جستجوی مقالات مرتبط با کلیدواژه "یادگیری ماشین" در نشریات گروه "علوم انسانی"

This study explores the phenomenon of deepfakes as a consequence of rapid advancements in artificial intelligence, machine learning, and deep learning technologies over the past decade. The primary objective is to analyze various methods for detecting deepfakes and examine their social and legal implications.

The research categorizes and evaluates four types of deepfake detection

deep learning-based, classical machine learning-based, statistical, and blockchain-based approaches. It also assesses the performance of these methods on different datasets.

The findings indicate that deep learning-based methods are more effective in detecting deepfakes compared to other approaches. Furthermore, the study analyzes the impact of deepfakes from multiple perspectives, including media and society, media production, representation, dissemination, audience, gender, law, and politics. The results reveal that society is currently unprepared to effectively combat deepfakes, due to a combination of technological, educational, and regulatory shortcomings.

Originality/Value: 

This research provides a comprehensive and comparative analysis of deepfake detection methods, offering valuable insights for policymakers and researchers. The study highlights the urgent need for effective strategies to address the rapidly evolving challenges posed by deepfakes in both social and legal contexts.

The urban space is the most important human-made spatial structure on the planet earth. The history of urban development shows the path of human development, political system evolution and technological, technical and industrial developments. The physical development of urban areas is one of the main drivers of global changes that have important direct and indirect effects on environmental conditions and biodiversity. In the process of physical development of the city, due to the transformation of natural and semi-natural ecosystems into impermeable surfaces, it often causes irreversible environmental changes. One of the new approaches in urban planning is the use of remote sensing techniques and geographic information system. The emergence of remote sensing and machine learning techniques offers a new and promising opportunity for accurate and efficient monitoring and analysis of urban issues in order to achieve sustainable development. The process of processing satellite images can generally be divided into two approaches: pixel-based image analysis and object-based image analysis. The pixel-based analysis technique is performed at the level of each pixel of the image and uses only the spectral information available in each pixel. On the other hand, the object-based analysis approach is performed on a homogeneous group of pixels, taking into account the spatial characteristics of the pixels. One of the basic problems in urban remote sensing is the heterogeneity of the urban physical environment. The urban environment usually includes built structures such as buildings and urban transportation networks, several different types of vegetation such as agricultural areas, gardens, as well as barren areas and water bodies. Therefore, in the pixel-based processing approach, the existence of heterogeneity in the urban biophysical environment causes spectral mixing and also spectral similarities in the classification operation of satellite images in such a way that in a place where a pixel is If the surrounding environment is different, it causes Salt and Pepper Noise. Therefore, according to the problems in the pixel-based processing approach, the aim of this research is to compare the accuracy of machine learning algorithms based on object-based processing of satellite images in extracting the physical development area of Hamedan city using Sentinel 2 satellite image.

  The remote sensing data used in this research is a multi-spectral satellite image with a spatial resolution of 10 meters from the Sentinel 2 satellite, including bands 2 (blue), 3 (green), 4 (red) and 8 (near infrared) related to the date is the 23 of August 2023 in the city of Hamadan. The image of the Sentinel 2 satellite was downloaded from the website of the European Space Agency. In ENVI software, the pre-processing operation was performed on the satellite image. Then, in the eCognition software, the segmentation process was performed based on the appropriate scale, shape factor, and compression factor with the aim of producing image objects. After segmenting and converting the image into image objects, using machine learning classifiers based on object-oriented processing of satellite images including Bayes classification algorithms, k-nearest neighbor, support vector machine, decision tree and random trees, the classification process was carried out and maps of urban physical development area were produced. After the segmentation operation and the production of visual objects, three classes of built-up urban land, vegetation and barren land were defined, and some of the built objects in the segmentation stage were selected as training points and some were selected as ground Truth points.Results & DiscussionAfter downloading the satellite image from the website of the European Space Organization, in order to apply the radiometric correction of the image and also with the aim of matching the value of the gray levels of the image with the value of the real pixels of the terrestrial reflection, the gray levels are converted to radiance and then, using atmospheric correction, to coefficients. They became terrestrial reflections. In order to apply radiometric correction, Radiometric Calibration tool was used, and to apply atmospheric correction, FLAASH model was used in ENVI software. In order to classify the satellite image based on machine learning algorithms based on object-based processing, eCognition software was used. The satellite image of the study area, which was pre-processed and saved in TIFF format, was called in the environment of this software and saved as a project. In order to produce visual objects, segmentation operations were performed in different scales, shape factor and compression ratio to reach the most appropriate segmentation mode. In this step, the multiple resolution segmentation method was used to segment the image. The most appropriate segmentation included the scale of 100 and the shape factor of 0.6 and the compression factor of 0.4. Because in scales higher than 100, the construction of the visual object was not done correctly, so that several distinct complications were placed in one piece, and in scales less than 100, in some cases, one complication was placed in several pieces. In order to classify the generated image objects, machine learning algorithms were defined separately and after training each algorithm, the classification operation was performed. In this step, the classification was done based on the nearest neighbor method and by selecting the average and standard deviation parameters for each image band. After producing a map of the city physical development range through machine learning classifiers based on object-based processing of satellite images, the classification accuracy of each of the used algorithms was calculated. In order to calculate the accuracy of the above algorithms in eCognition software, using selected ground Truth control points, the overall accuracy and kappa coefficient were calculated for each of the algorithms.

Based on the results of the research, it is possible to produce a map of Hamedan's urban physical development using machine learning algorithms based on object-based processing of satellite images with acceptable accuracy. Also, among all the algorithms used in this research, k-nearest neighbor with overall accuracy of 97% and kappa coefficient of 0.96 provided more accuracy.

According to the 22 selected features (which are checked during the research) with machine learning methods, this study predicts the cash holding of companies admitted to the Tehran Stock Exchange. 201 companies were investigated from 1396 to 1400. Multiple linear regression, K-nearest neighbor, support vector regression, decision tree, random forest, extreme gradient boosting algorithm and multilayer neural networks are used for prediction. The results show that the multiple linear regression methods provide the k-nearest neighbor of the root mean square error (RMSE) and the mean absolute error (MAE) of the high error. Meanwhile, more complex algorithms, especially support vector regression, achieve higher accuracy; The findings indicated that by reducing to 15 variables, machine learning methods, especially K-nearest neighbor, provided better results. Based on the paired sample t-test, support vector regression has a better performance than other supervised machine learning algorithms except decision tree. Also, the most important variables were company size and capital expenditures (CapEx). The World Uncertainty Index and inflation were also relatively important variables; Therefore, by using the support vector regression algorithm, we may predict the amount of cash to a significant extent.

Investment portfolio theory is an important foundation for portfolio management, which is a well-studied but not saturated topic in the academic community. Integrating return forecasting in investment portfolio formation can improve the performance of portfolio optimization model. Since machine learning models have shown a superiority over statistical models, in this research, a approach of forming the stock portfolio in two stages is presented. first step, by implementing neural network, suitable stocks are selected for purchase, in the second step, using the (MV) model, the optimal weight in investment portfolio is determined for them. In particular, the stages of selecting suitable stocks and forming a stock portfolio are the two main stages of the model developed in this research. first step, a convolutional neural network model is proposed to predict stock buy and sell points for the next period.second step, stocks that are labeled as buys are selected as stocks suitable for buying, and MV model is used to determine their optimal weight in the stock portfolio. The results obtained using 5 shares of Tehran stock market as a study sample show that the efficiency and Sharpe ratio of proposed method is significantly better than traditional methods (without filtering suitable stocks)

The intersection of artificial intelligence (AI), psychology and applied linguistics particularly in the realm of language learning, has opened up a fascinating avenue for exploring the intricate processes and mechanisms underlying human cognition. Machine learning algorithms have the potential to shed light on the fundamental principles of learning language specially on reading comprehension as the core language learning variable.

In this research, we used supervised machine learning techniques in order to discover the most important syntactic and lexical features affecting the reading comprehension of English language learners.

The design of present study is causal comparative type (ex post facto). the population includes all second secondary level students who learn English in language training institutions. To select the participants, language training institutes in Tehran were referred. The participants (n=360) answered BALA exam (Young, 2022) questions in written and spoken form.

260 features were extracted from the computer texts prepared from the speech and writing learners responses by natural language processing (NLP) algorithms. We used learning models of decision tree, nearest neighbor, support vector method, neural network and regularized linear method to predict reading comprehension using extracted linguistic features.

The results showed that the variance of language learners' reading comprehension can be well modeled using the extracted grammatical and lexical features, and in addition, twenty features that play the most important role in explaining the variance were identified. This study shows that ML methods can determine the detailed investigation of language processes related to reading comprehension.

As technology advances and the use of social networks and the Internet grows, there's a significant increase in the production of data. The concept of "big data" is becoming more prominent in various scientific studies. However, understanding how big data is used in tourism is a challenge addressed in this study. The research investigates the application of big data in tourism studies published from 2016 to 2022, aiming to identify research gaps and potential areas for future development. To provide an overview of published works in tourism, we used the systematic mapping study method. Articles were assessed based on various factors, including research topic, conceptual and theoretical descriptions, data sources, data types and sizes, data collection methods, data analysis techniques, and reporting and visualization methods. The findings indicate a growth in research within hospitality and tourism management using analytical techniques for big data. Despite this, the research in this field is scattered and limited in scope and methodology, revealing several gaps. Additionally, the study suggests that the full potential of big data remains untapped in tourism, offering numerous research opportunities for the industry. A notable challenge highlighted is the .researchers' incomplete knowledge of data-driven analysis and decision-making tools in this field

Forecasting the stock price due to its attractiveness has always been on the focus of experts and capital market activists. In such a way that various prediction models such as technical and fundamental analysis and data mining are increasingly used to predict stock prices. Past studies indicate the possibility of stock price prediction by machine learning models, but the prediction accuracy of these models has not been evaluated. The purpose of this research is to accurately compare the intelligence of five commonly used data mining algorithms, including neural network, Logestic regression, k-nearest neighbors, support vector machines and cross validation. Among the 385 active companies in Tehran Stock Exchange, 72 companies have been selected by the method of systematic elimination and the above models have been implemented to predict stock prices on the daily prices of selected stocks for the years 2009 to 2020. The stock price is used as a dependent variable and changes in the opening price, closing price, highest price, lowest price and volume of trade, daily price of forign currency, gold and oil price are used as independent variables. Three indicators R2, MSE and RMSE are used, to evaluate the accuracy of models, and analysis of variance using F statistics is used to fit the accuracy of the models, and t-student statistic is used to compare two models. The results are showed that among the smart algorithms used, the support vector machine algorithm has the highest accuracy. Matlab software is used to analyze the data and compare the models.

Portfolio optimization as a popular research field has received many attention from researchers in recent decades. Quality of portfolio optimization helps investors generate more sustainable returns. In this research, Convolutional Neural Network (CNN) is used to build a portfolio optimization model based on prediction. This model not only benefits from deep learning technology, but also benefits from modern portfolio theory. In this approach, CNN is first used to predict the future return of each stock. Then, the prediction error of CNN is used as the risk measure of each stock. Integrating the predicted return with the semi-absolute deviation of the prediction error leads to the construction of the portfolio optimization model. This model is compared with an equally weighted portfolio whose stocks are selected with CNN. Also, two prediction based portfolio models with support vector regression (SVR) are used as benchmark portfolios. The empirical data of this research includes the companies in the index of 50 most active companies of Tehran Stock Exchange. The experimental results show that the prediction-based portfolio model with CNN shows a superior performance compared to SVR in the conditions of different returns. Also, the increase in the expected return can improve the performance of this model. This research clearly states the positive role of deep neural networks (DNNs) in creating portfolio optimization models.

 The lack of certainty in the movement of the stock market has made forecasting a challenging task in the field of financial time series forecasting. On the other hand, it is not easy to analyze the time series data of stock prices due to non-linearity and high noise. Therefore, the aim of this research is to predict the long-term and short-term trend of the capital market. To achieve this goal, artificial intelligence algorithms of particles and genetics have been used in a comparative manner. The studied variable is the total stock price index in Tehran Stock Exchange in the period of 2016 to 2021 and on a monthly basis. The data have been reviewed after collection using the smoothing method for holidays, and in order to increase the accuracy of the models, the optimal window length of each algorithm has been calculated. The findings indicate that the genetic algorithm by minimizing the prediction error is a suitable algorithm for predicting the short-term and long-term trend of the total price index compared to the particle algorithm in the studied time period.

The current research has been conducted with the aim of comprehensively investigating the effects of artificial intelligence on the future of educational management and drawing future prospects in the horizon of 1410. This study has collected and analyzed the opinions of 50 experts in the field of educational management, artificial intelligence and future research by using the mixed (quantitative-qualitative) method and using the fuzzy Delphi technique. The findings of the research show that artificial intelligence will have deep and extensive effects on various dimensions of educational management. The most important influential areas include personalization of education, assessment and measurement of educational progress, human resource management in the educational system, and data-based decision-making at different management levels. The main challenges identified are: ethical and privacy issues, the need for extensive retraining of manpower, and structural changes in the education system. Key opportunities also include increasing the efficiency and effectiveness of educational processes, wider access to quality education, and the possibility of lifelong learning. By providing strategic solutions to prepare the educational management system in the face of future developments, this study can help policy makers and educational managers in long-term planning.

The current research was conducted to find the optimal portfolio for investing in stock exchange stocks, and one of the methods that is currently very popular among analysts and researchers in this field is methods based on artificial intelligence, followed by methods aimed at reducing risk metrics. The aim of the current research is to form a portfolio using machine learning methods, risk measurement and its combination with fuzzy theory, which has a better return than the average return of the market. The output of each method is entered into the random forest algorithm and prediction is made by this algorithm, and in the last step, the prediction output is entered into the value-at-risk and value-at-risk optimization model with the fuzzy theory approach to form the capital portfolio. Shares information is daily and its time period is from the beginning of 2014 to the middle of 2018. At the end of each of these methods and steps, it was compared with the real return of the market. the CVAR risk measure has a better ability than the VAR risk measure, and the random forest algorithm among the used machine learning algorithms has achieved better results in choosing the investment portfolio.

In this research, based on the historical statistics of the deposits of a private bank (Ansar bank before merging) into four different categories (funds), we are looking for the results and prediction of bank deposits in each of the categories, using machine learning prediction methods for the first time in internal studies. In this forecasting method, the superiority of the machine learning method is confirmed due to the examination of all the time changes of the data in a large volume (from the beginning until now). Forecasting the above deposit statistics for bank decision makers in the future according to the economic conditions can be the basis of macro banking decisions. The results show that partly stable long-term deposit and stable deposit will take a downward trend in the future in the Ansar bank, and unstable and interest-sensitive deposit will also take a downward trend. Qarzol-ha-sa-ne deposit has an upward trend, although it has always fluctuations similar to sin-cos fluctuations. It should be noted that this machine learning prediction method is a stable method without sensitivity analysis and has high reliability. According to this research, paying attention to long-term and stable deposits as well as Qarzol-ha-sa-ne is of great importance in the bank’s budget.

The purpose of this article is to predict the potential financial distress of the listed companies on Tehran Stock Exchange (TSE) and Iran Fara Bourse (IFB). To do so, a wide range of features including accrual accounting variables, cash-based accounting variables, market-based variables, corporate governance mechanisms, and macroeconomic indicators have been identified to prospectively predict the financial distress in the companies.The final sample includes 421 firms leading to 3,670 firm-year observations. The prepared data, was then split into a train and test data set using a 70/30 ratio.In this research, various data pre-processing machine learning techniques i.e., Z-score standardization, one-hot encoding, stratified K-fold validation combined with feature engineering are applied to improve classifier performance. Stratified K-fold cross validation method, (with k = 5) was used for estimation of model prediction performance during training phase. During the training phase, hyper-parameter tuning of a model was carried out using a grid-search. Furthermore, SMOTE technique in conjunction with the proposed imbalance-oriented metric i.e., F1 score were used to overcome the extreme class imbalance issue.Based on the experimental results, the tuned XGBoost model achieved a f1-score, MCC, recall and precision of respectively, 90%, 90%, 100% and 82% on the training set. Finally, the proposed model was tested on the hold-out test set which resulted in a f1-score, MCC, recall and precision of 52%, 52%, 73% and 41%, respectively. This information provides a powerful tool for predicting the financial distress of companies.

Identifying and mapping crops provides important information to agricultural lands management and cultivation area estimation of crops. Optical and radar images are valuable resources for classifiying agricultural land. Features deriverd from optical images contain information about the reflectance signatures of various products, while radar images provides information about the structural characteristics and distribution mechanisms of products. The combination of these two sources can create a complementary dataset with a significant number of spectral, texture and polarized temporal features for the classification of agricultural land

This study aims to explore the significance of red edge bands for the segregation of crops such aswheat, barley, alfalfa, beans, broad beans, flax, corn, sugar beet and potatoes using the random forest method and support vector machine. To conduct the analysis, a time series of Sentinel-1 and 2 images 2019 in the northwest region of Ardabil was retrieved from the Google Earth Engine (GEE) platform. The study evaluates the effectiveness of spectral and temporal information, plant indices and backscatter information on the crop mapping by examining different combinations of bands. Through the random forest feature selection method, essential features are identified and utilized as inputs for both the random forest and support vector machine classifiers.

The random forest provided the most favorable outcomes across all scenarios. The results revealed that incorporating red edge wavelengths and red edge-based vegetation indices proved more beneficial than other bands and vegetation indices for differentiating between barley, beans, broad beans, and flax. The most optimal outcome among various feature combinations was associated with the time series of spectral features from Sentinel-2 images combined with the time series of Sentinel-1 images, resulting in an overall accuracy of 84.67% and a kappa coefficient of 82.31%.  Furthermore, the results demonstrated that red edge bands and red edge-based vegetation indices effectively distinguish between different types of crops

It is recommended to carefully consider the selection of specific spectral bands to achieve higher accuracy in separation of crops. It is important to highlight that combining radar and optical images consistently yields superior results compared to classification methods based on a single sensor, leading to increased classification information.

The aim of the current research is to provide a comprehensive model for providing a model for operational budgeting in fossil power plants based on machine learning and ground theory methods. In this research, the impact of the participatory approach in performance-based budgeting on accountability in the public sector was investigated. Budgeting on the basis of performance provides a basis for more accountability against the use of the organization's resources. In addition, issues such as environmental protection, high initial investment cost and optimal use of resources have a direct impact on the performance of the players in this market. Unfortunately, in the new competitive conditions, the traditional and classic models that assume the economic factors (producers or consumers) to be homogeneous and without interaction do not have the necessary efficiency and do not provide acceptable results. Operational investigation explains the behavior of heterogeneous production units and allows interaction and learning in a dynamic environment, but due to the breadth of the model and its complexities, an analytical solution cannot be used to obtain variables in equilibrium conditions. Due to the advancement of computing technology, the results of agent-based models can usually be evaluated with simulation methods in the context of different scenarios. Production units present their proposed production schedule for every hour to the operating entity. According to the consumption demand forecast for the next twenty-four hours and by implementing the auction mechanism, this institution determines the winning units and the market price, and then by implementing the settlement mechanism, sales and settlements are done. Considering the necessity of balance of consumption and production in the whole country as well as the geographical distribution of production and consumer units, in addition to the amount of consumption, economic restrictions and technical restrictions related to production, transmission and distribution are also considered in the auction mechanism. The findings of the research show that the existing infrastructure of performance-based budgeting does not affect financial accountability in the public sector, while it has a negative effect on operational accountability. The establishment of performance-based budgeting infrastructure and the use of a participatory approach in budgeting will increase the level of accountability of the government in the public sector.

One of the key priorities of the Ministry of Agriculture Jihad is the mapping of croplands to estimate crop acreage and annual yield. In recent decades, remote sensing technology has proven to be highly effective in estimating the extent of crop cultivation through the use of timely images and synchronized data with diverse spatial, temporal, and spectral resolutions, leveraging advanced machine-learning algorithms. This study presented a framework for crop mapping in Marvdasht, Fars Province, by utilizing time series of Landsat-8 satellite images and advanced machine-learning algorithms. The employed algorithms included Decision Tree (DT), Random Forest (RF), Rotation Forest (RoF), Support Vector Machine (SVM), and Dynamic Time Warping (DTW) analysis. The results indicated that the dynamic time warping and random forest methods outperformed others, achieving significantly higher accuracy (with an overall accuracy improvement of 10-12%) in generating the agricultural land-use map of the study area. Furthermore, this research demonstrated the effectiveness of Bands 2-5 of Landsat-8 satellite in confidently identifying all crops in this region using the mentioned methods.

In this paper, we investigate how novel approaches can improve the predictions made by traditional econometric approaches in the field of forecasting. Autoregressive integrated moving average (ARIMA) is known as one of the most widely used methods for predicting economic and financial time series, providing a good framework, especially for short-term linear predictions of time series. However, the assumption of nonlinear effects in time series and the emergence of novel deep learning algorithms, which can extract complex features of time series and model them, have motivated researchers to examine the predictive power of traditional and novel modeling approaches. In this study, two methods are examined for predicting the prices of the four most valuable cryptocurrencies. ARIMA and three approaches in the field of deep learning, including (RNN, LSTM, and GRU), are investigated. In addition, a hybrid model of deep learning and ARIMA has been introduced, which is a combination of the strengths of both models to increase the accuracy of predictions. The results show that the hybrid models perform better in predicting future time series than each of the ARIMA and deep learning models separately. Also, the ARIMA-GRU model has fewer prediction error values than all estimated models.

The warming of the urban environment is one of the consequences of unsustainable growth. This research aims to investigate the possibility of modeling the effect of the structural parameters on the city’s surface temperature in the summer season in Tehran. For this purpose, the Landsat-8 image taken in 2018 was used to calculate the surface temperature. In order to determine the study units in this research, the segmentation method was used on the Sentinel-2 image of 2018, and the ratio of the vegetation cover and the separation of built-up areas from non-built-up ones were extracted using this image. The multi-layer perceptron neural network and the convolutional neural network methods were used to model the effect of urban structural parameters on the surface temperature during the summer. The results obtained from random forest feature selection  for the summer indicates that the presence of vegetation and urban uses that include residential and industrial areas, the presence of mixed residential/commercial/administrative areas, and the presence of vegetation affect changes in the urban surface temperature. Further, the information layers of road and population density in this season have an effect on the changing temperature of the earth's surface. Additionally, the results obtained through modeling and t-test of paired samples demonstrate the superiority of the convolutional neural network method, with a root mean square error of 0.61, determination coefficient of 0.62, and 17.75% estimation error, compared to the multi-layer perceptron model, which had 0.82 root mean square error, 0.26 determination coefficient, and 23.34% estimation error.

The digital world provides many opportunities for marketers to reach customers. However, in the fast-paced world, finding new and innovative ways to advertise and sell products and services is very important. Due to the advancement of artificial intelligence and its development in the field of advertising and sales, professionals now have the tools to completely redefine the current understanding of branding, marketing, advertising and sales. The growing popularity of the Internet and the increased use of mobile devices are generating massive amounts of consumer data that feed artificial intelligence-based systems. This research is a type of mixed research with a qualitative and quantitative approach, which is a survey descriptive study in terms of its purpose, application, and in terms of data collection. The statistical population of the research was managers and experts in the field of digital marketing and IT in the field of advertising and sales, who were selected using the snowball sampling method. In the qualitative part, the tools for collecting information were library and articles review, interviews, and in the quantitative part, questionnaires. In the qualitative part of the data analysis method, using the theme analysis that was compiled with MAXQDA software and using the coding method, and in the quantitative part, the analysis method was based on Kendall's correlation test. According to the results of the research, 7 main themes, 22 sub-themes and 44 codes were discovered, which included the consequences of using artificial intelligence and machine learning in advertising and sales

  As one of the most essential needs of living beings, clean air quality has been threatened by natural and human activities. In recent years, dust storms have been increasing spatially and temporally, causing numerous damages to social, economic, and environmental health for the residents of the southern and southwestern regions of Iran. In the present study, MODIS sensor data were used to investigate dust storms and detect horizontal optical depth.

The  advantages of MODIS sensor data include high spectral and temporal resolution. Additionally, meteorological station data were collected based on the study period. After preprocessing the data and preparing field observations, the necessary features for modeling were extracted using the differential method between selected bands of each MODIS sensor image, along with features extracted from ground-based meteorological station sensors. After further investigations and evaluations and using the viewpoints of meteorological experts, 36 differential features from various MODIS image bands and six features from ground-based meteorological station data, totaling 42 features, were extracted. Subsequently, using feature selection techniques, the best features were identified. A novel method named ML-Based GMDH, which improves the GMDH neural network by altering partial functions with machine learning models, was employed to detect dust concentration and horizontal optical depth. To achieve optimal accuracy, the hyper-parameters of this model were heuristically tuned using the TLBO optimization algorithm. Additionally, machine learning methods such as Basic GMDH, SVM, MLP, MLR, RF, and their ensemble models were implemented to compare with the main approach. According to the results, the TLBO-tuned ML-Based GMDH method provided superior accuracy in detecting dust concentration compared to the aforementioned machine-learning methods.

The SVM-PSO method was selected as the best method in the feature selection phase, the RF method was chosen as the best method among basic classification methods, and the Ensemble SVM and Ensemble RF methods were selected as the best methods in the ensemble and classification phase. It was also observed that using the ensemble approach led to a desirable improvement in horizontal optical depth classification. In the second approach, a method titled ML-Based GMDH, which improves the GMDH neural network by altering partial functions with machine learning algorithms, was used for estimating dust concentration. Additionally, to achieve suitable accuracy, the hyper-parameters of this model were finely tuned using the TLBO optimization algorithm. The results showed that this method provided appropriate accuracy in estimating dust concentration and horizontal optical depth, out performing the best-selected methods from the first approach