1. Introduction

2. Materials and Methods

This study employs a simulated dataset generated from an UDWDM FSO–FTTx system to evaluate the performance under varying atmospheric conditions. The dataset consists of 4800 rows and nine columns, incorporating features such as modulation schemes (OOK–NRZ, QPSK, PDM–QPSK, and PolSK), input power levels, FSO link lengths, and atmospheric attenuation. Key performance indicators include OSNR, QF, BER, data rate, and received power. Data preprocessing involved normalization to mitigate the impact of outliers and ensure uniform scaling across features. Data normalization was applied, followed by an 80:20 split for training and testing sets with a fixed random state to ensure reproducibility. Python was the primary software for system model data generation and ML implementation. ML models applied in this analysis include an ELM with log-sigmoid activation, a SVM with a polynomial kernel, and GB algorithms. These models were trained on simulated datasets to predict performance metrics, including weather-induced attenuation levels, link distances, modulation schemes, and input power variations. The evaluation focused on diverse atmospheric conditions such as foggy, rainy, hazy, and clear weather scenarios to assess the system’s robustness under real-world environmental influences.

2.2. Dataset Preprocessing and ML Models

Data preparation and collection for this study are accomplished by obtaining the simulated dataset using essential performance parameters such as OSNR, QF, BER, data rate, and received power of the signal under various atmospheric conditions, which were measured by the proposed UDWDM FSO–FTTx system for various input parameter interactions, which included modulation formats, atmospheric attenuation levels, input power values, and the length of the FSO link. Then, data are normalized in a way that eliminates the outliers. The final data shape is formed of 4800 rows and nine columns, where the features are modulation schemes (OOK–NRZ, QPSK, PDM–QPSK, and PolSK), input power levels, FSO link lengths, and atmospheric attenuation, and where the labels are OSNR, QF, BER, data rate, and received power. Data are then broken down into a 80:20 split ratio between a train set and a test set to ensure robust model evaluation, and a fixed random state is applied to maintain reproducibility [34]. As shown in Table 3, the dataset description provides a clear overview of the dataset’s features and structure. In addition, Table 4 presents the components and variables used in our model, along with a clear description of their respective meanings.

Table 3. Dataset description.

Description		Values
Data source		Simulated dataset by utilizing the proposed UDWDM FSO–FTTx system
No. of instances		4800
No. of variables	Numerical	8
	Categorical	1

Table 4. Model components/variable’s description.

Model components/variable’s full name		Data type	Variables
Modulation	OOK–NRZ	Categorical	Input feature
	QPSK
	PDM–QPSK
	PolSK
Atmospheric attenuation (dB/km)		Numerical	Input feature
FSO length (km)		Numerical	Input feature
Input power (dBm)		Numerical	Input feature
Quality factor (dB)		Numerical	Output label
Bit error rate		Numerical	Output label
Optical signal-to-noise ratio (dB)		Numerical	Output label
Received power (dBm)		Numerical	Output label
Data rate (Gbps)		Numerical	Output label

The process begins with generating datasets by simulating performance metrics such as OSNR, QF, BER, data rate, and received power under varying atmospheric conditions (foggy, rainy, hazy, and clear weather). Normalization was performed on modulation schemes such as OOK–NRZ, QPSK, PDM–QPSK, PolSK, input power levels, lengths of the FSO link, atmospheric attenuation, OSNR, QF, BER, data rate, and finally received power to standardize these data to ML models. Methods for ML, including SVMs, ELMs, and GB, are then employed to predict target metrics. Figures 1 and 2 show a structured flow diagram that outlines the steps from dataset generation to preprocessing, model training, and evaluation for accurate prediction of optical transmission parameters under diverse environmental conditions. The task involves predicting multiple target variables $$ based on input features $$, where n is the number of samples, p is the number of features, and m is the number of target variables.

2.2.1. Data Preprocessing

Preprocessing prepares clean and relevant data for training ML models. This study used the imputation technique to examine the dataset for inconsistencies and missing data. To ensure compatibility with ML algorithms, categorical features were encoded using one-hot encoding.

Feature selection was made to retain only the most essential variables for input features and target metrics. Redundant or irrelevant features were removed to enhance the model’s performance. The data were then normalized, and the scaling was performed uniformly to be suitable for ML algorithms.

The mathematical expressions for normalization inputs and targets for both original features or label values normalized feature or label values in the dataset are calculated by using both original features or label values, respectively, and it is given by [35]

$$ ()

$$ ()

where X and Y are the original features or label values, X_norm and Y_norm are the normalized feature or label values, X_min and Y_min are the dataset’s minimum values for a feature or label, X_max and Y_max are the dataset’s maximum values for a feature or label, and max_n and min_n are the desired ranges of scaling

2.2.2. ML Models Building

The study developed ELM, SVM, and GBM ML models to predict target metrics such as OSNR, QF, BER, data rate, and received power in FSO communication systems. Input features included modulation schemes (OOK–NRZ, QPSK, PDM–QPSK, and PolSK), atmospheric conditions (foggy, rainy, hazy, and clear), input power levels, and FSO link lengths. Researchers split the dataset into training and testing subsets while ensuring balanced feature representation. The model performance was optimized using MSE and R². The best-performing model achieved minimal MSE and maximum R² through iterative refinement. The selection of ELM, SVM, and GB was motivated by their computation speed, interpretability, and stability for a medium-scale dataset. ELM’s randomized hidden-layer composition enables high-speed training and is suitable for real-time application without demanding significant computational resources [36]. SVM’s kernel-based methodology ensures resilience to outliers while maintaining transparent decision boundaries [36], a critical advantage over less interpretable neural networks for smaller datasets. GB minimizes overfitting risks through iterative error correction in its ensemble framework, achieving near-perfect parameter generalization [37]. While neural networks excel with large datasets, their complexity and resource requirements made ELM, SVM, and GB more practical choices for balancing accuracy and efficiency in our study scope.

Dataset preprocessing: This study generated the dataset by simulating UDWDM FSO–FTTx systems under various weather conditions, modulation schemes, and input power levels. Key performance metrics such as OSNR, QF, BER, data rate, and received power were collected. The dataset was preprocessed and prepared for the normalization of input features and target metrics to ensure compatibility with ML models. Biases were reduced, and better accuracy in the models was achieved through this step. Feature selection: Relevant features such as modulation schemes (OOK–NRZ, QPSK, PDM–QPSK, and PolSK), weather conditions (foggy, rainy, hazy, and clear), input power levels, and FSO link lengths were the selected impacts of these features on the system performance metric used as the basis for their selection.

Training set preparation: The preprocessed dataset was divided into subgroups for training and testing to make model creation and assessment more manageable. A typical split ratio of 80:20 was used to ensure adequate data for training while keeping some for validation. ML models: Three ML techniques, ELM, SVM, and GBM, were employed to predict system performance metrics. Each model was trained using the selected features as inputs to predict outputs such as QF, BER, OSNR, data rate, and received power. Predicted outputs: The trained models predicted the target metrics under varying conditions. These predictions were compared against actual simulated values to assess model accuracy. Performance evaluation: The model performance was evaluated using MSE and R². The model with the highest R² scores and the lowest MSE values across all projected outputs was determined to be the top performer.

2.2.2.1. Model for ELM ML

This study uses an ELM, a single-hidden-layer feedforward neural network (SLFN). Hence, the detailed mathematical model for the proposed system is given in the following [38, 39]: the first step is to generate the input weight and biases randomly, and the input weight is calculated as

$$ ()

where N_h is the number of hidden neurons and N_i is the number of input features. Mathematically,

$$ ()

In addition, biases are calculated as

$$ ()

Moreover, the second step is hidden-layer outputs’ (H) calculation, and for this study log-sigmoid activation function is used, for each training data X_k as

$$ ()

where f(z) represents the log-sigmoid activation function, given by f(z) = 1/(1 + e^−z). Finally, the study used the matrix form for all training samples that is expressed by

$$ ()

where $$ is the input data matrix.

2.2.2.2. Model for SVM ML

The second ML model used by the study is a SVM along with a kernel with a polynomial that predicts the target variable using the following mathematical formulations [2, 39, 40]: so, the regression function of the expected value for input is given by

$$ ()

where n is the number of support vectors, x_i are the support vectors from the training data, $$ are the Lagrange multipliers determined during the training, and K(x, x_i) is the kernel function that is polynomial. So, the kernel function polynomial is given by

$$ ()

where d is the degree of the polynomial, r is the independent term in the kernel function, γ:1/n_features, and b is the biases term. Since our training separates the SVM model for each output dimension, so for the output dimension j, the prediction of each output dimension and the final prediction for each output of the study are expressed as

$$ ()

2.2.2.3. Model for GBM ML

Lastly, the ML model GB is used, which produces a progressive, step-by-step additive model, optimizing a loss function by combining weak learners. The mathematical formulation for our study is given by the authors in [41, 42]. Thus, for each target variable y_i, we open with an initial prediction, which is frequently the target values’ mean that is given by F₀(x) = c, where c = E[y] is the expected value of y_i. At each iteration n = 1, 2, …, N, we fit a new decision tree to approximate the negative gradient of the loss function with respect to the current predictions: let the true target values be $$. The GB algorithm minimizes a differentiable loss function that is given by L(y_i, F_n−1(x)), where F(x) represents the model’s prediction. Thus, the model starts with an initial guess

$$ ()

At each iteration n = 1, 2, …, N, a new weak learner is added to correct residual errors and is expressed as

$$ ()

where L(y_i, F_n−1(x)) is the loss function and F_n−1(x) is the current model prediction. A new tree is fitted to these residuals, and its output is calculated as follows: h_n(x) = f(x; w_n), where parameter w_n minimizes the residual error. Then, the updated model predictions by adding scaled outputs of the new tree become

$$ ()

where η > 0 is a learning rate. The process repeats until convergence or until reaching maximum iterations, and after training all iterations, the final prediction is given by

$$ ()

This process is also repeated independently for each output dimension for the data points ($$), and hence, our final boosted model prediction is expressed as

$$ ()

Finally, this study calculated two performance metrics for all models [4, 38, 39] and for each target variable (i = 1, 2, …, n), both MSE and R-squared (R²) scores are used and given as

$$ ()

$$ ()

where $$ is the residual sum of squares, $$ is the total sum of squares, y_ij is the true value of sample j for target variable I, and y_ij,pred is the predicted value by the models.

3. Results and Analysis

The system’s performance has been comprehensively evaluated by analyzing key parameters such as QF (dB), BER, OSNR (dB), Data rate, and received power (dBm) in relation to FSO range and input power under varying atmospheric conditions, including foggy, rainy, hazy, and clear weather from Figures 1, 2, 3, and 4. These metrics were recorded for different FSO link lengths ranging from 50 to 12,500 m while maintaining consistent system parameters. The study highlights the influence of increasing FSO link length on system performance across various modulation formats under diverse atmospheric conditions and turbulence levels. An increase in FSO link length resulted in the degradation of critical metrics such as OSNR, QF, BER, and received power across all examined modulation formats under the specified atmospheric conditions.

Figures 5(a), 5(b), 5(c), and 5(d) highlighted the performance of different modulation schemes in an UDWDM FSO–FTTx system under clear, hazy, foggy, and rainy weather conditions. For clear weather, PDM–QPSK achieved the highest QF of 30.617 dB at a link length of 12250 m with a minimal BER of 1.728765e − 52 and an OSNR of 57.77 dB, demonstrating superior performance compared to other schemes. In contrast, OOK–NRZ, while simpler, showed a lower QF of 24.617 dB at the given link length but suffered from a higher BER of 7.22e − 05. Under hazy conditions, all modulation schemes experience degradation; however, PDM–QPSK remains robust with a QF of 30.56 dB at a link length of 12,050 and an OSNR of ∼57 dB, while OOK–NRZ exhibited significant performance loss with a QF reduced to 24.56 dB over the same distance. The QF is highest for PDM–QPSK, reaching 29.85 dB in foggy conditions and 30.06 dB in rainy conditions, demonstrating superior signal quality compared to other schemes. The BER values indicated that PDM–QPSK achieved the lowest error rates, 2.47e − 44 in foggy and 1.84e − 46 in rainy conditions, due to its robustness against atmospheric impairments. Regarding OSNR, all modulation schemes maintained consistent levels around 48–50 dB, with slight variations across weather types. The received power remained stable at approximately −10 dBm for foggy conditions and improved to around −8 dBm during rain, reflecting better link performance in less dense atmospheric scattering scenarios. The results underscore that advanced modulation formats such as PDM–QPSK have resisted more unfavorable weather scenarios than simpler formats such as OOK–NRZ.

Figures 6(a), 6(b), and 6(c) underscored the effectiveness of different modulation schemes OOK–NRZ, QPSK, PDM–QPSK, and PolSK under clear and hazy weather conditions in terms of data rate, OSNR, log(BER), and QF. PDM–QPSK consistently achieved the highest data rate of 5120 Gbps with superior QF values exceeding 30 dB in clear (30.624 dB) and hazy (30.56 dB) conditions, demonstrating its robustness against atmospheric impairments. PolSK offered a balanced performance with a data rate of 2560 Gbps and high QFs above 28 dB (clear: 28.62 dB; hazy: 28.56 dB), making it suitable for moderate link lengths.

In contrast, OOK–NRZ exhibited lower data rates (1280 Gbps) and reduced QFs by around 24 dB (clear: 24.617151 dB; hazy: 24.558711 dB), reflecting its limitations in adverse weather scenarios despite simpler implementation. PDM–QPSK consistently outperformed other schemes with the highest data rates of 5120 Gbps, superior QFs (29.85 dB in foggy conditions and 30.06 dB in rainy conditions), and minimal BER values (2.47e − 44 in foggy conditions). In contrast, OOK–NRZ, while simpler, showed lower data rates (1280 Gbps) and higher BER values (2.51e − 04 in foggy conditions). The results also revealed that adverse weather impacted all metrics; however, rain has a slightly less detrimental effect compared to fog across all modulation schemes. Notably, OSNR remains stable around 48–50 dB, indicating robust signal quality despite environmental challenges. Across all schemes, BER values remained significantly low under clear conditions but increased slightly under haze, foggy, and rainy conditions due to signal attenuation effects on the FSO link. Notably, the data rate for PDM–QPSK is double to that of QPSK or PolSK (5120 vs. 2560 Gbps) due to polarization multiplexing, while OOK–NRZ exhibited the lowest data rate (1280 Gbps) but maintained acceptable performance metrics at the given link lengths.

Figures 7(a), 7(b), and 7(c) highlighted the effectiveness of OOK–NRZ, QPSK, PDM–QPSK, and PolSK modulation schemes under varying conditions of the weather about input power. In the situation of clear weather, QPSK at 5 dBm input power achieved a QF of 26.617 dB with an OSNR of 57.77 dB, while PDM–QPSK reached a higher QF of 30.617 dB and an extremely low BER (1.73e − 52) at the same input power range. PolSK demonstrated robust performance with a QF of 28.62 dB and comparable OSNR (57.77 dB) at 5 dBm input power over a 12000m link. However, OOK–NRZ showed lower efficiency with a QF of 24.62 dB and a higher BER (7.22e − 05) despite maintaining similar OSNR levels. Under hazy conditions, all modulation schemes experience slight degradation; QPSK at 5dBm input power yielded a reduced QF of 26.56 dB, while PDM–QPSK still outperformed others with a QF of 30.56 dB and minimal BER (8.22e − 52). For the input power of 5 dBm, PDM–QPSK consistently outperformed other schemes with the highest QF values of 29.85 dB in foggy weather and 30.06 dB in rainy weather, alongside negligible BER values (2.47e − 44 and 1.85e − 46, respectively). PolSK also demonstrated robust performance with QFs exceeding 27 dB in both conditions. OOK–NRZ exhibited the lowest QF (23.85 dB) and higher BER (2.51e − 04) under foggy conditions, indicating its susceptibility to adverse weather effects. The OSNR was similar for all schemes in the 48–50 dB range, which evidenced that UDWDM FSO–FTTx system’s adaptability to noise disruption is unaffected by diverse weather conditions. These results underlined the superior resilience of advanced modulation formats such as PDM–QPSK and PolSK in adverse weather conditions in contrast to conventional OOK–NRZ.

Figures 8(a), 8(b), and 8(c) depicted that the performance comparison of BER, QF, and OSNR against channel spacing for the UDWDM FSO–FTTx system under clear, hazy, foggy, and rainy conditions exhibited significant variations in the modulation techniques. When it comes to the channel spacing of 0.2 nm, PDM–QPSK consistently fared better than alternative methods for the top QF, 30.62 dB in clear weather and 30.56 dB in haze, along with the lowest BER, and 1.73e − 52 in clear weather and 8.22e − 52 in haze, making it the most suitable for high-capacity links at a data rate of 5120 Gbps. PolSK demonstrated strong performance with a QF of 28.62 dB (clear) and 28.56 dB (hazy) while keeping low BER values at 2.98e − 27 and 6.57e − 27, respectively.

In comparison, the OOK–NRZ has depicted lesser QF values of 24.62 dB in clear weather and 24.56 dB in haze for higher BER values of 7.222e − 05 and 8.00e − 05, respectively; consequently, it may only be appropriate for modest data rates, such as 1280 Gbps. At a narrow 0.2 nm channel spacing, the PDM–QPSK appears to perform better than others do for the highest QF in both foggy (29.85 dB) and rainy (30.06 dB) conditions at the lowest BERs in such conditions, 2.47e − 44 and 1.85e − 46, respectively. PolSK also came out very well with a QF of 27.85 dB for foggy and 28.06 dB for rainy conditions; in contrast, the lowest QF was brought about by OOK–NRZ: 23.85 dB for foggy and 24.06 dB for rainy weather yet continued to be viable for lower data rates such as 1280 Gbps. The results showed that UDWDM FSO–FTTx system levels were combined with advanced modulation formats, such as PDM–QPSK or PolSK, by utilizing the given channel spacing to maintain optimal performance under different atmospheric conditions.

The study’s results given above enable the accurate multiparameter performance predictions and system improvements required to predict the performance parameters of systems subject to diverse weather conditions through the use of ML techniques. This guarantees increased channel awareness and better QoS in a way that successfully alleviates difficulties brought on by unfavorable weather conditions. Thus, building on that basis, the subsequent examination of the use of ML to enhance the predictive model performance for the estimation of different parameters makes use of simulated datasets that consider the system performance as a whole. The study provides evidence of ML’s substantial contribution to the operational efficiency and reliable advancement of UDWDM FSO–FTTx systems. The regression fit plots in Figure 9(a) demonstrated that ELM had achieved the closest alignment between predicted and true QF values, with a near-perfect linear relationship. The result showed the excellent capability of this model in effectively simulating the QF in different scenarios.

By contrast, GBM in Figure 9(c) is also doing well but slightly deviating from the true values. On the other hand, it is found from Figure 9(b) that the SVM presented a more significant prediction error, showing that their precision is relatively low. Figure 10 further corroborated these findings by illustrating that ELM and GBM predictions closely matched the actual QF values across all data points, outperforming the other model. Figures 11(a), 11(b), 11(c), and 12 revealed that GBM consistently outperformed other models in predicting BER, achieving minimal error and a strong regression fit. The near-linear relationship of the predicted and actual BER values for GBM underlined its robustness in handling this parameter. ELM was the second-best performer, while SVM in Figure 11(b) displayed higher deviations from actual BER values, reflecting reduced accuracy.

For OSNR predictions, Figures 13 and 14 highlighted GBM’s dominance with precise alignment of predicted versus true values. The regression fit plot for GBM in Figure 13(c) showed minimal residuals, confirming its high accuracy in modeling OSNR. ELM again demonstrated in Figure 13(a) competitive performance but lagged slightly behind GBM. Figure 13(b) depicts SVM struggling to achieve comparable precision, reflecting noticeable prediction errors. Furthermore, supporting evidence is given in Figure 14, indicating that the GBM and ELM predictions closely fitted the true OSNR(dB) values across all data points, while the case was not so with SVM.

Figures 15 and 16 emphasized GBM’s and ELM’s exceptional predictive capability for data rate, achieving near-perfect regression fits with minimal error margins. Figure 15(c) of the result confirmed that GBM is highly effective at capturing complex relationships influencing data rates in FSO–FTTx systems; Figure 15(a)’s result of ELM provided reliable predictions but was less precise than GBM. Figure 15(b) revealed that SVM was less consistent, with more significant deviations from actual data rates. For received power (dBm) prediction, Figures 17(c) and 18 underscored that GBM demonstrated superior performance in predicting received power levels with high accuracy. Its regression fit plot showed a strong correlation between predicted and true values with minimal variance. While Figure 17(a) showed that the result of ELM performed reasonably well, it did not match the precision of GBM. Figures 17(b) and 18 revealed that SVM exhibited significant prediction errors compared to ELM and GBM.

Thus, Figures 9, 10, 11, 12, 13, 14, 15, 16, 17, and 18 showed that across all parameters, QF, BER, OSNR, data rate, and received power, GBM consistently outperformed other ML models by achieving the lowest MSE and highest R-squared values. These results underscored the potential of GBM for predicting performance metrics and optimizing FSO–FTTx systems in various environmental settings because of their capacity to deliver accurate predictions across critical performance metrics. Moreover, the comparative analysis of ML models ELM, SVM, and GBM was conducted using key performance metrics such as MSE and R-squared values across multiple predicted parameters, including QF, BER, OSNR, data rate, and received power. Figures 19 and 20, and Table 5 further supported the findings from Figures 9, 10, 11, 12, 13, 14, 15, 16, 17, and 18 by showing that GBM and ELM have consistently outperformed SVM in terms of accuracy and error minimization, as reflected by their superior R² values nearing one and significantly lower MSE values.

Figures 19 and 20 and Table 5 present the QF prediction analysis of the ML models revealing that ELM had achieved an outstanding MSE of 1.6785e − 04 with an R² value of 0.999965, whereas GBM had come close with a value of 9.3905e − 04 and an R² value of 0.999803. In contrast, the SVM model had shown an increased MSE of 3.2810e − 01 and a lower R² value of 0.9311, indicating less accuracy in prediction for this parameter. In the case of BER, GBM demonstrated the best performance with a very impressive MSE of 3.521367e − 12 and a high R² value of 0.998672, outperforming ELM, which had recorded an MSE value of 7.6909e − 12 and an R² value of 0.997099. SVM had lagged significantly behind with an MSE of 4.413622e − 10 and an R² value of only 0.833542. For OSNR, both ELM and GBM had shown strong predictive capabilities with respective MSE values of 7.9385e − 03 and 2.020315e − 03, alongside high R² values of 0.998946 and 0.999732, respectively; however, SVM again had underperformed with a considerably higher MSE value of 5.526608e − 01 and a lower R² score of 0.926604.

When analyzing the data rate, GBM achieved near-perfect results with an MSE as low as 1.252711e − 03 and an ideal R² value of precisely 1, showcasing its unparalleled precision for this parameter compared to ELM’s already excellent performance with an MSE of 2.2248e + 01 and an R² score close to perfection at 0.999987. SVM’s performance was also notably weaker here, reflected by its exceedingly high MSE value of 1.474559e + 05 and a relatively low R² score of just 0.916419.

Lastly, for received power, both ELM and GBM maintained their dominance with identical performances to those observed for QF, BER, OSNR, and data rate: ELM recorded an MSE value of 7.9385e − 03 paired with an R² score of 0.998946, while GBM achieved a superior MSE value at just 2.020315e − 03 along with a near-perfect R² score at 0.999732; meanwhile, SVM trailed behind once more with an elevated MSE at approximately 5.526608e − 01 accompanied by a lower predictive accuracy indicated by its corresponding R² score at only about 0.926604. In conclusion, the results from Figures 19 and 20 and Table 5 conclusively established that both GBM and ELM models are the most effective models among those tested for predicting various optical communication parameters such as QF (dB), BER, OSNR (dB), data (Gbps), and power (dBm). The fact that it yielded near-perfect R² values and minimum MSE highlighted its suitability in highly accurate and reliable applications. It was far from the high error rates emanating from SVM and lessened the predictive power it had.

4. Discussion

In this study, cutting-edge ML algorithms were applied to estimate the performance of UDWDM FSO–FTTx systems under various weather conditions. A comparative analysis with prior studies, as summarized in Table 6, highlighted significant advancements and contributions of our research regarding system modeling, atmospheric conditions, ML algorithms employed, modeling targets, and modulation schemes. In comparison to previous works such as [4], which utilized RoFSO systems under clear, haze, and rain weather conditions with ML algorithms lsuch as ANN, KNN, and DT targeting SNR and BER metrics using 16-QAM and 32-QAM modulation schemes, the present study extends its scope by incorporating a UDWDM FSO–FTTx model. This model can operate effectively under various weather conditions, including foggy, rainy, hazy, and clear sky environments. Similarly, the study in [6] focused on MDM/FSO systems in rainy weather using KNN for BER prediction. While effective for specific scenarios, it lacks the versatility demonstrated in the current work’s ability to handle multiple adverse weather conditions. This versatility ensures higher reliability in real-world scenarios. Similarly, the authors in [7] employs RoFSO systems under strong turbulence conditions with ANN, GBM, and DT models targeting BER improvements through BPSK modulation. While effective for specific turbulence scenarios, it lacks the broader adaptability of the proposed system. The present work not only addresses BER but also optimizes additional multiple parameters such as OSNR, QF, data rate, and received power in various weather conditions by taking into account a variety of modulation schemes, including OOK–NRZ, QPSK, PDM–QPSK, and PolSK modulation schemes. This multitarget approach ensures comprehensive performance evaluation across diverse operational scenarios.

Furthermore, the authors in [8] investigated hybrid FSO/RF links under rain and fog using RF ML models for link availability, spectral efficiency, and BER predictions. While innovative in addressing hybrid link challenges, it does not leverage ultra-dense WDM or address FTTx applications. The present work surpasses this by integrating UDWDM technology into FSO–FTTx systems to enhance spectral efficiency while maintaining robust performance under challenging atmospheric conditions. Lastly, the authors in [10] addressed IsOWC systems with scintillation and turbulence effects using ANN but limited its focus to received power predictions. The current study builds upon this by addressing multiple parameters, such as OSNR, QF, BER, and data rate, in various weather conditions by taking into account a variety of modulation schemes, including OOK–NRZ, QPSK, PDM–QPSK, and PolSK modulation schemes, while ensuring compatibility with ultra-dense WDM configurations.

The comparison shown in Table 7, a comparative evaluation of the achievement in current studies, is a direct consequence of observations framed in Table 6, which compares the proposed work with other studies. This table defines the new contributions, improved methodologies of the current study, and notable improvements over existing research.

In conclusion, the proposed UDWDM FSO–FTTx framework offers significant advancements over existing FSO communication systems, and Table 7 provides a detailed comparison of the proposed framework with other studies, which shows that the proposed framework significantly advances existing FSO communication systems by addressing critical scalability, adaptability, and performance analysis limitations. This work introduces a holistic approach, unlike prior studies that focused on narrow metrics such as BER or SNR [4, 6] or limited their evaluations to specific weather conditions such as rain or turbulence [8, 10]. The system achieves unprecedented spectral efficiency and scalability by integrating ultra-dense WDM with FSO–FTTx, a key gap in hybrid FSO/RF and MDM/FSO architectures [4, 7]. Moreover, rigorous testing across different atmospheric conditions, including fog, haze, rain, and clear skies, validates robustness in scenarios where conventional models falter, particularly under foggy environments with severely lower received power [8]. The adoption of ensemble and kernel-based ML models (ELM, SVM, and GBM) outperforms traditional ANN and DT methods [4, 7] in predicting OSNR and QF, enabling real-time adaptive modulation. In addition, the addition of PDM–QPSK and PolSK schemes improves noise resilience and data rates compared to standard QAM/BPSK modulations [6, 10]. This multidimensional optimization of OSNR, BER, and throughput addresses a critical trade-off overlooked in single-metric studies, making the framework a versatile solution for next-generation optical wireless networks.

5. Conclusion

This study comprehensively evaluates UDWDM FSO–FTTx systems under various weather conditions, and it established a rigorous framework for positioning UDWDM FSO–FTTx systems in next-generation networks, considering critical challenges posed by atmospheric turbulence and weather-induced attenuation. Systematic evaluations across clear, hazy, rainy, and foggy conditions demonstrated that polarization-based modulation schemes (PDM–QPSK and PolSK) consistently outperformed conventional formats, achieving superior performance metrics: QF (> 27 dB), OSNR (> 48 dB), and data rates (≥ 2560 Gbps) while maintaining BER below the (10e − 12) threshold. In contrast, OOK–NRZ demonstrated severe degradation under adverse weather conditions, predominantly in fog, where BER > (10e − 5), emphasizing the necessity of advanced modulation for reliable FSO communication.

In addition to these findings, ML algorithms, particularly GB, significantly enhanced predictive accuracy across performance metrics, achieving minimal MSE (< 0.002) and near-perfect R² values (> 0.998). While ELMs approached GB’s accuracy, SVM lagged in overall performance. The results showed that using ML algorithms greatly enhanced the prediction capability for UDWDM FSO–FTTx systems under different weather conditions while promoting channel awareness and enhanced QoS. These ML–driven insights enable real-time channel dynamics forecasting, which could lay the groundwork for adaptive strategies, such as SVM/ELM–guided microelectromechanical systems (MEMS) beam steering to counteract beam wander while preserving ultra-dense WDM channel spacing. Furthermore, GB’s ultra-low BER predictions could facilitate autonomous hybrid RF/FSO switching to maintain QoS during extreme weather events.

In light of these developments, the proposed system aligns with 6G efficiency targets through ultra-dense spectral alignment and ML–driven adaptability, offering a scalable optical backbone for emerging applications such as quantum key distribution (QKD)–compliant OSNR levels for quantum-secure transmission. The generated multiparameter datasets provide foundational frameworks for training deep reinforcement learning (DRL) agents to autonomously control polarization switching, modulation reconfiguration, and power optimization in response to real-time weather fluctuations. Future work should focus on experimental validation of ML–guided adaptive systems in turbulent urban settings and expansion into terahertz capacity regimes. This study positions ML–enhanced WDM FSO–FTTx as a transformative, intelligently adaptive solution for next-generation networks capable of sustaining robust performance under dynamic atmospheric conditions.

Conflicts of Interest

The authors declare no conflicts of interest.

Funding

There is no any research fund in this study.