Machine Learning-Based Identification of Key Predictors for Lightning Events in the Third Pole Region

Abstract

The Third Pole region, particularly the Hindu–Kush–Himalaya (HKH), is highly prone to lightning, causing thousands of fatalities annually. Skillful prediction and timely communication are essential for mitigating lightning-related losses in such observationally data-sparse regions. Therefore, this study evaluates kilometer-scale ICON-CLM-simulated atmospheric variables using six machine learning (ML) models to detect lightning activity over the Third Pole. Results from the ensemble boosting ML models show that ICON-CLM simulated variables such as relative humidity (RH), vorticity (vor), 2m temperature (t_2m), and surface pressure (sfc_pres) among a total of 25 variables allow better spatial and temporal prediction of lightning activities, achieving a Probability of Detection (POD) of ∼0.65. The Lightning Potential Index (LPI) and the product of convective available potential energy (CAPE) and precipitation (prec_con), referred to as CP (i.e., CP = CAPE × precipitation), serve as key physics aware predictors, maintaining a high Probability of Detection (POD) of ∼0.62 with a 1–2 h lead time. Sensitivity analyses additionally using climatological lightning data showed that while ML models maintain comparable accuracy and POD, climatology primarily supports broad spatial patterns rather than fine-scale prediction improvements. As LPI and CP reflect cloud microphysics and atmospheric stability, their inclusion, along with spatiotemporal averaging and climatology, offers slightly lower, yet comparable, predictive skill to that achieved by aggregating 25 atmospheric predictors. Model evaluation using the Technique for Order of Preference by Similarity to Ideal Solution (TOPSIS) highlights XGBoost as the best-performing diagnostic classification (yes/no lightning) model across all six ML tested configurations.

1. Introduction

The Indian subcontinent experiences a lightning maximum during the pre-monsoon season (March to June), with a secondary peak in the late monsoon season (August to October) [1,2]. Daggar, Pakistan (34.45° N, 72.35° E), located in the western Himalayas, is Asia’s top-ranked lightning hotspot [2]. The Tropical Rainfall Measuring Mission (TRMM)—Lightning Imaging Sensor (LIS) observations reveal that the Himalayan region is one of the world’s most lightning-prone zones; the intensity and frequency of lightning events vary from the West to the East Himalayas [3,4]. The rough terrain, distinct topography, and varied atmospheric circulation (westerlies, monsoon) cause significant spatiotemporal variability in lightning flash density over the Himalayan region [3]. Lightning events in the western Himalayas are mainly influenced by westerlies and orographic lifting, whereas those in the eastern Himalayas are primarily driven by afternoon convection over the Bay of Bengal [4,5]. The Tibetan Plateau experiences a much lower frequency of lightning events compared to the Himalayan regions.

Lightning is an electrostatic discharge that produces an illuminated flash of light in the sky, often accompanied by thunder. Broadly accepted, the non-inductive charging mechanism explains that lightning occurs within thunderstorm clouds due to charge separation between ice particles, as collisions of ice crystals and graupel in the presence of supercooled liquid water facilitate the transfer of electrical charge [6,7]. Charged particles separate due to gravity and convection, forming oppositely charged regions within the cloud, leading to lightning when the electrical potential reaches a sufficiently high threshold [7]. Thunderstorm clouds, typically 5–10 km wide and at least 4 km deep, can generate multiple lightning flashes per second for 30–60 min [4]. Lightning poses a significant hazard in the Nepal Himalayas due to strong monsoonal activity and the region’s complex orography, with 2501 reported lightning fatality events between 1971 and 2019 resulting in 1927 fatalities and over 20,000 people affected [8]. The intense heating from positive lightning discharges, reaching temperatures of up to 30,000 K, underscores their destructive potential [9].

Given the substantial human and socio-economic impacts of lightning in regions like the Himalayas, improving lightning forecasting has become a critical priority. This urgency has driven a transition from traditional observational approaches to more sophisticated numerical weather prediction (NWP) models, with recent advances increasingly leveraging ML techniques to enhance forecast accuracy [10,11]. Price and Rind (1992) developed a lightning parameterization, known as PR92, for the WRF model based on cloud top height and vertically integrated ice water path and CAPE, resulting in better capture of spatial lightning patterns [12]. A previous study predicted lightning by post-processing NWP ECMWF ensemble output for the European Eastern Alps region by proposing a statistical lightning prediction model [13]. Furthermore, researchers widely acknowledge the critical role of data assimilation in NWP models for extending forecast horizons and improving accuracy. Mueller and Barleben (2024) reviewed the experimental ICON-RUC (Rapid Update Cycle) model and emphasized that combining it with the nowcasting method JuliaTSnow significantly improves short-term (0–6 h) forecasting skill. This enhancement, evidenced by increased Critical Success Index (CSI) scores (>0.5), stems from the model’s rapid update cycles and higher spatial resolution [14]. The author later showed that for RUC, the CSI score remains above the critical threshold (0.5) for forecasts within 3 h of data assimilation but drops below this value for longer lead times [14]. The decline underscores limitations in the model’s physics, which struggle to accurately predict thunderstorms due to the inherent unpredictability described by chaos theory. Studies also show that deep learning methods attain enhanced thunderstorm prediction (POD = 0.60 ± 0.02) through purely data-driven approaches, surpassing current NWP capabilities for short-term forecasts [14,15]. A previous study reported improved lightning prediction during the monsoon season using a random forest (RF) regression model (R-squared score = 0.81), with predictors derived from LIS satellite observations, ERA5 reanalysis, and MODIS data [16]. Further, Geng et al. (2021) developed LightNet+, a data-driven lightning forecast model using ConvLSTM, which effectively integrates WRF simulation data, past lightning data, and Automatic Weather Station (AWS) data [11]. The results indicate that LightNet+ outperforms the three established schemes PR92, F1, and F2, achieving higher POD by 22%, 36%, and 26%, respectively. This improvement reflects enhanced forecasting skill with the integration of additional data sources.

Significant advances in lightning prediction and modeling have emerged in the last few years, from simple ML models to enhancing observational data with ML and post-processing NWP models using statistical and ML techniques. Recent studies have trained ML models for lightning prediction using observation data [16,17]. However, in data-sparse regions such as the Himalayan–Tibetan complex terrain, high-elevation terrain and northwestern areas face critical gaps in ground observations and satellite-derived climate records exhibit persistent uncertainties [18]. We trained our ML models using the ICON-CLM NWP model only, an approach that remains largely unexplored. In this framework, ML-based lightning prediction with lead time relies entirely on the forecasting skill of the NWP model, rather than on extensive observational data. Therefore, NWP models coupled with an ML approach can prove beneficial in advancing climate studies. Building on our previous research by Singh and Ahrens (2023) [4], our aim was to enhance the accuracy and realism of the ICON-CLM dynamical model by post-processing the dynamic parameters governing lightning occurrence using simple ML models. We also examined the dependency of various variables in predicting lightning occurrence and identified the most robust and best-performing ML model for this highly dynamic and uncertain phenomenon. This paper is structured as follows: Section 2 details the data and methodology, including the datasets and models used. Section 3 discusses our experimental results, key findings, and potential limitations. Finally, Section 4 offers a concise conclusion summarizing our study.

2. Data and Methodology

2.1. Lightning Observation

In this study, we used International Space Station (ISS)-based Lightning Imaging Sensor (LIS) observations from October 2019 to September 2020. The LIS is a satellite-based instrument that detects the global distribution and variability of total lightning, including cloud-to-cloud (CC), intra-cloud (IC), and cloud-to-ground (CG). LIS detects storm scale resolution (∼4 km) at millisecond timing with a narrow band filter (777 nanometers) in conjunction with a high-speed charge-coupled device (CCD) detection array [19]. This enables ISS-LIS to detect lightning during both the day and the night [19]. A flash comprises groups of lightning events that occur within 330 ms and 5.5 km, where each group represents simultaneous events in adjacent pixels. The flash child count quantifies the number of constituent groups per flash [20] and were used as lightning flash counts in the present study to train and test ML models.

2.2. Numerical Model Setup

The Icosahedral Nonhydrostatic Weather and Climate Model in Climate Limited Area Mode (ICON-CLM), version 2.6.4 [21,22], simulated the Third Pole region from October 2019 to September 2020 (with September 2019 as the spin-up period). The simulations were conducted at a high resolution with a horizontal grid spacing of 3.3 km and 60 vertical levels extending up to 10 hPa, covering the domain 22.5–42.5° N and 67.5–117.5° E [4,5]. However, our study focused on the subdomain 25–40° N and 70–115° E for the ML model’s training, testing, and further analysis. The ICON-CLM simulation used ECMWF’s fifth-generation atmospheric reanalysis (ERA5) as initial and boundary conditions. The performance of the ICON model was comprehensively evaluated in previous studies, including Singh and Ahrens (2023) [4] and Collier et al. (2024) [23], which provide detailed assessments of its predictive capabilities and limitations. The model simulated 25 variables on the hourly temporal scale used in the ML models, representing key physical processes associated with lightning, including atmospheric instability (4 variables), cloud properties (2), moisture and precipitation (3), and column-integrated water (2), adapted from a previous study [17] (see

Table 1. Summary of experiments with different spatial–temporal coverages and variable sets. (See Abbreviations before Appendix A).

Experiment Set	Variables Included
	CAPE, prec_con, LPI, RH_300, RH_500, RH_850, vor_300, vor_500,
S1	vor_850, T_300, T_500, T_850, sfc_pres, t_2m, clcm, clch, cin_ml,
	shfl_s, qhfl_s, lhfl_s, thb_s, sob_s, tqc, tqi, z
S2	CP, LPI
S3	CP, LPI, TRMM Climatology
S3t	$\overline{CP}$, $\overline{LPI}$, TRMM Climatology (i.e., time-averaged LPI and CP)
Spatial Coverage	Names of Experiments
	S1	S2	S3	S3t
Grid Point (GP)	S1-GP	S2-GP	S3-GP	-
60 km	S1-60	S2-60	S3-60	S3t-60-1h
90 km	S1-90	S2-90	S3-90	S3t-90-2h

). Our study further augmented the set with variables related to surface and radiation fluxes (5), vorticity (3), temperature (3), lightning potential index (LPI), surface elevation (z), and precipitation. Together, they formed a physically meaningful and comprehensive predictor set for lightning activity.

The CAPE variable quantified the thermodynamic energy available for convective updrafts, serving as a primary indicator of atmospheric instability essential for thunderstorm and lightning development. As discussed in Section 1, the lightning parameterization PR92, which remains widely used, including in the WRF model, relied on CAPE to estimate lightning distribution [12]. In addition, Romps et al. (2018) demonstrated that CP serves as an effective proxy for lightning, accurately capturing flash rate density over land [24]. Tippett et al. (2019) further applied the CP proxy to predict daily cloud-to-ground flash counts across CONUS (2003–2016), reporting strong regional correlations between CAPE and lightning activity [25]. While CP captures essential aspects of lightning generation, other factors, such as atmospheric moisture and wind shear, also play a critical role. To account for wind shear and vertical dynamics, we incorporated the LPI, which integrated vertical velocity information within convective clouds. LPI reflects the strength of updrafts and the microphysical processes that drive charge separation, both crucial for lightning generation [26,27,28]. Multiple studies have validated the predictive utility of LPI in various regions [4,25,27,29]. Building on these insights, we adopted a physically informed feature selection strategy that prioritized interpretability and computational efficiency. In this context, we selected CP and LPI as potentially important predictors, aligning with the findings of Singh and Ahrens (2023) [4], and formed the basis of our physics-aware modeling approach.

2.3. Machine Learning Models

In the present study, the following ML models were used: extra tree (ET), gradient boost (GB), K-nearest neighbor (KNN), RF, support vector machine (SVM), and extreme gradient boost (XGBoost). RF is an ensemble method combining multiple decision trees built using bootstrap sampling and random feature selection to reduce overfitting [30]. ET is a variant of RF that builds trees with fully random splits, increasing diversity and robustness [31]. GB builds trees sequentially, with each one correcting the errors of the previous, optimizing a chosen loss function [32]. Its improved version, XGBoost, adds regularization, handles missing data, and supports parallel computation [33]. KNN is a non-parametric approach that classifies data based on the majority label among the number of nearest neighbors [34]. SVM is a classifier that finds the optimal margin between classes, often mapping data into higher dimensions for better separation [35]. Our supervised learning approach employed these six models, specifically selected for their distinct strengths in handling tabular lightning data. The selected ML models, ranging from interpretable algorithms like KNN and SVM to more complex ensemble methods such as RF, ET, GB, and XGBoost, are well-suited for structured tabular data. Ensemble models effectively capture nonlinear relationships and handle noise, while SVM delineates decision boundaries and KNN adapts to localized patterns, supporting region-specific variation. This diversity enabled a balanced trade-off between interpretability and predictive performance. A detailed comparison of model characteristics is provided in Appendix A 

Table A1. Detailed overview of ML models.

Model	Algorithm Type	Key Features	Strengths	Limitations
Random Forest	Ensemble (Bagging)	Builds multiple decision trees and combines their outputs (majority vote).	- Handles high-dimensional data well. - Robust to over-fitting. - Provides feature importance.	- Slower with large datasets. - Less interpretable compared to simple models. - Memory-intensive.
Extra Trees	Ensemble (Bagging)	Similar to Random Forest but uses randomized splits for trees.	- Faster than Random Forest. - Lower variance. - Provides feature importance.	- May underperform in scenarios requiring precision tuning. - Sensitive to noise in data.
Gradient Boosting	Ensemble (Boosting)	Combines weak learners sequentially to reduce errors iteratively.	- High accuracy. - Effective for imbalanced datasets. - Can optimize custom loss functions.	- Slower to train. - Sensitive to overfitting without tuning. - Requires parameter tuning.
XGBoost	Ensemble (Boosting)	Highly efficient implementation of gradient boosting with regularization.	- Fast and memory-efficient. - Handles missing data. - Regularization helps prevent overfitting.	- Complex to tune. - Less interpretable. - Sensitive to noise in the data.
K-Nearest Neighbors	Instance-Based	Assigns class labels based on the majority vote of neighbors (k nearest points).	- Simple and intuitive. - Effective with small datasets. - No training phase.	- Computationally expensive for large datasets. - Requires proper scaling of features. - Sensitive to irrelevant features.
SVM (Support Vector Machine)	Discriminative	Finds a hyperplane to separate classes with maximum margin (can use kernels for non-linear problems).	- Effective in high-dimensional spaces. - Robust to overfitting (with kernel trick).	- Memory-intensive. - Slower for large datasets. - Requires careful parameter selection and kernel choice.

.

To ensure a fair comparison, all six structurally distinct models (RF, ET, GB, XGBoost, KNN, SVM) were initially evaluated using default hyperparameters. This approach isolated the influence of model architecture on performance, supported reproducibility, and promoted a transparent, physics-aware framework. The best-performing model was subsequently fine-tuned, with the results discussed in Section 3.

2.4. Experimental Setup

The flow chart in Figure 1 outlines the selection process for spatiotemporal lightning events and extraction of input data for ML models in this study. Firstly we filtered the ISS-LIS swaths to focus exclusively on the Third Pole region (25° N–40° N, 70° E–115° E). We subsequently segregated 36,975 verified lightning flash counts recorded over the domain region between October 2019 and September 2020. To enhance robustness and mitigate bias, we incorporated an equal number of false instances (no lightning detected) from ISS-LIS measurements. These were randomly selected while preserving proportionality to the monthly distribution of true lightning events. This approach ensured a balanced dataset, with an equal number of YES (lightning detected) and NO (no lightning) events for each month, aligned with the observed monthly frequency of lightning occurrences. Using this balanced ISS-LIS dataset, we identified the time and location of lightning activity and, at these points, extracted the corresponding 25 ICON-CLM model variables for training and testing the ML models (detailed discussion in Section 3). Furthermore, the dataset was randomly split into 80% for training and 20% for testing; therefore, all further analysis and metric calculations in Section 3 were performed with the test set. Furthermore, TRMM lightning climatology data was used to supplement the observed lightning events.

We evaluated the predictive capacity of 25 atmospheric variables for lightning using three experimental setups: direct grid point data (S1-GP) and spatial averages computed over 60 km (S1-60) and 90 km (S1-90) radii centered on each grid point, collectively termed setup S1 (Table 1). Setups S2 and S3 followed the same structure as S1, differing only in the selected input features. Additionally, an extended experiment, S3t, evaluated the effect of temporal averaging, using two approaches: one with 1 h averages within a 60 km spatial radius (S3t-60-1h) and another with 2 h averages within a 90 km radius (S3t-90-2h). Notably, both S3 and S3t incorporated climatology data to enrich the analysis. These setups were designed to account for spatial and temporal uncertainties, acknowledging that lightning may occur with slight spatial or temporal offsets from predicted locations.

3. Results and Discussion

The Third Pole region, encompassing the Himalayan mountains and the Tibetan Plateau (70°–110° E, 25°–40° N; Figure 2), emerges as a prominent lightning hotspot in Asia, as indicated by both our analysis and long-term observations from the TRMM satellite [36]. Figure 2a presents the TRMM lightning climatology over the Third Pole region, derived from LIS and optical transient detector (OTD) observations spanning 1998 to 2013. TRMM suggests that the western Himalayan region has the highest number of lightning incidents, followed by the Brahmaputra Valley (26° N, 90° E) and the central/eastern Himalayan ranges, as reported in previous studies [4,36]. Figure 2b shows the WWLLN-based daily frequency of lightning strokes per unit area (km²) climatology from 2010 to 2021, which displays spatial patterns broadly consistent with the TRMM climatology. WWLLN observations show fewer lightning events over the western Himalayan region compared to the Brahmaputra valley (Figure 2b). In addition, the WWLLN record shows more lightning incidents over the east of the Tibetan Plateau than in the TRMM climatology. WWLLN offers the advantage of continuous global coverage, in contrast to satellite-based sensors LIS and OTD, which only sample lightning activity during satellite overpasses (see Virts et al., 2013, and references therein) [37]. Given the extended temporal coverage of WWLLN data, spanning up to 2021, compared to TRMM LIS climatology, an increase in lightning activity in recent years is particularly evident in regions such as the Tibetan Plateau. This trend reflects WWLLN’s inclusion of more recent data. Differences in the observed magnitude between the TRMM and WWLLN datasets arise partly from the attenuation of very-low-frequency radio waves used by WWLLN, which is strongest during the daytime (see Virts et al., 2013, and references therein) [37]. Additionally, WWLLN preferentially detects strong cloud-to-ground strokes [19]. Although strokes and flashes are distinct, since a flash may consist of multiple strokes [38], for annual climatology, the number of WWLLN-detected strokes is generally close to the number of flashes observed by LIS/OTD. As such, climatological comparisons between WWLLN strokes and LIS/OTD flashes remain informative [37]. We used lightning flash density data from ISS-LIS to assess lightning activity across the study region. A limitation of ISS-LIS was its 90 min revisit time over a given region, which restricted the temporal frequency of observations, while the detection efficiency was ∼60%, 10% lower than that of TRMM-LIS, its prior platform [19]. The spatial distribution of ISS-LIS lightning activity during the study period closely resembled the TRMM climatology, indicating consistency in the observed patterns (Figure 2c), thereby reinforcing the reliability of ISS-LIS as a valid ground truth dataset for training ML models. TRMM climatology and ISS-LIS lightning observations were gridded to a 0.5° × 0.5° spatial resolution to match the native grid of the WWLLN dataset.

We evaluated the performance of the ML models using several standard classification metrics (see

Table 2. Summary of evaluation metrics. Note: $FPR={\displaystyle \frac{FP}{FP+TN}}$.

Metric	Formula	Interpretation
Accuracy	${\displaystyle \frac{TP+TN}{TP+TN+FP+FN}}$	Overall correctness of the model
Precision	${\displaystyle \frac{TP}{TP+FP}}$	Measured how many predicted events actually happened
POD (Recall)	${\displaystyle \frac{TP}{TP+FN}}$	Measured how well the model detected actual events
FAR	${\displaystyle \frac{FP}{TP+FP}}$	Measured how many predicted events were false alarms
CSI	${\displaystyle \frac{TP}{TP+FP+FN}}$	Balance between false alarms and missed events
ROC-AUC	${\int }_0^{1}POD\left(FPR\right)dFPR$	Measured model’s ability to distinguish classes (lightning and no-lightning)

). In this context, a True Positive (TP) referred to correctly predicting lightning when it actually occurred. A False Positive (FP), or false alarm, occurred when the model predicted lightning that was not observed in the ISS-LIS data. A False Negative (FN), or missed event, occurred when the model failed to predict lightning that was observed. A True Negative (TN) referred to correctly predicting no lightning when no lightning was observed. All metric values ranged from 0 (worst) to 1 (best), except for the False Alarm Rate (FAR), where lower values indicated better performance; thus, we report 1-FAR for consistency. ROC-AUC stands for Receiver Operating Characteristic–Area Under the Curve.

Based on the lightning patterns observed in Figure 2c, we used six ML models to reconstruct and evaluate the spatial lightning activity, aiming to assess their predictive skill across the study domain. Figure 3 presents a heat map depicting the performance of the S1 experiments across multiple metrics. The rows represent different models and experiments, such as ET, GB, KNN, RF, SVM, and XGBoost. The columns correspond to different performance metrics, including accuracy, CSI, POD, precision, ROC-AUC, and 1-FAR. Each model was evaluated using point data and aggregated data at spatial scales of 60 km and 90 km. These spatial scales were selected based on prior studies: Lynn et al. (2010) found that LPI performs best when averaged over 20 × 20 km² [27], while Singh and Ahrens (2023) particularly demonstrated that over the Third Pole region, ICON-simulated LPI captures observed lightning activity with only 18% success at the grid point scale. However, when averaged over 60 × 60 km² and 90 × 90 km², the success rate significantly improves to 52.47% and 61.26%, respectively, with similar enhancements seen in the CP [4].

The XGBoost model consistently achieved top performance across most evaluation metrics, particularly for spatially averaged datasets (60 km and 90 km). For instance, XGBoost-S1-90 attained the highest accuracy (0.63), POD (0.66), and ROC AUC (0.68), highlighting its ability to effectively utilize aggregated data. Spatial averaging enhanced overall model performance, with the KNN and RF models showing notable improvements in accuracy and CSI at S1-60 and S1-90 compared to the grid-point configuration (S1-GP), suggesting that aggregation aided in capturing broader spatial patterns and mitigating noise. While GB and XGBoost performed robustly across all metrics, ET demonstrated stable performance across scales but showed less sensitivity to spatial averaging. In contrast, SVM consistently underperformed, particularly in accuracy and CSI, indicating limited capability in detecting TP events.

With a comparative analysis of multiple metrics alone, it was difficult to definitively identify the optimal model and discern the underlying influence of specific predictors. Thus, it was equally important to examine which of the 25 predictors most strongly influenced the ML predictions. In order to evaluate feature importance across models, we employed methods tailored to each algorithm’s characteristics. For tree-based models RF, ET, GB, and XGBoost, we used the built-in ‘feature_importances_’ attribute, which quantified the contribution of each feature (input variable) in model prediction based on the total reduction in impurity, typically Gini impurity for RF, ET, and GB and gain for XGBoost. For non-tree-based models such as SVM and KNN, which lack intrinsic feature importance measures, we used permutation importance. This method estimated the impact of each feature on model performance by measuring the change in accuracy when the feature’s values were randomly permuted, with accuracy used as the scoring metric in this study. Figure 4 illustrates the relative contribution of each of the 25 input variables (predictors) to the predictive outputs of all six models in the experimental setup S1-90. The feature importance of other experiments (S1-GP, S1-60) are presented in the Supplementary Results in Figure S1. In the given Figure 4, the horizontal axis represents the variables (features) used to train the model; see the list of variables in Table 1. RF and ET, belonging to the ensemble learning family of models, exhibited nearly parallel trends that closely aligned with fluctuations, including spikes and lows with respective features. The ET and RF models exhibited relatively uniform feature importance distributions, which could make them less effective at identifying the most influential features. Due to their ensemble nature averaging across many weak learners, they tended to assign importance in a moderate and somewhat diffuse manner, limiting their ability to distinctly highlight the most relevant features. The random assignment of weights to 25 variables showed that all features received nearly equal importance (see Figure 4). Such a uniform distribution of feature importance suggests that the model lacked reliability in distinguishing key predictors. Unlike ET/RF, GB produced sharper importance peaks for key features like T_850 and precipitation, likely reflecting its iterative residual learning process that progressively emphasized the most influential predictors [39]. Although in the real physical world, almost 50% of the variable features exhibited strong inter-correlation and interdependence that could negatively influence the learning process of the ML models. The inconsistent emphasis on features such as z, thb_s, and RH highlights the ambiguity in model interpretation, underscoring the need for physics-informed feature selection to ensure meaningful and domain-relevant insights. Therefore, as discussed in Section 2.3, we tested the predictability of the ML models using the key predictors CAPE and LPI in experiment S2.

The S2 experiment simulation yielded results comparable to the 25-feature experiment S1. While accuracy and other metrics showed slight variations, models like XGBoost and GB maintained a POD of 0.6. Figure 5 presents the S2 experiment heat map, evaluating the isolated importance of CP and LPI. Despite relying solely on these two predictors, the ML models achieved performance comparable to that of the S1 experiment, reaffirming their dominant dynamic influence on the prediction task. A sensitivity experiment incorporating TRMM climatology data (S3) revealed no measurable improvement in model prediction skill (see Supplementary Results Figure S2). Although the integration of multi-source data was expected to enhance model learning [11], the findings indicated minimal additional benefit from climatological inputs in the context of this specific prediction objective.

Lightning is a localized phenomenon wherein thunder clouds can traverse significant distances before striking the ground. Consequently, models may correctly detect lightning events yet misidentify their precise location or timing. Therefore, averaging predictions over 1 and 2 h intervals can enhance detection consistency and reduce localization errors. The S3t experiment (Figure 6) incorporated time-averaged CP and LPI variables, building upon the previous experimental framework. This configuration demonstrated superior performance across all models compared to S2 and S3 while maintaining high predictive accuracy with only a marginal reduction 5–6% relative to S1. The results suggest that temporal averaging of these key dynamical variables effectively preserves predictive skill while enhancing model stability. Meanwhile, the margin-based optimization model (SVM) showed a rise in the POD, which was crucial. The ensemble boosting models (XGboost, GB) still maintained similar accuracy to sets 1, 2, and 3. It is important to note the remarkable similarity (especially for XGBoost, POD ∼0.6) in S1 and S3t. This indicates that the six ML models were optimized to generate consistent predictions for the 25 variables, as well as the two time-averaged diagnosed variables (CP and LPI), when combined with the climatological data. Spatial averaging (grid → 60 km → 90 km) systematically improved skill scores (accuracy, precision, POD, ROC-AUC) individually in S1, S2, S3, and S3t, confirming mesoscale integration as critical for lightning predictability. In S1, the XGBoost model showed a 3–4% improvement in accuracy and ROC-AUC from S1-GP to S1-90 (Figure 3), a trend similarly observed in S2 and S3. Additionally, in S2 and S3, RF and ET also exhibited a 3% accuracy increase over the same spatial range in comparison to S1. Thus, time and space averaging consistently improved model performance over grid-point values.

Logistic regression (LR) is a linear model commonly used for binary classification. It estimates class probabilities using the logistic (sigmoid) function and fits model parameters through maximum likelihood estimation [40]. In this study, LR served as a baseline for comparison with more advanced models. Given the constraints on computational resources, we performed hyperparameter tuning for the XGBoost model. The optimally tuned configuration is referred to as XGBoost_best. Figure 7 illustrates the performance improvement during model development, starting from the LR baseline. The transition to the default XGBoost model resulted in an 8% increase in accuracy, with a further 1% improvement achieved using XGBoost_best. This additional 1% gain was likely attributable to a sixfold increase in the number of trees and a 16% reduction in the learning rate compared to the default XGBoost settings.

To identify the best-performing ML model within each experiment set and across all experiments, we employed a multi-criteria decision analysis method, TOPSIS (see Supplementary Results for Equation (S1)) [41]. A model with a high TOPSIS score was closer to the ideal solution (value ≈ 1), indicating strong performance across multiple metrics. It also had a smaller Euclidean distance from the ideal solution and a larger distance from the anti-ideal solution (value ∼ 0). Here, the XGBooost_best metrics were used to define ideal/anti-ideal solutions. Figure 8 illustrates the top 5 ranking models across individual sets of experiments (S1 to S3t). For S1, XGBoost achieved the highest TOPSIS score for S1-90. In contrast, KNN emerged as a model that performed better for S1-60, with a TOPSIS score of 0.76. For the S2 experiment set, ET performed poorly, while SVM topped the rank with a score of 0.41. RF scored the highest in the most relevant experiment in this study, that is, S3t. Therefore, RF could be also considered the second-best model in this study.

The spatial prediction performance of our top-performing XGBoost model (Figure 8) is presented across the Third Pole region (Figure 9). Figure 9a shows the density of the observed lightning event from the test set (20% unseen data), while Figure 9b shows regions without lightning during ISS-LIS overpasses. Therefore, Figure 9c,d illustrates the spatial distribution of the number of predicted lightning events per km² by the XGBoost model on the test dataset. This can also be interpreted as a spatial representation of the confusion metrics, highlighting regional variations in model performance. XGBoost reproduced lightning hotspots in the western Himalayan region (Figure 9c) but underrepresented the Brahmaputra Valley region. At the same time, Figure 9d shows the reverse, where no lightning (clear sky) events were predicted well over the Brahmaputra Valley region. This regional performance contrast reflects fundamental differences in convective regimes: the western Himalayas are characterized by deep convection generating intense but infrequent lightning, while the Brahmaputra Valley features shallow convection producing frequent but weaker events [4,36]. Potential model biases are also discussed in Section 4. These distinct regimes may require tuning of the NWP and ML approaches, enhancing parameterizations for mountainous terrain. Figure 9e shows locations where the model falsely predicted lightning that did not occur in observations (FP) and Figure 9f shows the missed events (FN) that the model failed to predict when lightning was observed in reality. Although FP and FN occurred with lower intensity, this suggests that the model struggled with predictor sensitivity or misclassified convective signals. The FP (false alarms, Figure 9e) consequently identified by XGBoost may in fact have corresponded to actual lightning events that were not captured by the ISS-LIS instrument due to its temporal limitations. This could be attributed to the fact that it introduced a degree of bias in the training of the ML model, which increased false positive rates.

Overall, the results showed slight performance reductions when using the two variables for training the ML models; this could be attributed to the fact that ISS-LIS observations exhibited a bimodal lightning distribution, with peaks in the afternoon and early morning over the study region. In contrast, ICON-simulated CP showed a single afternoon peak, likely due to its strong dependence on CAPE, which was sensitive to convection parameterization in the NWP models. On the other hand, LPI produced two peaks, but its performance was sensitive to the microphysical scheme in the ICON model [4]. The present study employed ISS-LIS satellite observations exclusively as ground truth for lightning identification, with model training restricted to one year of ICON-CLM simulation data. Unlike previous studies that employed neural networks and reanalysis data, our approach relied on simpler ML models. Despite these constraints, our model’s performance metrics—POD, FAR, CSI, and ROC-AUC—remained comparable to or slightly lower than those reported in past research [42,43,44]. Vahid Yousefnia et al. (2024) applied ICON data with a feedforward neural network for pixel-wise thunderstorm prediction up to 11 h ahead, achieving skill metrics (CSI, POD) comparable to our results [17]. The present study similarly leveraged ICON-CLM data but adopted a computationally efficient XGBoost model tailored for tabular inputs, identifying key predictors for lightning forecasts while maintaining robust performance.

The ISS-LIS satellite observations used in this study were inherently limited by their temporal resolution and orbital coverage, capturing lightning events only near the satellite overpass times and locations. As a result, many lightning occurrences could go undetected, not due to their absence but due to observational constraints. Consequently, treating the observed imbalanced dataset as a definitive representation of event frequency could introduce bias into model evaluations. To mitigate this, we employed a balanced dataset with equal representation of lightning (YES) and non-lightning (NO) events for both training and initial testing phases. To further ensure model robustness and evaluate performance under more realistic conditions, we conducted a supplementary experiment, motivated by the approach in a previous study [17], using the same balanced training set, but a test set reflecting the true (practical) class distribution, with 75–80% no-lightning events. As shown in Supplementary Figure S3, the XGBoost_best model maintained consistent performance, with an accuracy of 0.62 and POD of 0.63, with somewhat lower precision, which depended on the quantity of real YES lightning counts fed to the model test dataset. This indicated a reliable generalization of climatologically realistic scenarios. While adapting our original approach to such realistic testing distributions may have altered the results slightly, addressing this would have required modifications to the ML framework, such as introducing class imbalance penalties or adjusting the decision threshold, to better evaluate metrics like precision, depending on the acceptable trade-off between false alarms and missed detections. However, these aspects can be explored in more detail in future studies.

4. Conclusions

The present study demonstrated the potential of ML models to predict lightning activity in the Third Pole region by leveraging dynamically relevant atmospheric variables (predictors). Among the six evaluated models (ET, RF, XGBoost, GB, SVM, KNN) applied across four experimental setups, the best performance was achieved when training with all 25 predictors (accuracy = 0.64, POD = 0.66). However, a comparable performance (accuracy = 0.58, POD = 0.62) was obtained using only two spatially and temporally averaged predictors, CP and LPI, with just a 5–6% reduction in accuracy. This highlights the efficiency and practical utility of using simplified input features, which significantly reduces computational costs while retaining predictive skill. Figure 4 shows that when all 25 predictors were used, the feature significance of CP and LPI were minimal, likely overshadowed by other features. However, when CP and LPI were used exclusively in a physics-based approach, the model achieved a POD of ∼0.60. The results also show that applying spatiotemporal averaging, particularly at 60 km (POD = 0.60) and 90 km (POD = 0.62) scales (Figure 3 and Figure 6), consistently improved all model performances. Among the models evaluated, XGBoost emerged as the top-performing model, with TOPSIS scores of 0.87 (S1-60) and 0.92 (S1-90). Spatial evaluation showed that the XGBoost model successfully identified lightning hotspots aligned with mountainous topography while correctly classifying stable regions, though some false alarms were encountered at some point (Figure 9e). The FN category shown in Figure 9f likely reflected missed lightning detections due to the spatio-temporal limitations of the satellite instrument. This limitation could be mitigated by incorporating continuous ground-based lightning observations into the training process, thereby enhancing model accuracy. Future work should also consider extending the training data period and adopting hybrid modeling approaches that integrate additional physics-aware parameters to improve predictive skill, particularly over regions with complex topography.