The statistical and machine learning models are a part of the data-driven models in predicting tropical cyclone genesis. The study by [20] used a statistical framework to determine important environmental factors of short-term TC genesis prediction. They have modeled the problem as a classification problem, where the validation was done using graphical model structure learning (PC algorithm) and logistic regression. However, the data imbalance and discretizations of statistical dependency are key limitations of statistical models. The study [21] formulated a statistical-dynamical tropical cyclone genesis guidance tool based on global model outputs by carrying out a multiple logistic regression. The problem of genesis prediction was conceptualized in the study as a classification problem and found major predictors in basins. Even though the forecasts were traditionally well-calibrated, they had the drawbacks of the biases in models and the prediction imperfection of regional specificity.

Although these developments have occurred, there are complex nonlinear and spatio-temporal relationships that can hardly be represented using traditional statistical models. Therefore, more advanced machine learning-based models have been applied to improve tropical cyclone genesis prediction.

Machine learning-based methods have been used for meteorological tasks. The study [22] used an ensemble approach for TC path prediction. The study [23] used a ConvGRU model for TC intensity prediction up to 24 h.

The study [24] proposes a machine learning methodology to define the possibility of mesoscale convective systems (MCSs) developing into a TC. Positioned as a classification task, it utilized AdaBoost as the classification algorithm, achieving an F1-score of 97.2 with a 6-h lead time. A major contribution is the combination of environmental and MCS-specific predictors, which provides better prediction than conventional methods such as the Genesis Potential Index (GPI).

The study [25] proposes a machine learning framework for long-term TC genesis prediction using Support Vector Machine (SVM) and AdaBoost algorithms. By accounting for five important meteorological factors, the model successfully maps input features to genesis outcomes as a classification task. It balances the data and training splits and provides better spatio-temporal precision and generalization.

The study [26] uses multiple environmental variables to forecast TC genesis with a Maximum Entropy (MaxEnt) machine learning model. The model, trained on reanalysis data and tested on historical observations, provides more spatially accurate potential genesis index values compared to conventional genesis indices. When applied to CMIP6 future climate scenarios, it demonstrates a statistically significant reduction in TC genesis prediction error, indicating a complex nonlinear correlation between potential intensity and TC genesis.

The study [27] put forward a multiscale regression model to predict the frequency of TC genesis (TCGF) in the Northern Hemisphere. The model represents variability of observed TCGF in 1960–2019 in six climate-related predictors at interannual, interdecadal, and global warming time scales. Together with CMIP6 climate data, it forecasts basin-scale climate change in the near future, which gives a hybrid statistical-dynamical model that has a better decadal predictivity than longer-range historical trajectory-based prediction methods.

The study [28] proposes a novel methodology that combines CNN and Long Short-Term Memory (LSTM) within a deep learning framework to predict typhoons. It combines 3D and 2D CNNs to get spatial information based on atmospheric and ocean surface data, and LSTM acquires time-dependent information. The model is assessed using three regional datasets, and it is found to be better than the conventional numerical, statistical, and machine learning approaches, and this offers an innovative spatio-temporal reservoir-based technique to typhoon prediction, disaster warning, and preparedness.

The study [29] introduces a deep learning model, TCGNet, that combines Convolutional Neural Networks (CNNs) with channel- and spatial-attention mechanisms to emulate tropical cyclone genesis (TCG) under climate change conditions. Based on two historical and CMIP6 datasets, TCGNet shows improved prediction ability and generalization ability compared to the conventional statistical models. It eliminates the need for manual feature selection, captures complex, spatio-temporal patterns in the environment, and extrapolates the lower frequency of TCG in the case of increased carbon emissions.

The study [15] establishes an enhanced spatiotemporal deep learning model, TC-Pred, in which TC intensity is predicted at high resolution. The most important contribution is a feature aggregation mechanism for multi-source environmental data and the incorporation of a convolutional transformer module into the sequence-to-sequence architecture to reduce sequential dependency problems. The task is positioned as a regression task by forecasting future TC intensities. The model uses the ConvGRU algorithm embedded within a deep neural network.

The study [30] leverage artificial intelligence to increase the predictability of extreme events on subseasonal to decadal timescales through the hybrid integration of artificial intelligence and climate models, along with best practices for robustness and interpretability.

The study [31] has proposed the XAI-GPI framework, which uses the combination of clustering, ensemble feature selection, and Shapley Additive Explanations (SHAP) to better estimate the genesis of tropical cyclones while providing transparent driver attribution. The study [32] suggests NeuralGCM, an ML-physics model to make efficient seasonal predictions of tropical cyclone activity with competitive performance. The study [33], the researchers suggest the use of uncertainty-sensitive cyclone forecasting using interpretability-driven white-box models, including decision trees and Bayesian rule lists.

Although the accuracy of tropical cyclone genesis prediction has been improved significantly with statistical and machine learning-based models, as shown in Table 1, they still have significant shortcomings. Most importantly, these models are binary classification models and do not directly predict TC genesis location. Additionally, these studies use a fixed grid size and fail to consider the best grid size. These gaps point to the need for a spatio-temporal modeling framework for predicting genesis coordinates as well as testing various grid sizes to identify a new grid size.

Let 𝒟CMA={(ti,lati,loni)}i=1Ndenote the CMA Best Track dataset [34], where each record captures the timestamp ti and coordinates (lati,loni) of observed TCs, and the statistics of the refined CMA data are shown in Table 2. We define a tropical disturbance Tbi as the initial record for each developing system with a classification status below the tropical depression threshold, as shown in Eq. (2).(2)Tbi=(ti,lati,loni),Tbi∈{Disturbance,Low}

These Tbi are used to extract temporal features, as shown in Eq. (3), and serve as anchors for subsequent spatio-temporal feature extraction. (3)wi=wi(T)∈RT×dwwhere T represents time, and dw represents the temporal features.

For each detected disturbance, Tbi, a square grid of size g×g (e.g., 10×10,15×15,30×30,40×40) is defined around its center (lati,loni). Let the grid be denoted as in Eq. (4):(4)𝒢i={(lati+Δym, loni+Δxn) | m,n∈[−g2,g2]}where Δym and Δxn are the spatial steps in latitude and longitude, respectively.

From the ERA5 reanalysis dataset [35], we extract multiple variables (e.g., zonal wind u, meridional wind v, temperature T, and relative humidity RH) across multiple pressure levels P={p1,p2,…,p37} at time ti.

These four variables were chosen because they are representative of the basic dynamical and thermodynamical processes that control TC genesis prediction. Previous studies have unanimously found mid-level vorticity, wind structure, temperature stratification, and moisture availability to be dominant predictors of genesis [12]. Moreover, some key indices, such as vertical wind shear, tilting, and moisture convergence, can be diagnostically established from u, v, T, and RH fields, which guarantees the physical completeness of the representation without redundancy.

For each variable vk and pressure level pj, as defined in Eq. (5):(5)xijk=vk(𝒢i,pj,ti)

This yields a 4D feature tensor for each sample i, as shown in Eq. (6):(6)𝒳i=[xijk]∈Rg×g×v×pwhere:

g×g is the spatial resolution.

Each tensor 𝒳i captures the spatial, vertical, and variable-wise atmospheric snapshot centered on the tropical disturbance at time ti, as shown in Fig. 1.

Sea surface temperature (SST) is extracted from the ERA5 single-level reanalysis dataset as a surface variable without vertical pressure dependence. Unlike multi-level atmospheric variables defined across pressure levels P={p1,…,p37}, SST is defined only at the surface level (p=0).

For each tropical cyclone instance at time ti and center (ϕi,λi), a spatial grid of size G×G (with G=3 in this study) is constructed as follows:(7)SSTi(m,n)=SST(ti,ϕi+Δϕm,λi+Δλn),m,n∈{1,…,G}

The resulting spatial tensor is represented as:(8)XiSST∈R1×G×G

To integrate SST into the temporal branch, the spatial grid is vectorized and concatenated with the temporal feature representation:(9)wi∈RT×dw⇒w~i∈RT×(dw+G2)where T denotes the temporal length, and dw represents the original temporal feature dimension.

We use a two-layer GRU to process the temporally wide input as in Eq. (10):(10)ht(1)=GRU(1)(wi)∈RT×H1,ht(2)=GRU(2)(ht(1))∈RT×H2where wi describes the temporal features at time T, ht(1) represents the intermediate hidden feature for the GRU layer, and ht(2) represents the deeper temporal representation after a second GRU layer. Final wide feature can be defined as in Eq. (11):(11)fw=hT(2)∈RH2

Each variable Xi at time step t is a 3D tensor of shape [C, H, W] and is passed through three convolutional blocks with channel-wise attention.

Convolutional Block: The initial input to the convolutional pipeline is denoted as z(0)=𝒳i(v)[:,:,:], where 𝒳i=[xi(v)]∈Rg×g×p represents the spatial grid for the vth atmospheric variable across g×g latitude-longitude cells and p pressure levels for the ith sample. The colon operator indicates that all spatial and vertical dimensions are included. This tensor is then passed through multiple convolutional layers. At the lth layer, the output feature map z(l) is computed as z(l)=ReLU(Conv(l)(z(l−1))), where Conv(l) denotes a 2D convolution operation applied with a learnable kernel, and ReLU is the Rectified Linear Unit activation function, as shown in Eq. (12). This non-linearity introduces the ability to learn complex spatial representations. (12)z(l)=ReLU(Conv(l)(z(l−1))),z(0)=𝒳i(v)[:,:,:]

Channel-Wise Attention: At each block, attention is applied as in Eq. (13). The attention map A(l) is computed by applying a 1×1 convolution followed by a sigmoid activation σ on the feature map z(l). The output feature z(l) is then scaled element-wise using ⊙ with the attention map. (13)𝒜(l)=σ(Conv1×1(z(l))),z(l)=z(l)⊙𝒜(l)

After 3 layers of Conv-Attention-Pooling, the spatial feature is flattened as in Eqs. (14) and (15):(14)ft(v)=FC(Flatten(z(3)))∈RHc

The CNN-encoded features for each variable across T are:(15){f1(v),f2(v),…,fT(v)}∈RT×Hc

Eq. (16) represents the use of a Gated Recurrent Unit (GRU) to process a sequence of feature vectors [f1(v),…,fT(v)] for each variable v. The GRU captures temporal dependencies across the sequence and outputs hidden states ht(v) that are further utilized for the final feature vector fn, as shown in Eq. (17). (16)ht(v)=GRU(v)([f1(v),…,fT(v)])∈RT×Hv (17)fn=hT(v)∈RHv

We concatenate the wide and deep features from each variable as in Eq. (18):(18)fconcat=[fw,fui,fv,ft,fr]∈RH2+4Hv

This vector is passed through two dense layers as in Eqs. (19) and (20):(19)z=ReLU(W1fconcat+b1)∈R64 (20)y^i=W2z+b2∈Rwhere yi is the predicted location of the TC genesis.

The model is trained using Mean Absolute Error (L1 loss) as in Eq. (21). (21)ℒ=1N∑i=1N|y^i−yi|

The performance of the proposed model, as shown in Table 4, indicates clear evidence that the proposed model ((u, v, r, t, wide) + attention) demonstrates strong predictive capability for TC genesis locations. Particularly, it has the lowest MAE for latitude (1.7995∘) and longitude (3.5566∘). The results indicate minimal error between predicted and actual genesis coordinates. The model achieved R2 values (0.8399 and 0.901 for latitude and longitude, respectively), implying that it explains more than 84% of latitudinal variation and almost 90% of longitudinal variation, as shown in Fig. 3. Results indicate that component-wise fusion plays an important role in model performance. Wide features capture temporal dynamics, while the deep CNN layers extract spatial patterns from multi-channel reanalysis variables (u, v, r, t). The combination of both temporal and spatial patterns achieves the lowest MAE. The fusion network enhances the capability of the model by learning complex spatio-temporal patterns for TC genesis location prediction.

The performance of the proposed STAG-Net model was assessed using two real-time cases of TC genesis: MINDULLE (2016) and YAGI (2013). In the case of TC MINDULLE, the actual genesis coordinates were (17.3∘, 142.5∘), compared to the predicted coordinates (17.88∘, 142.09∘). This results in a latitude MAE of 0.5826∘ and a longitude MAE of 0.4147∘, which reveals significant model performance. Similarly, the actual coordinates of TC YAGI were (17.8∘, 129.9∘), and the predicted coordinates by STAG-Net were (17.38∘, 130.68∘), with a latitude and longitude difference of 0.4209∘ and 0.7813∘, respectively.

These findings indicate that STAG-Net has a high ability to forecast TC genesis locations with sub-degree error in latitude and longitude, as shown in Fig. 4. Fig. 5 shows the distribution of prediction errors, demonstrating right-skewed behavior, with most predictions clustered around the median (3.91∘) and mean (4.59∘), and a small number of extreme cases exceeding 10∘.

To further examine the temporal robustness of the proposed model, the time gap between the start of the disturbance and the time taken to reach TC genesis was divided into four slots, i.e., 0–24 h, 24–48 h, 48–72 h, and more than 72 h. As can be seen from Table 5, the accuracy of the prediction gradually declines as the time interval increases. The lowest mean MAE is shown in the first 24 h (3.91∘), whereas errors are noticeably higher for events longer than 72 h (7.73∘). This trend shows that disturbances in the early stage are more predictable; on the contrary, long-duration genesis cases are more variable and uncertain, and thus, prediction error and dispersion increase.

To evaluate the effectiveness of the channel-wise attention layer, we compare the results of the STAG-Net model with ((u, v, r, t, wide) + attention) and without (u, v, r, t, wide) attention layer, as shown in Table 4. The results indicate that the baseline model obtained 1.9902∘ and 4.0082∘ MAE for latitude and longitude, and the channel-wise attention managed to improve the MAE to 1.7995∘ and 3.5567∘ for latitude and longitude, respectively. This shows a performance increase of up to 10% for latitude and 11.93% for longitude with respect to MAE. Furthermore, the model shows a performance increase of up to 19.85% and 20.25% for latitude and longitude, respectively, based on MSE.

The Gradient-weighted Class Activation Mapping (Grad-CAM) visualization, as shown in Fig. 6, highlights the spatial regions that are most important for predicting cyclone genesis for the TC named “Cecil.” The activation map indicates an area of localized focus, which indicates that the model is perceiving physically meaningful structures in the atmosphere instead of diffuse structures.

The great reason for this enhancement is that channel-wise attention enables the model to focus on the most useful feature maps at each convolutional block. In the prediction of the location of TC genesis, various meteorological factors (such as components of wind, temperature, relative humidity, etc.) play different roles at different locations. These feature channels are adaptively adjusted by the attention mechanism, which favours significant features and suppresses irrelevant or noisy ones. Such a selective focus is a part of more effective representation learning and therefore leads to less prediction error and better generalization.

To evaluate the impact of grid size in the proposed approach, we compare the performance of multiple grid sizes. In our experimental evaluation, we compare four different grid sizes (10×10,20×20,30×30,40 × 40). These grid sizes are taken from smaller to larger to evaluate the impact on TC genesis location prediction. The results shown in Table 6 and Fig. 7 indicate that the grid size (30×30) achieves an average MAE of 2.67∘, which is 7.54% lower than grid size (10×10), 6.55% lower than (20×20), and 2.5% lower than (40×40). Furthermore, the grid size (30×30) achieves an average R2 score of 0.87, which is 4.8%, 3.75%, and 1.28% higher than the other three grid sizes, respectively.

The 30×30 attention grid offers a trade-off between spatial detail and stability, allowing meaningful areas of focus to be captured without injecting too much noise, as shown in Fig. 8. Such a tendency indicates that larger grids provide a more enriched spatio-temporal setting, with additional information about the surrounding atmospheric conditions of the disturbance being used. Consequently, this makes the model more competent in capturing the dynamics; thus, learning the spatial patterns that affect cyclone formation is improved.

However, further increasing the grid size beyond 30 × 30 may not necessarily ensure further performance improvement. Extremely large grids could add non-informative or noisy environmental information from areas far away from the cyclone core dynamics, which could reduce the signal-to-noise ratio. Additionally, the computational complexity scales quadratically with the grid dimension, which could result in decreased accuracy relative to the increased training cost.

Most of the existing work on TC genesis prediction is presented in the context of a binary classification problem (i.e., genesis vs. non-genesis); therefore, these studies do not allow direct comparison with our coordinate prediction problem. Hence, we adjust the most suitable baseline models to fit our regression-based model in order to make a fair comparison.

We modified three baseline models: a temporal model (GRU) [36], Saf-Net [37], and the Dynamic Spatio-temporal model (DST) [16]. To ensure a fair comparison, only the last classification layers of the baseline models were changed, while the original feature extraction architectures were kept unchanged. Specifically, the sigmoid/softmax output layers were replaced with fully connected linear layers having two outputs (latitude and longitude), and the binary cross-entropy loss was replaced with Mean Squared Error (MSE). All convolutional, attention, and recurrent modules remained unchanged to maintain architectural consistency.

The results shown in Table 7 and Fig. 9 provide a clear indication that STAG-Net efficiently outperforms the baseline models with regard to all important performance metrics, including MAE, MSE, RMSE, and R2. The results indicate that STAG-Net achieved an average MAE of 2.678∘, which is 9.75% lower than the DST model (which achieved the highest error among the baseline models). Furthermore, the proposed model achieved 26.25%, 12.95%, and 4.27% performance improvement with respect to MSE, RMSE, and R2, respectively. This comparison shows that the combination of spatial and temporal interdependencies, along with attention mechanisms as in STAG-Net, provides a valuable contribution toward enhancing TC genesis location prediction.

The results of the study illustrate the efficiency of the proposed STAG-Net model for the prediction of TC genesis location. STAG-Net exceedingly performs better than the baseline temporal and spatio-temporal models in different evaluation metrics, namely MAE, MSE, RMSE, and R2. Such increased performance can be explained by the ability of the model to simultaneously take into account the spatial and temporal patterns using u-wind, v-wind, relative humidity, temperature, and large-scale contextual information (wide). The combination of these different features allows the model to learn general representations of tropical disturbances and their evolution to cyclones.

The addition of channel-wise attention mechanisms is another reason why the model has achieved better performance. Such attention modules help the model to focus on the most relevant feature maps by giving it importance scores, in this way minimizing the influence of irrelevant or noisy data. Consequently, the model shows a better performance in capturing more informative and discriminative features, which is very beneficial for the generalization ability.

Besides, the grid resolution of 30×30 is also optimal and favorable with respect to the trade-off between spatial resolution and computational speed. It ensures that adequate environmental context is incorporated during training without introducing too much irrelevant information into the model. Moreover, the finding that the model successfully and consistently runs real-time cyclone case studies (e.g., Mindulle and Yagi) indicates strong quality and real-world applicability in an operational forecasting environment.

The proposed model processes a dense tensor but is nonetheless computationally feasible for early-warning applications. For the case of 30×30, as shown in Table 8, the total training time was 110.76 s, with an average training time per epoch of 0.87 s. The total inference time was 9.35 s, or 55.01 ms per sample. These results show that once trained, the model is able to make predictions in milliseconds, which meets near-real-time operational needs. Training is performed offline, while inference speed is fast enough to be used in early-warning systems.

Although the proposed STAG-Net model shows positive results, there are several limitations. For example, the model is based on a fixed-grid resolution and does not dynamically adapt to different spatial scales of various tropical disturbances. Further, the model is likely to perform poorly in areas or seasons where past cyclone activity is limited due to insufficient training data. Addressing such issues in future work would improve the robustness and extensibility of the model.

Furthermore, the current work mainly focuses on multi-level atmospheric dynamics to study vertically structured predictors of cyclogenesis. While considering model dimensionality and computational cost, single-level thermodynamic spatial fields such as ocean heat content were not fully integrated into the spatial branch, which may limit the representation of oceanic energy contribution to genesis.