The research’s study area was gathered from Air Quality Monitoring Stations (AQMS) in the states listed in Tab. 1. For data gathering, active stations were segregated. These study sites were chosen based on the availability of CPCB air quality data, satellite images, and the fact that the areas chosen were the most polluted, trafficked, and prone to industrial development activities. Some states have inactive AQMS and so they are identified first before selecting the sites. States having active air monitoring stations details are isolated. For the planned job, the network was trained using AQMS from various states across the country. The proposed paper focus on overall air pollution prediction of the country which can further be narrowed to particular state or city area. In our study nearly 21 states are selected for experiment.

The features to be collected from the specific site are processed once the study area has been established. It is critical to comprehend the data in order to recognize the features. As a result, self-reviewing data is required, and it is assessed for all of the chosen states or cities. Fig. 2 shows a flow diagram of the data selection process.

For the available number of daily Air Quality Index data per city, about 37000 records for each station are taken on an hourly basis for the specified study areas from 2016 to 2020. The data was collected for three seasons: summer, rainy season, and winter. Before preprocessing, data collected from the KAGGLE website is rigorously scrutinized. Tab. 2 lists the features that have been identified for the proposed work. It is vital to comprehend the government-mandated averaging monitoring hours and minimal ambient concentration of air pollution levels.

Once the necessary data has been gathered, it is standardized to eliminate the effects of missing numbers. Fig. 3 shows the stages involved in normalizing. Missing data is critical in preprocessing and has a significant influence on its own, thus diagnosing missing values with adequate data is critical. For these reasons, unknown values other than numbers are deleted from input data before transformation for complex numbers with a special number called Not a Number (NaN).

For training and testing purposes, we divided the input data into two portions. Nearly 70% of the 37000 records gathered are used for training, and 30% are used for testing. Ground truth parameters are collected during training, and the network is trained using the Stochastic Gradient Descent with Momentum (SGDM) optimizer. In comparison to other current algorithms, this best approach finds the model parameters that best fit the expected and actual outputs, calculates faster, and converges better with longer training time. Before training LSTM categorization, soft max is employed for activation layer during input data testing.

Finally, the survey data is analyzed using methods from the Statistical Package for Social Sciences (SPSS). This SPSS software suite was used to conduct a detailed analysis of the data collected. The measurements done often includes mean, median, Standard Deviation (SD), Mean Absolute Percentage Error (MAPE), Mean Square Error (MSE), Mean Absolute Error (MAE), Root Mean Squared Error (RMSE) and Mean Squared Error (MSE) helps to predict the performance level of classifier which enables to conclude AQI.

In comparison to simple RNN, the LSTM network is capable of performing long-term dependencies, which was first described by Hochreiter and Schmidhuber in 1997. It allows avoiding the long-term reliance problem. The core idea behind LSTM is as follows:

The key feature that goes horizontally through the diagram at the top is cell state. It’s similar to a conveyor belt, but with a few more interactions. This cell state can be added or withdrawn based on the information and is regulated accordingly using a three-gate structure. As shown in Fig. 5 this regulation consists of a (σ) sigmoid neural net layer and a (x) point-wise multiplication operation. The main purpose of this sigmoid layer is to output values that are either zero (to signal “allow nothing through”) or one (to indicate “let everything through”).

Initially, the input data, as well as the input data concentration sequence before and after transformation, are defined. For sequence to label classification, layer array is created which includes sequence input layer, LSTM layer, fully connected layer, soft max layer and classification output layer. Sequence input layer represents total number of input features taken for study and the classes required for algorithm as decided is specified by fully connected layer. The basic block diagram of LSTM classification and regression is shown in Fig. 6:

First the gender of the subject is analyzed for the given input x_t and the output value h_t. The sigmoid layer checks h_t−1 and x_t and accordingly gives the output of number between 0 and 1 for each numbers in the cell state C_t−1 as per Eq. (1). The new candidate value vector is created by tanh.

(1) ft=σ(Wf.[ht−1,Xt]+bf)

Next C_t is added to the new state followed by adding gender of the subject to the cell state as given in Eq. (2):

(2) C~t=tanh⁡(Wc.[ht−1,Xt]+bc)

Now old cell state is updated C_t−1 into cell state C_t. Later forgetting of previous information is performed by multiplying f_t with old state and adding it with C_t as shown in Eq. (3):

(3) Ct=σ(ft∗Ct−1+it∗C~t)

Finally the output is decided from the cell state C_t.

The LSTM starts with the details of the input data that will be given to the network, and this choice is made by a sigmoid layer dubbed the “forget gate layer” (h_t−1 and x_t), which produces a number between 00 and 11 for each cell state (C_t−1). The ‘input gate layer’ analyses the new information that needs to be stored in the cell state and determines which values need to be changed. The tanh layer follows the input gate, creating a vector of new added values C_t−1, which is then concatenated to provide an update to the cell state. We usually set the input values to tanh between −11 and 11 and multiply with the output sigmoid gate to only consider a certain section of the state [21].

Data collected contains some unusual or missing data and so this impact may create side effects on the whole records and so data cleaning is very important before data preprocessing. There are numerous frequent methods to replace the missing values such as mean-median of previous or next value of current data R interpolation. Data acquisition frequently involves aberrant or missing data, which might have unforeseen repercussions for the full set of records. As a result, prior to data preparation, data cleaning is essential. Missing data is removed and relevant gaps are filled in using command tools. R-interpolation, mean-median of the previous or next value of the current data is all common ways for substituting missing values. The normalized input data is shown in Tabs. 3 and 4.

Statistical Validation of Extracted Features is done before classification, each piece of data that is used as an input must be evaluated for its importance. Tab. 5 shows the proposed characteristics and their accompanying metrics following validation. Number of samples (N), Standard Deviation (SD), Standard Error (SE), degree of freedom (df), Mean Square (MS), (measure of test accuracy) F1 Score, and significant are among the evaluation measures. The data was tested for the normality using Shapiro Walik Test and it was found that all the data was normally distributed and its significance of air pollutants is less than 0.05.

One-way Analysis of Variant (ANOVA) was used to validate the input features. It can be used for further processing if the significant value is less than 0.05. Following validation, it was determined that all of the input features used in the study were significant, implying that all of the input characteristics used in the proposed study can be used for further classification using machine learning and deep learning algorithms.

All of the significant features that have been validated using SPSS tools are used for classification. The corresponding sequence of air pollutant PM_2.5 for a given set of time T is defined as X, and these values are filled with record means to get PM_2.5 concentration sequence X¯, and afterwards these data sequences are translated into supervised learning format. Because of its lengthy temporal dependency problem, the simple RNN cannot cope with large amounts of data. To overcome this, we use LSTM, which takes a lagging observation t−1 as an input variable and uses it to forecast the current time step T. The modified data set is represented by a X¯, while the output variable is represented by a Y¯. These sequences are then used to forecast individual PM_2.5 series, and the process is repeated for all of the other features in the proposed study. Finally, SPSS tools are used to compare the prediction outcomes, and the performance of the classifier is evaluated using various attributes such as root mean squared error, mean, median, standard deviation, and so on. Using these assessment markers, LSTM is found to be superior to other models in processing time series data, indicating that the current model is useful in AQI prediction. Fig. 7 depicts the steps involved in defining the LSTM algorithm prior to network training.

LSTM begins with initialization of sequence of input layers needed, fully connected layer, soft-max layer and classification layer. Then after training options are given which includes initial learning rate, (Ridge Regression) L2 regularization, drop periods, drop factors, epochs needed, batch size and SGDM. Once relevant initialization is completed then the input data’s are converted to array format and later on input and ground truth are compared in activation layer. Finally the output is predicted based on the metrics such as accuracy, precision, error rate, sensitivity, specificity, F1score.

With proper initialization of training options the network is trained for classification. Defining LSTM layers includes input sequence (fully connected layer), LSTM 120 (soft-max) and LSTM 60 (classification layer). Initialization of learning rate (0.1), L2 regularization (0.0001), schedule (piecewise), drop factor (0.1), drop period (100), maximum epochs (500), mini batch size (128), and shuffling for every epoch plots are some of the learning rate of training options.

Each epoch is trained using 200 iterations, and it was discovered that the mini-batch loss and iterations are inversely proportional, with the batch loss reducing as the number of iterations grows. Tab. 6 shows the network’s initial stage of training.

The accuracy and other characteristics are assessed for various epochs after the network has been trained for the above configurations constructed according to the suggested LSTM model. Fig. 8a through Fig. 8h illustrate the relevant network training plots.

Over or under fitting can create to classification issues in a training network, hence regularisation is crucial. In machine learning, regularisation is used to solve this problem, and in deep learning, dropout regularisation is used to prevent over-fitting and under-fitting by removing random neurons from hidden layers. For large data sets hold-out validation works good compared to cross-out validation.

In general, having too many epochs might lead to the model overfitting the training data. It signifies that the model memorises rather than learns the data. The accuracy of validation data is checked for each epoch or iteration to see if it over-fits or not. The number of epoch determines how the network’s weights are changed. As the number of epochs grows, so do the number of times the neural network’s weights are modified, and the border shifts from underfitting to optimal to overfitting.

For better performance, training data is shuffled for every epochs. As CPU is the available source mini batch size can be implemented that represents short sequences. Once all the desired configuration is inserted the network starts training. For every epoch and iterations the accuracy level and corresponding error rate is plotted. During run time the behaviour of network is analyzed by its accuracy level and error rate.

A 64-bit operating system AMD A4-5000 APU with Radeon (TM) HD graphics with 1.50 GHz and 8:00 GB RAM is utilized in conjunction with MATLAB 2019a for modeling, processing, comparisons and visualizing the experimental numbers and findings through various deep learning algorithms such as Support Vector Machine (SVM), Neural Network (NN), K-Nearest Neighbor (KNN), Naive Bayes (NB), Ensemble (EN) and LSTM.

The study’s shortcoming was that the computation time was prolonged. The proposed study also has the disadvantage of not monitoring air pollutant concentrations in conjunction with other AQMS around or adjacent to it. Normally, both physical and chemical features of aerosols are used to predict air quality, however biological components and qualities are limited in this case. To increase the measurement level, future work can be expanded by including more air contaminants and additional data such as satellite images and industrial emissions into the atmosphere. To further understand the consequences of air pollution and human action, the article can be expanded by looking at specific states in relation to the current pandemic, as well as the situation before and after lockdown. Also, harmful air pollutants can be projected in advance for specific sites such as homes or roads, and the same can be combined with Internet of Things (IoT) and updated in real time in cloud computing for the benefit of people.