons in Fig. 1. The correlations are shown for locations where the relationship is significant at the 95% confidence level.
Considerable variations occur throughout the year in the relationship between ENSO and seasonal lightning activity. The NINO3.4 index is significantly related to lightning activity at 53% of all grid-point locations for one or more seasons throughout the year, noting that about 19% of the study region on average could be expected to have a significant correlation for at least one or more seasons at the 95% confidence level based on random chance. 50% of land locations and 54% of ocean locations have a significant correlation between lightning activity and NINO3.4 for one or more seasons, noting that about two thirds of the study region is oceanic. For land regions, a significant positive correlation occurs for one or more seasons at 35% of locations, and a significant negative correlation occurs at 18% of locations. Similarly for ocean regions, a significant positive correlation occurs for one or more seasons at 37% of locations, and a significant negative correlation occurs at 23% of locations.
Maps similar to Fig. 1, but for the 6 indices other than NINO3.4 (i.e. NAM, SAM, NAO, PNA, DMI and QBO), are shown in Supplementary Figures 1-6. These maps show relatively few locations where significant correlations occur at locations different to the case for NINO3.4. Although there are some regional exceptions to this, such as for the DMI in SON in the equatorial Indian Ocean (Supplementary Figure 5), the percentage of the study region for which this occurs is similar to what would be expected based on random chance alone (as noted previously from Table 1).
Modes of variability such as NAM and SAM are characterised by shifts in atmospheric mass between polar and mid-latitude regions, such that associated teleconnections may take some time to become apparent in locations away from their region of origin, noting that the majority of lightning activity occurs in the tropics 51   . Correlation coefficients are shown (right panels, 'x' symbols) for the relationships between NINO3.4 and three atmospheric parameters (based on reanalysis) associated with thunderstorm occurrence: specific humidity (SH), temperature lapse (TL) and convective available potential energy (CP). Correlation coefficients are also shown (right panels, '□ ' symbols) for the relationships between lightning flash density and the three atmospheric parameters. The NINO3.4 region is shown in blue (upper panel: from 5°N to 5°S in latitude and from 120°W to 170°W in longitude).
Scientific RepoRts | 6:20874 | DOI: 10.1038/srep20874 Lightning flash densities versus NINO3.4 anomalies are shown in Fig. 2 (left panels). Also presented in Fig. 2 (right panels) is an examination of NINO3.4 and lightning flash density in relation to three different atmospheric parameters associated with favourable environments for thunderstorm occurrence: specific humidity at 925 hPa (SH), temperature lapse at 850-500 hPa (TL) and surface-based Convective Available Potential Energy (CAPE: CP). The parameters are based on ERA-Interim reanalysis 49 and were selected to represent commonly used measures of moisture, heat and momentum in the atmosphere. These three parameters are not expected to entirely explain the exact physical cause of variations in lightning activity throughout the world as there are a large number of different factors that can potentially influence convective storm development 16,27,28,36,52,[55][56][57][58][59] .
The atmospheric parameters are strongly correlated with NINO3.4 in many cases, as well as with lightning flash density. This is the case for specific humidity and CAPE in all four locations shown in Fig. 2 as well as for temperature lapse in the Philippines and China examples. The sign of these correlations is consistent with what would be expected if ENSO-related variations in these parameters were influencing lightning activity, thereby providing a degree of physical plausibility for the relationships between ENSO and lightning activity. For example, the results for the Philippines in DJF (Fig. 2a) show that lightning activity and specific humidity are both typically lower than average during El Niño conditions (i.e., negatively correlated with NINO3.4), with a positive correlation between specific humidity and lightning activity. Temperature lapse has a somewhat stronger relationship to ENSO and lightning flash density for the Philippines and China cases than for the Indonesia and Brazil cases, with this being similar to the case for CAPE (which could be expected to some degree given that CAPE can be influenced by temperature lapse rate as well as by atmospheric moisture content).  Figure 14).
The study region examined by that previous study 28 as well as by another similar study 34 encompasses the location in eastern China that was examined here in Fig. 2b. The results presented in these two previous studies indicate increased lightning activity for El Niño events in this eastern China region. With broad similarity to these previous studies, Supplementary Figure 15 shows that more lightning activity tends to occur during El Niño than La Niña conditions in this region due to the relatively strong ENSO influence on lightning activity during the first half of the year (DJF and MAM), with relatively little ENSO influence on lightning activity during the second half of the year (JJA and SON). It also shows that the lightning activity in this region during the 1998 El Niño was the highest on record for DFJ throughout the 18-year study period and third highest on record for MAM.
One previous study 33 examined global lightning flash counts in relation to ENSO with a focus on two 12-month time periods to define El Niño conditions (during 1997/1998 and 2002/2003) and two 12-month time periods to define La Niña conditions (during 1995/1996 and 1998/1999), indicating about 10% more lightning globally for El Niño than La Niña conditions. That study also examined individual regions in relation to lightning variability associated with ENSO, such as in their Fig. 8 for South America which indicated more lightning for El Niño than for La Niña conditions at around 30°S, while at around 10°S more lightning occurs for La Niña than El Niño conditions. This change in sign of the ENSO/lightning relationship in South America is broadly consistent with the results presented here in Fig. 1, while noting various differences in method between their study and this study (such as their use of average lightning activity over large regions, as well as their use of 12-month time periods to examine specific El Niño and La Niña events). It is also noted that there can be considerable differences in the lightning response between individual ENSO events, as is apparent between two different years they classified as La Niña in their Figure 9 for the Ganges River valley.
In addition to these previous studies which use a case study type of approach to examine the ENSO/lightning relationship (i.e., examining a number of individual ENSO events and/or individual regions), a recent study 29 applied a more systematic approach to examine this relationship. That study examined the ENSO/lightning relationship for each individual location globally (based on the OTD/LIS sensor data coverage) and for all times throughout their study period (from 1997 to 2006). Their results show that there is considerable spatial variation in the magnitude and sign of the relationship, based on examining the ENSO/lightning relationship for all time periods combined throughout the entire year (rather than for individual seasons). Building on the results of previous studies such as these mentioned above, the results presented here examine for the first time the seasonal variation in the ENSO/lightning relationship (in addition to the spatial variability in this relationship), thereby allowing an investigation of the potential for seasonal forecasting of lightning and thunderstorm activity throughout tropical and temperate regions of the world.
Potential for seasonal forecasting. Hindcasts (i.e., reforecasts) of NINO3.4 values were obtained from NOAA for the period 1996 to 2010, with a three-month lead time prior to the start of each season 60 . These hindcasts are used here to examine predictions of lightning flash density as either higher or lower than their seasonal median value, based on whether the hindcast NINO3.4 values are higher or lower than their seasonal median value. The accuracy of the predictions is calculated as the number of correct predictions divided by the total number of predictions. Figure 3 shows the accuracy of the predictions, presented for all locations where a significant relationship exists between lightning activity and ENSO (based on Fig. 1). The accuracy is higher for DJF, MAM and SON than for JJA: the mean value of the accuracy is 52% for JJA as shown in Fig. 3 (i.e. similar to that of random chance) compared with a mean value of 64-70% for each of the other three seasons. There are some regions and seasons where the accuracy shown in Fig. 3 is higher than 70%, including for three of the four examples examined previously from Fig. 2 (Philippines for DJF, China for MAM and Brazil for SON), with relatively low accuracy of 53% for one of those four examples (JJA in Indonesia). The low accuracy during JJA is associated with a period of low hindcast skill for NINO3.4 in the months preceding the JJA season (i.e. March, April and May) 60 and is an example of the challenges that remain in understanding ENSO variability and predictability. Figure 4 presents seasonal anomalies of lightning flash densities for three different phases of ENSO (El Niño, Neutral and La Niña), shown for locations where a significant relationship occurs (from Fig. 1). The anomalies are calculated as the percentage difference to the seasonal mean at each individual grid-point. A benefit of examining these anomaly fields (Fig. 4) is that they show the relative magnitude of the variation in lightning activity associated with ENSO, in contrast to the correlations (Fig. 1) which indicate the strength of the relationship between lightning activity and ENSO. The phases of ENSO are defined using three-month average values of NINO3.4 with values greater (less) than or equal to 0.5 (− 0.5) defined as El Niño (La Niña) conditions if this occurs for at least two consecutive seasons, otherwise the conditions are defined as Neutral (similar to the convention used by NOAA 61 ). The period of data available for this study is such that  Table 2).

Anomaly maps for ENSO phases.
The lightning anomalies tend to be larger in magnitude for the two extreme phases of ENSO (i.e. El Niño and La Niña) than for the Neutral phase. Large regions with positive anomalies are particularly apparent in the cases shown for El Niño conditions, while the anomalies for neutral conditions are predominantly negative. The lightning anomalies are of the order of 25% to 75% of the seasonal average in many cases for the El Niño and La Niña phases. Figure 4 could be used to provide guidance in relation to seasonal forecasting applications. For example, if El Niño conditions were forecast to occur for an upcoming DJF season, then it might be expected that the southeast China region could experience increased lightning activity, while the region near the Philippines could experience decreased lightning activity. Similarly, if El Niño conditions were forecast to occur for an upcoming SON season, then it might be expected that the equatorial Central Pacific region could experience increased lightning activity, while the western South Pacific region could experience decreased lightning activity.

Discussion
Of the seven different modes of variability examined here, ENSO was found to have the strongest relationship with lightning activity during each individual season, with relatively little relationship for the other modes of variability. ENSO is the strongest source of atmospheric and oceanic variability throughout the world at interannual time scales and provides a large source of seasonal forecasting skill 16,23,24,62 . Although ENSO originates in the tropical Pacific Ocean region, it can affect land and ocean regions in many regions throughout the world, including through the modulation of phenomena such as the Walker Circulation, extratropical storm tracks, jetstreams and the production of Rossby wave trains that propagate to mid-and high-latitudes 17,63-67 .
There is considerable evidence that variations in lightning and thunderstorm activity can be related to ENSO in some cases. For example, a number of previous studies examined lightning activity in relation to particular ENSO events, including many reports of anomalous lightning activity for the 1997-98 El Niño and subsequent La Niña conditions 20,[25][26][27][31][32][33][34] . In particular, the Maritime Continent region of Southeast Asia has been identified previously, as well as here, as being a region where strong relationships exist between lightning activity and ENSO 26,28,29,[32][33][34]68 . Other examples of strong relationships, as noted here and in previous studies, include northeast Brazil 29,37 , the western North Pacific region 25,[27][28][29] and the Gulf of Mexico 25,33 .
Variations in atmospheric parameters associated with thunderstorm activity were found to provide a plausible physical explanation for the ENSO-related lightning variability. A method for predicting anomalous lightning and thunderstorm activity was examined, while noting that there is a wide range of methods that could potentially be used for seasonal forecasting applications, including both statistical and dynamical modelling techniques. The accuracy of the seasonal predictions was found to be better for DJF, MAM and SON than for JJA. The relative weakness of the skill for JJA could be expected to some degree given that this is the typical transition period for ENSO signals.
An ability to predict variations in lightning and thunderstorm activity could be expected to lead to improved preparedness for associated impacts on built and natural environments. Given the wide range of impacts associated with lightning and thunderstorm activity throughout the world, this could have benefits for groups such as industry, government, insurance and emergency services. The results presented here show that there is potential for accounting for, and also predicting in advance, the variability of seasonal lightning and thunderstorm activity in various regions of the world.

Methods
Lightning data. The lightning flash data used here are based on observations from two NASA satellite sensors: the Lightning Imaging Sensor (LIS) and the Optical Transient Detector (OTD) 41,42 . The Optical Transient Detector (OTD) was launched into low Earth orbit in April 1995 on the OV-1 satellite (formerly MicroLab-1), with this mission ending in March 2000, such that the total period of available data covers about 5 years. The Lightning Imaging Sensor (LIS) on the Tropical Rainfall Measuring Mission (TRMM) satellite 69 was launched in December 1997 and ceased operation in April 2015. Due to its orbit, the LIS sensor provided coverage only within the region from about 38°S to 38°N. Details of the seasonal climatology of the LIS and OTD data were recently presented 70 . The 18-year period of data available for this study spans multiple occurrences of different phases of the various large-scale modes of variability examined here.
Lightning flash densities throughout the time period from 1996 to 2013 were obtained from the monthly gridded time series product (name "LISOTD_LRMTS", version "V2.3.2014") produced from a combination of the LIS and OTD data (available from http://thunder.nsstc.nasa.gov) 71 . The spatial resolution of the data is 2.5 degrees in both latitude and longitude, although prior to being released the data are smoothed with a 7.5° × 7.5° and approximately 3 month (111 days for OTD and 99 days for LIS) boxcar moving average (i.e. running mean) due to the approximate 3 month revisit time of the satellite coverage through the diurnal cycle at individual locations. Due to the effective temporal resolution of the satellite data (based on the approximate 3 month moving average), seasonal lightning activity is examined here based on the data for January (to represent the December, January and February season: DJF), April (to represent the March, April and May season: MAM), July (to represent the June, July and August season: JJA) and October (to represent the September, October and November season: SON). In contrast to this method, if three consecutive months of the data were averaged to try to represent seasonal lightning activity, it would actually represent lightning activity over a time period much longer than one season (i.e., a time period of over 5 months). The geographic study region used here covers all longitudes globally, and ranges in latitude from 35°S to 35°N within the field of view of both the OTD and LIS sensors, based on a 2.5 degree grid in both latitude and longitude.  72 . These seven large-scale modes of variability have a period of variability less than decadal and greater than monthly, as required given that the period of available lightning data is not long enough to allow decadal variations to be examined and that the effective temporal resolution of the lightning data is about 3 months. The seven indices used here represent commonly used measures of these modes of variability. Three month averages of these indices are used here for DJF, MAM, JJA and SON.

Modes
Reanalysis. ERA-Interim reanalysis 49 is used to derive three different atmospheric parameters, including specific humidity at 925 hPa, temperature lapse at 850-500 hPa and surface-based Convective Available Potential Energy (CAPE). Seasonal averages are produced for DJF, MAM, JJA and SON, with a grid spacing of 7.5 degrees in both latitude and longitude, consistent with that of the lightning data. Coastlines are shown in figures presented here based on the land-sea mask of the ERA-Interim reanalysis. Statistical analyses. The sample Pearson's correlation coefficient, r, is used to examine the dependence between datasets. The 95% confidence level is used to examine the significance of results, determined using a nonparametric bootstrap method based on 10 000 random permutations of the data, with two-sided confidence intervals based on percentiles.