Explainable Learning Framework for the Assessment and Prediction of Wind Shear-Induced Aviation Turbulence

Abstract

Wind shear-induced aviation turbulence (WSAT) remains a major safety concern during approach and takeoff phases at complex terrain airports. This study develops an interpretable Explainable Boosting Machine (EBM) framework to classify WSAT events at Hong Kong International Airport (HKIA). The framework integrates Differential Evolution with HyperBand (DEHB) for hyperparameter tuning and applies multiple data balance methods such as SMOTE, Borderline SMOTE, Safe-Level SMOTE, and G-SMOTE. The dataset consists of Pilot Reports (PIREPs) collected between 1 January 2007 and 31 July 2023, with 6838 wind shear events that include variables that relate to wind shear magnitude, altitude, runway distance, rainfall condition, and causal factors. Among all configurations, the EBM tuned via DEHB and trained with SMOTE-treated data achieved the highest predictive performance with BA = 0.710, MCC = 0.321, and G-Mean = 0.708, higher than untreated and other balance variants. EBM-based interpretation showed that wind shear altitude and wind shear magnitude were key predictors, and their interaction reflected a nonlinear pattern where WSAT probability rose under moderate-to-high shear conditions (wind shear altitude ≈ 0.5–2.5 and magnitude ≈ 30–35 knots). The DEHB-optimized EBM–SMOTE framework provides a transparent interpretive foundation for WSAT risk assessment and advances quantitative evaluation in aviation meteorology.

1. Introduction

Wind shear-induced aviation turbulence (WSAT) is one of the most critical aerodynamic hazards in flight operations. It results from abrupt spatial or temporal variations in wind speed vectors, which cause rapid changes in lift, thrust, and aircraft attitude, thereby reducing flight path stability. WSAT is most hazardous during low-altitude flight phases such as approach and takeoff, when aircraft operate with limited energy margins and minimal recovery altitude. The International Civil Aviation Organization [1] defines low-level wind shear as that occurring below 1600 feet above ground level or within 3 nautical miles of a runway threshold. In such zones, vertical and horizontal wind gradients can generate severe turbulence that often forces pilots to conduct go-around or missed-approach maneuvers to maintain safety. Aviation Turbulence Intensity (ATI) is expressed through the cube root of the Eddy Dissipation Rate (EDR), which represents the rate at which turbulent kinetic energy cascades and dissipates into smaller eddies within the atmospheric boundary layer [2,3]. The EDR metric provides a standardized, aircraft-independent measure of turbulence intensity that allows uniform reporting across different aircraft and flight conditions [4]. Low-to-moderate or insignificant aviation turbulence corresponds to EDR values between 0.3 and 0.5 m²s⁻³, while values above 0.5 m²s⁻³ indicate significant turbulence [5]. Such conditions can lead to rapid altitude fluctuations, increased control input requirements, and high structural loads on the aircraft.

Among major airports vulnerable to wind shear and low-level aviation turbulence, Hong Kong International Airport (HKIA) is one of the most affected [6,7,8]. Due to frequent events, HKIA has deployed the Wind Shear and Turbulence Warning System (WTWS), Doppler Light Detection and Ranging (Doppler LiDAR), and dense anemometer networks to monitor, identify, and warn of hazardous wind phenomena in real time [9,10]. These systems form one of the most comprehensive aviation meteorological infrastructures worldwide. In addition to these automated detection systems, Pilot Reports (PIREPs) act as an important operational tool for identifying and verifying wind shear and turbulence events [11]. Each PIREP provides a direct in-flight observation from flight crews on conditions such as turbulence, wind shear, or convective activity that may not be captured by ground-based sensors. At HKIA, where terrain-induced flow distortion occurs frequently, Pilot Reports (PIREPs) furnish real-time validation for WTWS and LiDAR-based alerts. Since operations began in July 1998, statistical assessments show that roughly one in every 500 arrival or departure operations has encountered significant wind shear, whereas about 1 in 3500 flights has reported significant aviation turbulence, on average [12]. However, despite their operational value in situational awareness, these systems lack inherent predictive capability, which establishes the necessity for advanced data-driven forecast strategies.

In recent years, the use of Artificial Intelligence (AI) has expanded rapidly across multiple domains, including healthcare and medicine [13,14,15], finance and economics [16,17], and engineering [18,19]. This growing prominence of AI is primarily due to the exponential increase in data availability, diversity, and computational capability, which together enable the development of advanced analytical frameworks capable of processing complex datasets with high precision and automation [20,21]. Although AI has shown remarkable success across many disciplines, its use within aviation has grown rapidly in recent years, with research now centered on meteorological forecast tasks, turbulence detection work, and improvements in operational safety. For instance, Stacked Temporal Convolutional Networks and Extreme Gradient Boosting Framework (TCNs + XGBoost) have been used for low-level wind shear prediction [22], Artificial Neural Networks (ANNs) for short-term wind gust forecasting [23], and Random Forest (RF) and Long Short-Term Memory (LSTM) [24], as well as hybrid Convolutional Neural Network–Long Short-Term Memory (CNN–LSTM) architectures for visibility forecasting and classification [25]. In addition, Principal Component Analysis (PCA) combined with the K-Means clustering method has been applied for the prediction of aviation turbulence risk [26]. Similarly, AI methods have been employed to improve airport efficiency and flight risk management. The decision tree (DT) approach has been used for estimating runway occupancy time [27], while Support Vector Machines (SVMs) [28], unsupervised clustering algorithms, including Density-Based Spatial Clustering of Applications with Noise (DBSCAN) [29], and CatBoost [30] have been utilized for flight delay prediction. Furthermore, Bayesian Neural Networks (BNNs) have been applied for the prediction of flight trajectory [31]. For Low-Visibility Conditions (LVCs), a decision support system based on the Analog Ensemble (AnEn) strategy was developed, which predicts LVCs up to 24 h ahead [32].

Despite the growing use of AI in aviation analytics, two major challenges remain in predicting WSAT from PIREP data. The first challenge lies in the imbalanced distribution of aviation turbulence classes, where reports of low/moderate or insignificant turbulence greatly exceed those of severe or significant turbulence. This imbalance biases the learning process and limits the predictive accuracy of models for the most safety-critical events. The second challenge concerns the lack of interpretability in conventional AI algorithms. Many existing machine learning models operate as opaque “black-box” systems [33,34] that provide predictions without revealing how different variables affect the final output, which restricts their practical value. To address these issues, this study employs the Explainable Boosting Machine (EBM), a transparent glass-box model within the Generalized Additive Model (GAM) structure, which combines interpretability with stronger predictive capacity [35,36,37,38]. To correct class imbalance in the PIREP dataset, SMOTE and its variants, including Borderline SMOTE, Safe-Level SMOTE, and G-SMOTE, are used to generate synthetic minority samples so the classifier forms a more balanced decision boundary [39,40,41,42]. Hyperparameter tuning is carried out using Differential Evolution with HyperBand (DEHB), which refines model configuration and improves prediction while reducing overfit risk [43]. The major contributions of this study are summarized as follows:

• Development of an interpretable EBM-based framework for predicting WSAT in the vicinity of airport runways.
• Application of data balancing techniques, including SMOTE and its advanced variants, to address class imbalance within the PIREP dataset and improve EBM model performance.
• Provision of transparent feature-level interpretation of WSAT prediction outcomes through the inherently explainable structure of the EBM.

The remainder of this paper is structured as follows. Section 2 provides a case study description as well as a theoretical overview of the DEHB–EBM framework with data balancing strategies. Section 3 discusses the EBM analysis results and its interpretation, while Section 4 concludes with recommendations for future research.

2. Materials and Methods

2.1. Case Study Description

In this study, HKIA is taken as the case location as it is situated within one of the most aerodynamically complex operational environments in global civil aviation due to the strong terrain–flow interaction around Lantau Island [44]. It is directly influenced by complex terrain features, including Tai Tung Shan (871 m), Yi Tung Shan (747 m), and Lantau Peak (934 m), which induce strong orographic flow disturbances, as shown in Figure 1. Under prevailing southerly or southeasterly synoptic flows, the surrounding terrain acts as a major perturbation source, which forces abrupt displacement of the approaching synoptic flow. This displacement then produces mechanically induced turbulence, rotor-type circulations, and localized wind shear within the approach and departure corridor [7]. The HKIA-based PIREPs also revealed that most wind shear events cluster very near ground level below 200 m, which shows the highest hazard density, as shown in Figure 2 [45]. The frequency drops rapidly with altitude, which shows far fewer reports beyond 600 m. This pattern reflects the aerodynamic instability that is strongest close to the runway environment during the takeoff and landing phases. Furthermore, studies also revealed that based on HKIA-based PIREPs, runway corridor 07LA at HKIA experienced the highest number of wind shear events, while some corridors such as 07LD and 25LD had very few, as shown in Figure 3 [46]. Similarly, wind shear occurrence shows clear seasonal variation across months at HKIA. Occurrence reaches higher levels in spring months, with April and March at the top range, as shown in Figure 4. Winter months remain low, while late summer and early autumn months show moderate levels.

2.2. Theoretical Overview of the DEHB–EBM Framework

The proposed framework integrates the EBM with DEHB and SMOTE and its variants to improve interpretability and prediction reliability in WSAT classification using PIREP data. The EBM provides a transparent additive structure that captures both nonlinear and interaction effects among features [48,49,50], while DEHB determines the best model hyperparameters [43]. SMOTE and its variants address class imbalance by creating synthetic minority samples, which allows the model to better detect significant WSAT events [51]. A 10-fold cross-validation (CV) procedure is applied to validate generalization capability and avoid bias from single-split evaluation. Together, they form an interpretable and data-driven analytical framework suitable for WSAT classification. The conceptual workflow is shown in Figure 5, and the theoretical formulation is presented below.

2.2.1. Explainable Boosting Machine Model (EBM)

The EBM is a GAM with pairwise interaction terms, designed to provide interpretability while preserving predictive flexibility. Given a set of predictor variables $X=\left[x_1,x_2,\dots ,x_n\right]$ and a target variable $y$, EBM models the prediction function as Equation (1).

$$\widehat{y}=f\left(X\right)={\beta }_0+{\displaystyle \sum _{i=1}^n{f}_i(x_i)+}{\displaystyle \sum _{i<j}{f}_{ij}(x_i,x_j)}$$

(1)

where ${\beta }_0$ is the global intercept, $f_i(x_i)$ represents the univariate shape function that captures the nonlinear contribution of feature $x_i$, and $f_{ij}$ denotes pairwise interaction terms between features $x_i$ and $x_j$.

Each univariate and interaction function is learned through gradient boosting of shallow decision trees trained in a cyclic additive manner. This structure allows EBM to approximate complex relationships while retaining interpretability. In the case of WSAT, the probability of significant aviation turbulence $\left({\psi }_{sig}\right)$ is modeled through the logistic link function, given as Equation (2).

$$P\left(y=1|X\right)={\displaystyle \frac{1}{1+e^{-\widehat{y}}}}$$

(2)

The predicted turbulence class $\tilde{y}$ is then determined by a threshold applied to the probability as shown by Equation (3).

$$y=\left\{\begin{array}{cc}1,\hfill & if P(y=1|X)\ge \tau ,\\ 0,\hfill & \mathrm{otherwise}.\end{array}\right.$$

(3)

where $\tau$ is the decision threshold, generally set to 0.5 for binary classification.

2.2.2. Data Balancing Through SMOTE and Its Variants

Due to the class imbalance in PIREP data, where instances of low or insignificant aviation turbulence $\left({\psi }_{in-sig}\right)$ dominate over significant turbulence $\left({\psi }_{sig}\right)$ events, data balancing is applied before model training. The SMOTE generates artificial instances for the minority class by interpolating between existing minority instances. Given a minority class instance $x_i$ and one of its $k-\mathrm{nearest}$ neighbors $x_{i,NN}$, the synthetic instance $x_{new}$ is created, as shown in Equation (4).

$$x_{new}=x_i+\lambda \left(x_{i,NN}-x_i\right),\lambda \sim U\left(0,1\right)$$

(4)

where $\lambda$ is a random number drawn from a uniform distribution $\mathrm{U}\left(0,1\right)$. This formulation makes sure that the new instances are distributed along the line segment connecting $x_i$ and its nearest neighbor, which enriches the minority class without duplication.

To enhance the minority class representation, four oversampling techniques are applied, including SMOTE, Borderline SMOTE, Safe-Level SMOTE, and G-SMOTE. SMOTE generates synthetic instances by interpolating between minority instances and their nearest neighbors to create a balanced dataset. Borderline SMOTE concentrates on minority instances near the decision boundary to strengthen class distinction and reduce misclassification risk. Safe-Level SMOTE adjusts the number of synthetic instances according to the safety level of each instance, minimizing the risk of noise amplification. G-SMOTE employs geometric constraints around minority instances to generate adaptive synthetic instances within a safe feature space, improving the diversity and boundary learning. The degree of oversampling $\left(\alpha \right)$ determines the number of synthetic instances added, as given in Equation (5).

$$N_{\mathrm{new}}=\alpha \times N_{\mathrm{minority}}$$

(5)

where $N_{\mathrm{minority}}$ is the number of original minority class instances.

2.2.3. DEHB for Hyperparameter Tuning of EBM

Hyperparameter tuning plays an important role in achieving high predictive performance for EBM. Traditional methods such as grid search or random search often require extensive computational effort and fail to efficiently explore complex, high-dimensional search spaces [52]. To address this limitation, the DEHB algorithm provides a reliable and resource-efficient approach for hyperparameter optimization by combining the global exploration capability of DE with the adaptive resource allocation mechanism of Hyperband. The DEHB framework provides several advantages for EBM tuning, including hybrid efficiency through the integration of exploration and resource allocation. Through this hybrid approach, DEHB provides a powerful and effective means of optimizing EBM hyperparameters while preserving interpretability and enhancing generalization capability.

Differential Evolution Component

The DEHB algorithm maintains a population of candidate hyperparameter vectors, denoted as $\mathrm{\Theta }$, given as Equation (6).

$$\mathrm{\Theta }=\left\{{\theta }_1,{\theta }_2,\dots ,{\theta }_N\right\}$$

(6)

where $N$ is the population size and each ${\theta }_i$ represents a vector of hyperparameters for the EBM.

At each generation $g$, new candidate solutions are generated through mutation and crossover operations. For a target vector ${\theta }_i^g$, a donor vector $v_i^g$ is created, as shown by Equation (7).

$$v_i^g={\theta }_{r1}^g+F\left({\theta }_{r2}^g-{\theta }_{r3}^g\right)$$

(7)

where ${\theta }_{r1}^g,{\theta }_{r2}^g,{\theta }_{r3}^g$ are distinct randomly selected population members, and $F\in \left[0,1\right]$ is the scaling factor controlling mutation strength.

The crossover step creates a trial vector $u_i^g$, as illustrated by Equation (8).

$$u_{i,j}^g=\left\{\begin{array}{cc}{v}_{i,j}^g,& if ran{d}_j(0,1)\le CR or j=j_{rand},\\ {\theta }_{i,j}^g,& otherwise\end{array}\right.$$

(8)

where $CR\in \left[0,1\right]$ is the crossover probability, and $j_{rand}$ makes sure that at least one parameter is inherited from the donor.

The selection step determines whether the trial vector replaces the target vector, shown by Equation (9).

$${\theta }_i^{g+1}=\left\{\begin{array}{cc}{u}_i^g,& if f\left(u_i^g\right)\le f\left({\theta }_i^g\right),\\ {\theta }_i^g,& otherwise\end{array}\right.$$

(9)

Hyperband Component

The Hyperband mechanism allocates computational resources efficiently across candidate configurations by performing successive halving. Each configuration is initially assigned a small training budget, and only the best-performing fraction proceeds to higher budgets.

For a given budget $b$ and reduction factor $\mathrm{\Gamma }>1$, Hyperband is defined as Equation (10).

$$n_i={\displaystyle \frac{n_{\max }}{{\mathrm{\Gamma }}^i}},b_i=b_{\max }{\mathrm{\Gamma }}^i,i=0,1,\dots ,s,$$

(10)

where $n_i$ is the number of configurations at stage $i$, and $b_i$ is the corresponding computational budget.

Combined DEHB Procedure

The DEHB algorithm integrates the evolutionary update rule of DE with the early-stopping mechanism Hyperband. Each population member’s fitness evaluation corresponds to a model trained on a given resource budget $b_i$. Configurations with lower performance are eliminated early, while promising configurations are further improved through DE mutation and crossover operations. Formally, DEHB can be represented as Equation (11).

$$\theta *=\underset{\theta \in \mathrm{\Theta }}{\arg \min f\left(\theta ,b_i\right)}$$

(11)

Equation (11) is subject to adaptive budget allocation and evolutionary search updates until convergence or maximum resource consumption.

Model Output and Interpretability

Once optimized, the DEHB–EBM framework produces both predicted classifications and feature contribution profiles. The effect of each factor can be visualized through partial dependence plots of $f_i\left(x_i\right)$ and pairwise surfaces of $f_{ij}\left(x_i,x_j\right)$. These representations quantitatively describe how each meteorological and operational factor influences aviation turbulence probability, as illustrated in Equation (12).

$$WSAT=\left\{\begin{array}{cc}{\psi }_{sig},& \mathrm{if} \widehat{y}\ge \tau ,\\ {\psi }_{in-sig},& \mathrm{otherwise}.\end{array}\right.$$

(12)

2.3. Performance Measures

The predictive and classification performance of the EBM was assessed using six metrics, including precision $\left(\mathit{Pn}\right)$, recall $\left(Rl\right)$, balanced accuracy $\left(BA\right)$, geometric mean $\left(GM\right)$, Matthews Correlation Coefficient (MCC), and Receiver Operating Characteristic (ROC) curve. These metrics were computed from the confusion matrix, which compares predicted and actual class outcomes. The corresponding performance formulas are presented in

Table 1. Summary of classification performance metrics used for model evaluation.

Metric	Description	Expression
$\mathrm{Precision} \left(Pn\right)$	Measures the proportion of correctly predicted positive instances out of all predicted positives.	$Pn={\displaystyle \frac{P_C}{P_C+P_E}}$
$\mathrm{Recall} \left(Rl\right)$	Quantifies how effectively the model identifies actual positive instances within the dataset.	$Rl={\displaystyle \frac{P_C}{P_C+N_E}}$
$\mathrm{Balanced} \mathrm{Accuracy} \left(BA\right)$	Measures the average recall for each class, which provides an unbiased estimate of overall classification.	$BA={\displaystyle \frac{1}{2}}\left({\displaystyle \frac{P_C}{P_C+N_E}}+{\displaystyle \frac{N_C}{N_C+P_E}}\right)$
$\mathrm{Geometric} \mathrm{Mean} \left(GM\right)$	Represents the balance between precision and recall by taking their geometric mean.	$GM=\sqrt{Pn\times Rl}$
Matthews Correlation Coefficient (MCC)	Evaluates the overall quality of binary classifications by considering true and false positives and negatives.	$MCC={\displaystyle \frac{\left(P_C\times N_C\right)-\left(P_E\times N_E\right)}{\sqrt{\left(P_C\times P_E\right)\left(P_C\times N_E\right)\left(N_C\times P_E\right)\left(N_C\times N_E\right)}}}$
Receiver Operating Characteristic (ROC) Curve	Plots the True Positive Rate (TPR) against the False Positive Rate (FPR) at various threshold settings.	—

Note: $P_C:$ True Positives, $P_E:$ False Positives $\left(P_E\right)$, $N_C:$ True Negatives, and $N_E:$ False Negatives.

.

3. Analysis and Results

This section presents the analytical outcomes derived from HKIA-based PIREPs collected between 1 January 2007 and 31 July 2023. During this period, PIREPs reported a total of 6838 WSAT events from both outbound and inbound flight operations. Among these, 1169 WSAT events were classified as ${\psi }_{sig}$ and the remaining 5668 were classified as ${\psi }_{in-sig}$. A binary classification framework was established in which each observation was assigned a categorical label, i.e., “1” represents ${\psi }_{sig}$ (minority class), while “0” represents ${\psi }_{in-sig}$ (majority class), as defined in Equation (12).

Table 2. Description and coding details of the factors used in this study.

Factors	Symbol	Description and Coding Details
Wind Shear Magnitude	${\mu }_{WS}$	Represents the magnitude of wind shear in knots, measured as a continuous value.
Wind Shear Distance from Runway	${\delta }_{WS}$	Indicates the distance of wind shear occurrence from the runway. If coded 0, it represents at runway; if coded 1, it represents 0–1 NM from runway; if coded 2, it represents 1–2 NM from runway; if coded 3, it represents 2–3 NM from runway; and if coded 4, it represents Beyond 3 NM.
Wind Shear Altitude	${\eta }_{WS}$	Indicates the altitude range at which wind shear occurred. If coded 0, it represents 0–399 ft; if coded 1, it represents 400–799 ft; if coded 2, it represents 800–1199 ft; and if coded 3, it represents 1200–1600 ft.
Wind Shear Causes	${\varphi }_{WS}$	Represents the primary cause of wind shear. If coded 0, it corresponds to terrain; if coded 1, it corresponds to sea breeze; and if coded 2, it corresponds to gust front.
Rainfall Condition	${\rho }_{WS}$	Indicates the rainfall condition at the time of the event. If coded 0, it represents no rain; if coded 1, it represents rain.
Aviation Turbulence Category	${\psi }_{sig}/{\psi }_{in-sig}$	$\mathrm{If} \mathrm{coded} 0, \mathrm{it} \mathrm{denotes} {\psi }_{in-sig}$$; \mathrm{if} \mathrm{coded} 1, \mathrm{it} \mathrm{denotes} {\psi }_{sig}$

provides the detailed description and coding scheme of the factors extracted from HKIA-based PIREPs that were used as input factors in the classification model.

Figure 6a–e present the frequency distributions of wind shear-related factors categorized by aviation turbulence severity $\left({\psi }_{sig}/{\psi }_{in-sig}\right)$. Figure 6a shows that most turbulence events occurred at altitudes below 400 ft, with 3952 ${\psi }_{in-sig}$ and 503 ${\psi }_{sig}$ cases, which indicates that near-surface wind shear contributes most to turbulence generation. Figure 6b shows that the majority of ${\mu }_{WS}$ fall within the 15–20 knots range (4806 ${\psi }_{in-sig}$ and 971 ${\psi }_{sig}$). Figure 6c compares ${\rho }_{WS}$ and shows that most ${\psi }_{in-sig}$ events occurred under no-rain conditions. Furthermore, Figure 6d,e illustrate the influence of ${\varphi }_{WS}$ and ${\delta }_{WS}$, respectively. Terrain-induced wind shear dominated the dataset, with 3671 ${\psi }_{in-sig}$ and 746 ${\psi }_{sig}$ cases, followed by sea breeze (1272 ${\psi }_{in-sig}$ t and 360 ${\psi }_{sig}$) and gust front (725 ${\psi }_{in-sig}$ and 63 ${\psi }_{sig}$). Regarding ${\delta }_{WS}$, most reports occurred at or near the runway, with 1815 ${\psi }_{sig}$ and 415 ${\psi }_{in-sig}$ cases “at runway” and 1857 ${\psi }_{in-sig}$ and 317 ${\psi }_{sig}$ cases within 0–1 NM from runway.

Table 3. Class distribution before and after data balancing.

Method	${\mathit{\psi }}_{\mathit{s}\mathit{i}\mathit{g}}$		${\mathit{\psi }}_{\mathit{i}\mathit{n}-\mathit{s}\mathit{i}\mathit{g}}$
	Before Balance	After Balance	Before Balance	After Balance
SMOTE	17.1% (818)	50.0% (3967)	82.9% (3967)	50.0% (3967)
Borderline SMOTE	17.1% (818)	50.0% (3967)	82.9% (3967)	50.0% (3967)
Safe-Level SMOTE	17.1% (818)	50.0% (3967)	82.9% (3967)	50.0% (3967)
G-SMOTE	17.1% (818)	50.0% (3967)	82.9% (3967)	50.0% (3967)

summarizes the class distribution before and after data balancing.

3.1. EBM Training and Testing

The EBM model was applied to classify WSAT severity using the extracted predictor variables. The full dataset contained 6837 samples, with 5668 ${\psi }_{in-sig}$ cases (82.9%) and 1169 ${\psi }_{sig}$ cases (17.1%), as shown in Figure 6. The data was divided into a 70% training subset (3967 ${\psi }_{in-sig}$ and 818 ${\psi }_{sig}$) and a 30% testing subset (1701 ${\psi }_{in-sig}$ and 351 ${\psi }_{sig}$), while retaining the original class proportions. Since the imbalance existed within the training subset, oversampling was applied only to this portion. SMOTE and its three variants (Borderline SMOTE, Safe-Level SMOTE, and Geometric SMOTE) were used, resulting in a balanced training distribution of 3967 ${\psi }_{sig}$ and 3967 ${\psi }_{in-sig}$ cases, as shown in Table 3.

Furthermore, in order to improve the predictive capability of the EBM, the DEHB strategy was adopted for hyperparameter optimization across four data balancing techniques, including SMOTE, Borderline SMOTE, Safe-Level SMOTE, and G-SMOTE, as shown in Figure 7a–d. Each subplot in Figure 7 presents the search range explored for six principal hyperparameters and highlights the optimal configuration (green marker) identified through the DEHB process.

The tuned hyperparameters include Max Bins, Learning Rate, Max Leaves, Min Samples Leaf, Outer Bags, and Inner Bags, which govern the interpretability, convergence behavior, and generalization capability of the EBM model. The Max Bins parameter determines the discretization level of continuous features, which influences how the model captures feature interactions. The Learning Rate controls the contribution of each iteration during boosting, balancing convergence speed and stability. Max Leaves and Min Samples Leaf define the complexity and depth of individual trees in the additive model, while Outer Bags and Inner Bags relate to the bootstrapping mechanism that improves ensemble stability and reduces overfitting.

For the SMOTE-balanced dataset, DEHB identified the optimal configuration as Max Bins = 90.021, Learning Rate = 0.043, Max Leaves = 16.225, Min Samples Leaf = 35.987, Outer Bags = 2.371, and Inner Bags = 9.099. For the Borderline SMOTE dataset, the optimal parameters were Max Bins = 164.444, Learning Rate = 0.014, Max Leaves = 24.502, Min Samples Leaf = 49.661, Outer Bags = 19.910, and Inner Bags = 5.715. Under the Safe-Level SMOTE configuration, DEHB achieved the best performance with Max Bins = 363.634, Learning Rate = 0.018, Max Leaves = 22.648, Min Samples Leaf = 15.346, Outer Bags = 2.945, and Inner Bags = 9.179. Finally, for the G-SMOTE case, the optimal settings obtained were Max Bins = 170.796, Learning Rate = 0.039, Max Leaves = 15.362, Min Samples Leaf = 27.018, Outer Bags = 17.111, and Inner Bags = 6.720.

The performance assessment of the EBM using untreated data (without any class balancing) is illustrated in Figure 8a,b. The results reflect the baseline predictive behavior of the EBM model when trained on the original, imbalanced dataset, where the majority of cases correspond to ${\psi }_{in-sig}$ and the minority to ${\psi }_{sig}$. The confusion matrix in Figure 8a shows that the model correctly classified 1675 instances (81.6%) as ${\psi }_{in-sig}$, while 52 instances (2.5%) were correctly identified as ${\psi }_{sig}$. However, 299 ${\psi }_{sig}$ cases (14.6%) were misclassified as ${\psi }_{in-sig}$, which indicates a tendency of the model to favor the majority class. Only 26 ${\psi }_{in-sig}$ cases (1.3%) were incorrectly predicted as ${\psi }_{sig}$. This imbalance in predictions shows that the untreated EBM model struggles to identify minority-class ${\psi }_{sig}$ events effectively. The ROC curve in Figure 8b shows an Area under the Curve (AUC) of 0.788, which indicates reasonable discrimination between ${\psi }_{sig}$ and ${\psi }_{in-sig}$ despite class imbalance. For EBM on the untreated dataset, BA (0.566), MCC (0.262), and G-Mean (0.382) collectively reveal moderate predictive performance, with limited sensitivity toward ${\psi }_{sig}$ cases.

The performance of the EBM after applying the SMOTE technique is illustrated in Figure 9a,b. The confusion matrix in Figure 9a shows an improved ability to detect significant turbulence events, with 266 (13.0%) correctly identified ${\psi }_{sig}$ cases and a reduction in missed detection compared with the untreated model. The ROC curve in Figure 10b achieved an AUC of 0.785, indicating comparable overall discriminative ability but improved class balance. The BA (0.710) and G-Mean (0.708) increased slightly, and the MCC (0.321) also improved, which indicates that SMOTE allowed the EBM to handle the minority class more effectively without reducing performance across the full dataset. The EBM with Borderline SMOTE achieved balanced performance, as shown in Figure 10a,b. The model correctly identified 261 (12.7%) ${\psi }_{sig}$ cases with an AUC of 0.765, BA of 0.701, G-Mean of 0.699, and MCC of 0.307, which indicated stable but slightly lower performance than standard SMOTE due to borderline sample complexity.

The Safe-Level SMOTE and G-SMOTE results are presented in Figure 11a,b and Figure 12a,b, respectively. For Safe-Level SMOTE, the EBM achieved improved minority class recognition, correctly classifying 243 (11.8%) ${\psi }_{sig}$ cases. The model yielded an AUC of 0.772, BA of 0.690, G-Mean of 0.690, and MCC of 0.295, which reflected balanced predictive behavior with steady sensitivity to ${\psi }_{sig}$ events. For G-SMOTE, the EBM maintained steady discrimination performance, which correctly classified 196 (9.6%) ${\psi }_{sig}$ cases with an AUC of 0.775, BA of 0.667, G-Mean of 0.658, and MCC of 0.279, as shown in Figure 12a,b.

Based on the above results, the EBM model trained with SMOTE-treated data achieved the best overall performance among all data balancing strategies and can be interpreted both globally and locally.

3.2. EBM-Based Interpretation

The EBM model trained with SMOTE-treated data identified wind shear altitude $\left({\eta }_{WS}\right)$ and magnitude $\left({\mu }_{WS}\right)$ as the two most influential predictors in distinguishing between ${\psi }_{sig}$ and ${\psi }_{in-sig}$ categories, as shown in Figure 13a. The global feature importance plot reveals that these variables, along with their interaction term $\left({\eta }_{WS}\times {\mu }_{WS}\right)$, dominate model explainability and indicate that both the strength and vertical position of wind shear critically affect aviation turbulence severity. Similar conclusions were reported in prior research that examined atmospheric turbulence dynamics, which identified vertical wind shear instabilities as one of the primary sources of aviation turbulence, along with convective and mountain wave activities [53]. A previous case study at the HKIA further substantiates this relationship. The mountainous terrain of Lantau Island, located south of the airport, often disrupts airflow under specific wind directions, generating marked low-level wind shear and turbulence during aircraft approach and departure [54].

Other factors including ${\rho }_{WS}$, ${\delta }_{WS}$, and ${\varphi }_{WS}$ also contributed meaningfully, though to a lesser extent. The clear separation between single-feature and interaction terms implies that the model effectively captures nonlinear and interdependent aerodynamic effects. In the ${\eta }_{WS}$ plot (Figure 13b), the model score transitions from negative to strongly positive when ${\eta }_{WS}$ exceed approximately 0.5 and maintains a stable high contribution up to 2.5 before slightly declining. This pattern indicates that low altitude between 0.5 and 2.5 considerably improves the likelihood of ${\psi }_{sig}$. Similarly, for ${\mu }_{WS}$, the model shows a nonlinear response with alternating positive and negative regions, as shown in Figure 13c. The probability of ${\psi }_{sig}$ begins to rise sharply once ${\mu }_{WS}$ exceeds 18 knots. However, at around 28–30 knots, the score drops, which implies that intermediate ${\mu }_{WS}$ may not always produce ${\psi }_{sig}$. Beyond 35 knots, a distinct positive peak appears, defining an optimal range where strong shear coincides with maximum turbulence amplification. This pattern aligns with empirical evidence from the HKIA, where recurrent low-level wind shear events of comparable magnitude result from terrain-induced flow distortion near Lantau Island [55,56,57].

Furthermore, the interaction plot between ${\eta }_{WS}$ and ${\mu }_{WS}$ reveals a strong interdependence in shaping turbulence likelihood, as shown in Figure 13d. The highest positive scores, shown in bright yellow, appear where both ${\eta }_{WS}$ and ${\mu }_{WS}$ reach moderate-to-high levels, typically around ${\eta }_{WS}$ ≈ 1.5–2.5 and ${\mu }_{WS}$ ≈ 30–35 knots. This indicates that when elevated wind shear occurs concurrently with a strong magnitude, the conditions are most conducive to ${\psi }_{sig}$ formation. On the other hand, the darker regions at the lower-left (low ${\eta }_{WS}$ and ${\mu }_{WS}$) and upper-left corners (high ${\mu }_{WS}$ but low ${\eta }_{WS}$) correspond to negative scores, which means that such combinations are likely to produce ${\psi }_{in-sig}$.

4. Conclusions and Recommendations

This study proposed a framework that integrates the EBM model with several data balancing strategies, such as SMOTE, Borderline SMOTE, Safe-Level SMOTE, and G-SMOTE, for WSAT classification and interpretation. Results indicate that an imbalance in the WSAT dataset affects model performance and that controlled synthetic enrichment of the minority class leads to higher accuracy and clearer interpretability outcomes. The EBM model trained with SMOTE reached the highest performance with a BA of 0.710, an MCC of 0.321, and a G-Mean of 0.708, with Borderline SMOTE, Safe-Level SMOTE, and G-SMOTE producing lower performance values. This confirms that the minority WSAT class required augmentation to reveal its underlying structure in a reliable manner.

Interpretability results from the EBM framework revealed that wind shear altitude and wind shear magnitude act as the dominant variables that drive WSAT classification. Their joint effect illustrates a nonlinear response area, where the highest WSAT probability takes place at a moderate to higher altitude range (0.5–2.5 units) together with wind shear magnitude between 30 and 35 knots. This interaction gives physical insight into the aerodynamic mechanisms that support WSAT generation close to the ground within runway operational zones.

4.1. Limitations of This Study

This study is based on HKIA only, which is characterized by strong terrain–flow interaction unique to Lantau Island. This domain specificity limits direct transferability of the learned WSAT behavior to airports with weaker or different topographic influence, different turbulence climatology, or distinct operational runway layouts. The modeling was also based on PIREPs, which depend on subjective pilot reporting behavior and do not provide quantitative turbulence magnitude distribution information beyond occurrence classification.

4.2. Future Recommendations

Future studies should expand the framework across multiple airports with different climates and topographic regimes so the generalization capability can be assessed beyond a single location. The integration of higher-resolution turbulence measurements from LiDAR, the ADS-B-derived Eddy Dissipation Rate, and aircraft-mounted sensor-based measurements can strengthen objective signal representation and reduce dependence on subjective pilot-based turbulence reporting. In addition, other atmospheric state variables such as vertical shear profile, temperature gradient, and humidity stratification can raise classification reliability further. Probabilistic uncertainty quantification should be incorporated in future model development because aviation safety applications require calibrated risk estimation rather than a single deterministic outcome.