FEATURE SELECTION AND DEEP LEARNING MODEL FOR AIR QUALITY PREDICTION ↔ FEATURE SELECTION AND DEEP LEARNING MODEL FOR AIR QUALITY PREDICTION

IJAIDL IAEME Publication Chennai, India editor@iaeme.com/ iaemedu@gmail.com https://iaeme.com/Home/journal/IJAIDL INTERNATIONAL JOURNAL OF ARTIFICIAL INTELLIGENCE AND DEEP LEARNING https://iaeme.com/Home/journal/IJAIDL 1 https://iaeme.com/Home/journal/IJAIML 62 editor@iaeme.com International Journal of Artificial Intelligence and Deep & Machine Learning (IJAIDL) (IJAIML) Volume 2, 4, Issue 2, July-December 2025, pp. 1-11, 62-72, Article ID: IJAIDL_02_02_001 IJAIML_04_02_005 Available online at https://iaeme.com/Home/issue/IJAIDL?Volume=2&Issue=2 https://iaeme.com/Home/issue/IJAIML?Volume=4&Issue=2 Impact Factor (2025): 22.45 (Based on Google Scholar Citation) Journal ID: 2311-3214 9339-1263; DOI: https://doi.org/10.34218/IJAIDL_02_02_001 Madhya Pradesh- 474011, https://doi.org/10.34218/IJAIML_04_02_005 Uttar Pradesh – 201303, India. ABSTRACT The growing number of vehicles and industry pollution leads to poor air quality, which brings serious health issues and lung problems for humans. There are safety measures followed by the Government to regulate the air quality, the detection of air quality is challenging due to the implementation cost and computations. The air quality measure is to be continuously monitored under urban areas to ensure healthy living. Artificial intelligence is the growing area in all industries, thus this work is proposed to exploit the advantages of deep learning algorithms for air quality detection. The proposed work used feature selection technique for selecting the important and relevant features brings high accuracy of air quality detection. The wrapper based technique, Recursive feature elimination (RFE) is used for feature selection and deep learning model Deep Neural Network (DNN) is used for predicting the air quality. This prediction gives an alert for the surrounding to work on further measures in controlling pollution. Experimental results showed that the deep learning model outperforms in air quality prediction with the highest accuracy. Keywords: Machine Learning, Deep learning, Recursive feature elimination, Feature selection, Air quality, Deep Neural Network. Agastya Saksham Shrivastava Gupta https://iaeme.com/Home/journal/IJAIDL https://iaeme.com/Home/journal/IJAIML 2 63 editor@iaeme.com Cite this Article: Agastya Saksham Shrivastava. Gupta. (2025). Feature Selection and Deep Learning model for Air Quality Prediction. International Journal of Artificial Intelligence and & Deep Machine Learning (IJAIDL), (IJAIML), 2(2), 4(2), 1-11. 62-72. DOI: https://doi.org/10.34218/IJAIDL_02_02_001 https://doi.org/10.34218/IJAIML_04_02_005 1. INTRODUCTION Air pollution is generally caused by high levels of PM2.5, which is called the particulate matter having an aerodynamic diameter less than 2.5 micron has been one of the growing threats around the world. Urbanization and industrial growth are the major cause for air pollution and less quality of air leads to the health issues to humans. Governments are taking necessary measures to control air pollution, abiding by the World Health Organization's standards. Air pollution prediction is crucial in modern cities to monitor the air quality and it is also challenging. Thus it has gained research attention these years, to study and overcome these issues. The emergency action to be taken for any high-level pollution events majorly depends on the prediction precision from the data centers and its promptness. Artificial intelligence is emerging in all industries such as health care, finance, agriculture, automobiles and more. The growth of Internet of Things (IoT) is also beneficial for regular monitoring of air quality throughout various monitoring centers with less hardware and using distributed cloud environments to handle collected data from various centers. The data- driven air pollution prediction is challenging with the collected data. The proposed work aimed at air pollution prediction with artificial intelligence techniques to alleviate accuracy of the prediction. The challenges in these systems is partial observation of attributes/factors that need to be carefully handled. This system has to predict the possible pollution event to handle emergency situations. The occurrence of various pollutants such as nitrogen dioxide, (NO2), carbon monoxide (CO), sulphur dioxide (SO2) and particulate matter (PM) are the major causes for air pollution. The PM has a diameter less than 2.5 and less than 10 microns are considered as pollutants.

These pollution causes lung problems and diseases to human beings. Thus there are monitoring stations available to closely monitor these levels on a regular basis. There are challenges in collecting these pollution information from these stations and identifying any potential threat to the environment. Collecting these information and monitoring involves high human resources in traditional systems. These data need to be monitored regularly to make any quick decision based on alert. The recent growth in IoT technologies has provided a better opportunity Feature Selection and Deep Learning model for Air Quality Prediction https://iaeme.com/Home/journal/IJAIDL 3 https://iaeme.com/Home/journal/IJAIML 64 editor@iaeme.com to collect data from small sensors at different stations and process the collected data through cloud servers. The growth of AI technologies provides an opportunity for predicting the air pollution types at minimal cost and in the most accurate manner.

The proposed work is a deep learning approach for predicting air pollution. There are several features involved in air pollution data, thus to avoid the problem of over-fitting, the proposed work used a wrapper based feature selection approach. Feature selection technique, Recursive feature elimination selects only high important features from dataset for learning.

The accuracy of the air pollution prediction is increased with this proposed model. The objective of study is to implement air quality prediction with machine learning algorithm and deep learning algorithm, with feature selection and compare their results. The main contributions of this paper are • Air quality prediction through machine learning model, logistic regression and deep learning model, Deep Neural Network (DNN) and evaluate its performance • Implement feature selection algorithm for arriving best features from the attributes. • Minimizing the over-fitting problem with feature selection and choosing best features according to wrapper technique. This paper is organized as Chapter 2 gives detailed study of literature survey and related work in air quality prediction. Chapter 3 discusses proposed algorithms. Chapter 4 gives detailed discussion on results arrived for air pollution prediction. Chapter 5 draws conclusions based on this work and further enhancement possibilities. 2. RELATED WORK Air pollution leads to serious health issues and lung disease for humans. Traditionally there are only a few monitoring stations for air pollution monitoring. The growth in industries and urbanization brings more pollution. The existing studies are performed based on traditional statistical models. There are few research done based on artificial intelligence are discussed in this chapter. The adaptive probabilistic prediction is proposed in [1] for prediction for air quality, the model used in generative adversarial network with three layers namely inferences, generators and discriminators. This model is built with three hidden layers and 256 neurons. Experimental Agastya results Shrivastava showed https://iaeme.com/Home/journal/IJAIDL that 4 the editor@iaeme.com mean results square showed error that (MSE) the is mean 1.05. square However, error the (MSE) model is 1.05. deep However, learning the model classification is accuracy deep is learning very the less classification thus accuracy the is MSE very loss less is thus very the high. MSE Air loss pollution is prediction very with high. a Air Bayesian pollution deep prediction learning with model a is Bayesian proposed deep in learning [2], model the is work proposed used in domain [2], specific the knowledge work integration.

used The domain specific specific pollutant knowledge types integration. PM2.5 The and specific PM10 pollutant are types used PM2.5 as and regularization PM10 features. are The used multi as step regularization forecast features. model The is multi used step to forecast project model the is air used quality. to Saksham project Gupta the https://iaeme.com/Home/journal/IJAIML air 65 quality. editor@iaeme.com The data pre-processing preprocessing handled missing data and data interpolation. The implemented model had two layers one with RNN one time prediction strategy and another with recursive prediction strategy. Experimental results proved that Bayesian mode model reduced the error by 3.7%. Recurrent neural network type, Long short term memory (LSTM) model for air pollution prediction is discussed in [3]. To increase the accuracy of prediction and forecast, kernel density estimation with sub division tuning is performed. Spatially adjusted multivariate imputation is developed to manage the incoming data from different monitoring stations. The dataset also considered meteorological variables such as wind speed, direction, humidity and rainfall.

Experimental results showed that MAE error is 8.12. IoT sensing based air quality monitoring is widely used, the collection of data is made more reliable with sensor networks. The work [4] studied air quality monitoring through IoT networks. The regression models, support vector regressor, random forest regressor and Gradient Boosting regressor (GBR) are used. RMSE error 7.8 for GBR, which is more effective in air quality prediction. Deep learning architectures are highly used for air quality prediction, the work [5] implemented ConvLSTM for air quality forecasting. The model is four stacked convLSTM with five hidden states to extract features from historical data. Dropout layer is added in each ConvLSTM block that enhances the speed of learning. The mode is compared with SVR, from experimental results it is shown that deep learning model perform well than machine learning regression. This literature survey studied air quality prediction using various artificial intelligence models. Some of the work studied machine learning models, some of them used deep learning models. The deep learning model most widely studies is Long Short Term Memory (LSTM)

model as it considered the time series dataset. Some of the studied used regression analysis for air quality prediction. The literature review showed that these works are effective in predicting air quality, however mean square error is the evaluation metric used in most of the work used, which is very high. The higher the MSE error, the lower the model accuracy. Thus it is necessary to propose an effective model, which can accurately predict air quality. Feature Selection and Deep Learning model for Air Quality Prediction https://iaeme.com/Home/journal/IJAIDL 5 editor@iaeme.com 3. PROPOSED WORK Air quality is an important measure to be monitored in every location of the city and other regions such as industrial regions. Air pollution type prediction is considered as the problem in this proposed work. The attributes used for this problem are SO2, NO2, RSPM, SPM are the major pollutants used for air pollution type prediction. Pollution risk and exposure levels can Feature Selection and Deep Learning model for Air Quality Prediction https://iaeme.com/Home/journal/IJAIML 66 editor@iaeme.com be identified with air quality prediction, however, it is highly challenging due to dispersion of particles and meteorological factors. Existing approach is manually handled, it takes high cost and resources for monitoring at different stations. DATASET DETAILS The air pollution dataset is the Historical Daily Ambient Air Quality Data from the Ministry of Environment and Forests and Central Pollution Control Board of India.

The dataset contains the following features features. Table 1: Air Pollution Dataset Details Exploratory data analysis Features Description stn_code Station code, given to each station for recorded data.

sampling_date The date when the data was recorded State States, which air quality data is measured Location City in which data is collected Agency Name of the agency that measured the data.

Type The type of area where the measurement was made. so2 Sulphur Dioxide measured. no2 Nitrogen Dioxide measured Rspm Respirable Suspended Particulate Matter Spm Suspended Particulate Matter location_monitoring_station Location of the monitoring area. pm2_5 Particulate matter less than 2.5 micron Date Date of recording Exploratory data analysis Exploratory data analysis (EDA) helps in visualizing data with plots and helps data scientists to understand better on data pattern and its behaviour. This proposed work, data is visualized through some plots. The following plot shows the state-wise air pollution level. This plot helps not only to understand which state has highest pollution and which is lowest Agastya Shrivastava https://iaeme.com/Home/journal/IJAIDL 6 editor@iaeme.com pollution, this plot also helps in analysing each pollutant factor such as SO2, NO2 presence in each state. In the below plot, the air quality among all states are plotted, the colours in each bar represents the pollutants name, for example, blue represents the presence of SO2. From this visualization it is visible that air pollution is high in Andhra Pradesh and the next highest state is Maharashtra. Saksham Gupta https://iaeme.com/Home/journal/IJAIML 67 editor@iaeme.com Figure 1: State wise Air Pollution data in India METHODOLOGY In this section, the methodology of the proposed system, air pollution prediction is discussed. The work implements a machine learning model, logistic regression and deep learning model, Deep Neural Network (DNN). The dataset with 12 features are taken for this study. This work is classified as a multi class classification problem with four pollution types Residential, Industrial, RIRUO (Residential / industrial / rural / other areas), sensitive. The feature selection technique is used to select important features from the dataset. There are five independent variables, the top three important features are selected for this implementation through a scoring function performed by the RFE algorithm.

Feature Selection and Deep Learning model for Air Quality Prediction https://iaeme.com/Home/journal/IJAIDL 7 https://iaeme.com/Home/journal/IJAIML 68 editor@iaeme.com Figure 2: System architecture of Airpollution prediction The above figure represents the proposed system architecture, the dataset is loaded and first applied RFE technique to select the top three important features from the dataset. There are two models implemented and an air pollution type of four classes is predicted. FEATURE SELECTION Recursive feature elimination (RFE), a Wrapper based feature selection technique is implemented to select the top three important features based on the scoring function. The machine learning algorithm to be given as base classification model for RFE, logistic regression is given for RFE for classification and scoring. RFE scoring function computes score for every attribute and the least important attributes are removed from the list, thus giving the highly important features. The algorithm chosen top three features are SO2, SPM and PM2.5. The following table shows the attributes and its ranking given by RFE and top ranked features are selected for training. Table 2: RFE Scoring given to air pollution dataset Feature Rank SO2 1 SPM 1 RSPM 2 NO2 3 PM2.5 1 Agastya Shrivastava https://iaeme.com/Home/journal/IJAIDL 8 Saksham Gupta https://iaeme.com/Home/journal/IJAIML 69 editor@iaeme.com For classification problems, the scoring metric used by RFE is accuracy. Accuracy is computed based on the proportion of correctly classified instances. The formula to calculate accuracy is given in equation 1. -(1) The selected features are given input to the machine learning model Logistic regression, this is a classification algorithm, which takes independent features from the dataset and classifies the air pollution into four types. This classification is a multi -class problem with 4 types are Residential, Industrial, RIRUO and Sensitive.

DEEP NEURAL NETWORK Deep neural network with three layers as shown in the below figure is added. The first layer is a dense layer with input dimension 2, units 4 and a 'uniform' kernel. The second layer accuracy metric. The number of epochs trained is 10 with batch size 1000. Figure 3: Deep Neural Network Layers for Air pollution prediction 4. RESULTS AND DISCUSSIONS The proposed work air quality prediction is implemented in Python 3. Logistic regression and deep neural networks are implemented in the air pollution data. The data contains all state pollution information registered from 1990 to 2015 provided in the Indian Government portal. The dataset is visualized with matplotlib, the visualization shows the highest, moderate and lowest pollution states with pollutant parameters specified for each state. RFE feature selection Feature Selection and Deep Learning model for Air Quality Prediction https://iaeme.com/Home/journal/IJAIML 70 editor@iaeme.com is implemented to reduce the number of attributes to three with highest scoring on accuracy Feature Selection and Deep Learning model for Air Quality Prediction https://iaeme.com/Home/journal/IJAIDL 9 editor@iaeme.com basis. The logistic regression model, which is a classification model, is build for RFE as base classification model. The results shows that the deep learning algorithm has achieved the highest accuracy of air pollution classification. The below plot shows an evaluation metric computed for a logistic regression model for air pollution prediction. The results show that the accuracy of the model is 62.3%, the error metrics computed for classification are Mean square error is 0.65, mean absolute error and Root mean square error is 0.80. Figure 4: Evaluation metrics of Logistic regression for air pollution prediction The following plot shows the training accuracy for Deep Neural network algorithm, the plot shows the accuracy is 62.8% with 30 epochs. Figure 5: Training accuracy for air pollution prediction using DNN Agastya Shrivastava https://iaeme.com/Home/journal/IJAIDL 10 Saksham Gupta https://iaeme.com/Home/journal/IJAIML 71 editor@iaeme.com The following plot shows the training loss for Deep Neural network algorithm, the plot shows the loss is 0.6 with 30 epochs. Figure 6: Training loss for air pollution prediction using DNN 5. CONCLUSIONS This paper proposed air pollution prediction based on machine learning and deep learning models. The proposed work used a feature selection algorithm, RFE for selecting the important feature and reducing the over-fitting problem. The base classification model used for feature selection is Logistic regression, it is a classification algorithm, classifies and gives scores based on accuracy. The top ranked features, three important features are selected by the model. These features are trained with machine learning model Logistic regression and tested the performance and similarly a deep learning model Deep Neural Network is trained and evaluated its performance. Experimental results showed that deep learning has performed better in classification accuracy. This work can be further enhanced by implementing other deep learning models such as Convolutional Neural Network and Long short term memory model. These deep learning models can be optimized for the parameters such as batch size, epochs and number of layers. Feature Selection and Deep Learning model for Air Quality Prediction