Classification Prediction of PM₁₀ Concentration Using a Tree-Based Machine Learning Approach

Abstract

The PM₁₀ prediction has received considerable attention due to its harmful effects on human health. Machine learning approaches have the potential to predict and classify future PM₁₀ concentrations accurately. Therefore, in this study, three machine learning algorithms—namely, decision tree (DT), boosted regression tree (BRT), and random forest (RF)—were applied for the prediction of PM₁₀ in Kota Bharu, Kelantan. The results from these three methods were compared to find the best method to predict PM₁₀ concentration for the next day by using the maximum daily data from January 2002 to December 2017. To this end, 80% of the data were used for training and 20% for validation of the models. The performance measure of the PM₁₀ concentration was based on accuracy, sensitivity, specificity, and precision for RF, BRT, and DT, respectively, which indicates that these three models were developed effectively, and they are applicable in the prediction of other atmospheric environmental data. The best model to use in predicting the next day’s PM₁₀ concentration classification was the random forest classifier, with an accuracy of 98.37, sensitivity of 97.19, specificity of 99.55, and precision of 99.54, but the result of the boosted regression tree was substantially different from the RF model, with an accuracy of 98.12, sensitivity of 97.51, specificity of 98.72, and precision of 98.71. The best model can assist local governments in providing early warnings to people who are at risk of acute and chronic health consequences from air pollution.

1. Introduction

Particulate matter is a well-known air contaminant that has been linked to a variety of negative health effects. When compared with the other criterion contaminations, particulate matter is the most prominent pollutant in Peninsular Malaysia, with the highest air pollutant index (API) value. According to one study [1], until 2017, PM₁₀ contributed the most to Malaysia’s API, whereas PM_2.5 had a substantial impact on APIs starting in the middle of 2017.

Particulate matter consists of microscopic liquid droplets or solids that can travel deep into the lungs of humans and cause serious health problems. Studies [2,3,4] have stated that particulate matter can cause lung or heart disease, asthma, irregular heartbeat, heart attacks, decreased lung function, and increased respiratory symptoms such as coughing, airway irritation, and breathing difficulties.

In other studies [5,6], it was stated that vehicle traffic is one of the main causes of PM₁₀ pollution in Malaysia due to primary and secondary emissions from exhausts, as well as suspended dust from the streets caused by circulation. In addition, power plants, open burning and wildfires, industrial facilities, and other sources also contribute to this pollution.

Several recent studies have developed models to predict PM₁₀ concentrations such as WRF-CMAQ and WRF-Chem. Weather research and forecasting (WRF) models combined with community multiscale air quality (CMAQ) models (WRF-CMAQ) were developed to predict the mesoscale meteorology condition and to simulate the regional air quality, respectively [7,8,9,10]. The WRF-CMAQ model system provides advances in terms of the capability to simulate complex atmospheric processes that transport and transform pollutants in a dynamic environment [11].

However, the interactions between air pollutants and meteorological conditions could decrease the accuracy of the WRF-CMAQ models, especially in heavily polluted episodes. Ref. [12] reported that the performance of the WRF model inaccurately reproduced low surface temperature and overestimated the wind speed as the pollution loading increased, thus affecting the CMAQ model. There was also another prediction model developed associated with WRF, which is the weather research and forecasting coupled with the chemistry module (WRF-Chem) system. WRF-Chem incorporated complex gas-phase chemistry, aerosol treatments, and photolysis scheme to investigate the influence of different chemical mechanisms on aerosol concentrations [13,14,15]. Chemical transport and transformations are incorporated into WRF so that the interactions between meteorology and chemistry can be investigated [16]. Different from the WRF-CMAQ model, the WRF-Chem system showed the worst performance, specifically during episodes dominated by coarse particles [17]. WRF-Chem model was found to underestimate aerosol optical depth (AOD) because of the misinterpretation of the coarse particles.

Advanced statistical models based on machine learning (ML) techniques have been widely applied in the field of air quality modelling [18,19] because they can perform large and complex data analysis to predict a potential outcome efficiently [20]. Artificial intelligence is used in these techniques, which learn the patterns in the dataset and then construct and train a predictive model [21].

The decision tree (DT) analysis is a data mining and machine learning analytical method that generates a tree-based classification model that classifies cases or predicts values of a dependent variable based on the values of independent variables. Ref. [22] introduced the term ‘classification and regression trees’ (CART) to refer to DT algorithms that can be used to solve classification or regression predictive modelling challenges. This algorithm is known as ‘decision trees’ in the traditional way, but in some platforms, such as R, it is referred to as ‘CART’. The DT technique was originally applied to statistics by [22]. Ref. [23] explained how the DT technique in machine learning was established, while the authors of [24] described decision trees from a statistical approach.

Two well-known ensemble approaches are bagging [25] and boosting [26,27]. They were developed to improve CART’s stability by generating multiple tree models and integrating their outputs to obtain a final prediction. According to [25], the random forest (RF) algorithm applies the bagging approach to an ensemble of DT that trains numerous trees in parallel and uses the majority decision of the trees as the RF model’s final decision. The boosted regression tree (BRT) algorithm, stated by [28], uses boosting to randomly resample training datasets (without replacement) and build a sequence of trees, with each new tree focusing on poorly fitted cases. As a rule of thumb, the authors of [28] suggest using 1000 trees. However, this is based on a detailed analysis of predictive stability for the specific dataset used in the paper.

Furthermore, there are various methods used by previous researchers to predict air pollution concentrations using classification-based predictions. For instance, a study by [29] was performed to determine the major pollutants present in India. The results showed that the DT algorithm provided the highest accuracy. A study conducted by [30] suggested that the multilabel classification be used to predict multiple pollutants since it computes more accurate posterior probabilities, which better supports the decision maker. Then, the authors of [31] stated that classification-based predictions can be measured by accuracy, and the results of this study showed that the multilayer perceptron gave the best results in predicting the levels of PM₁₀ concentration.

The aim of this study is to develop a new approach to predict and classify PM₁₀ concentration levels in Malaysia using machine learning tree-based techniques, i.e., decision tree (DT), boosted regression tree (BRT), and random forest (RF). This study also presents the results of the most relevant features in predicting PM₁₀ concentration levels in Kota Bharu, Kelantan.

2. Materials and Methods

2.1. Study Area

Kota Bharu is a town located in Kelantan, and the station is located at Sekolah Menengah Kebangsaan Tanjong Chat, Kelantan. It is the royal seat and the state capital of Kelantan. The latitude and longitude for this station are 6°6′28.42″ N and 102°15′5.01″ E. This location is classified as an urban area by the Department of Environment (DoE), Malaysia. Its location on the Peninsular’s east coast, facing the South China Sea, in the path of the cold surge, results in a harsh climate during the months of November, December, and January due to the annual monsoon season. Since Kota Bharu is Kelantan’s capital, its population is higher than the population of other parts of the state. The availability of jobs in this main city has resulted in rising population density and economic development, which might result in increased pollution.

2.2. Monitoring Records

In order to acquire a better grasp of PM₁₀ variability, this study used eight maximum daily parameters across a sixteen-year period (2002–2017). Gaseous parameters such as carbon monoxide (CO, ppb), nitrogen dioxide (NO₂, ppb), sulphur dioxide (SO₂, ppb), particulate matter with an aerodynamic diameter less than 10 µm (PM₁₀, µgm⁻³), and ozone concentration (O₃, ppb), as well as meteorological parameters such as wind speed (WS, Km/h), relative humidity (RH, %) and temperature (T, °C), were used as predictors for the next day’s PM₁₀ concentration. WS, RH, T, and previous PM₁₀ concentrations were selected as independent variables since, according to [32,33,34], they had the most significant effect on future PM₁₀ concentrations. On the other hand, Ref. [35] discovered that additional air pollutants, such as SO₂, NO₂, and CO, can similarly enhance PM₁₀ concentration predictions. Furthermore, Ref. [36] found no significant correlation between PM₁₀ and wind direction. Therefore, wind direction does not affect PM₁₀ concentration prediction.

This study used linear interpolation for missing data imputation. According to [37], this linear interpolation method estimates the missing data better for the air pollution data.

The monitoring records were obtained from Malaysia’s Ministry of Environment and Water’s Department of Environment (DoE). Then, 80% of the monitoring data were used for model training, while 20% was used for model validation. In this study, value 1 was set as indicative of a low level of air quality, if the concentration level did not exceed 100, which is the index value in the current 24 h PM₁₀ standard, while value 2 represented a high level of air quality if it exceeded 100;

Table 1. Class labels for the next day’s PM₁₀ concentration.

PM10 Concentration Value	Class Label
Low-level air quality with, PM10 < 100	1
High-level air quality with, PM10 ≥ 100	2

shows the labelling for the next day’s PM₁₀ concentration. Thus, in this study, the threshold value was considered based on the Malaysia Ambient Air Quality Guideline 2019 [38].

To avoid imbalanced classification in developing predictive models on these classification datasets, this study used the synthetic minority oversampling technique (SMOTE), as proposed by [39]. In the first step, the dataset was filtered to only consider the minority class, which was the high-level air quality, with 95 data points only. Following that, a search of the k-nearest neighbours was performed (k = 5). The algorithm then selected a random nearest neighbour for these data. A new data point was created, which was on the line between the two data points. The results before and after using the SMOTE technique are shown in Figure 1.

2.3. Tree-Based Machine Learning Approaches

Three models were used for the prediction of PM₁₀ concentration for the next day by using the maximum daily data in Kota Bharu, Kelantan. The models are decision trees (DTs), random forest (RF), and boosted regression trees (BRTs). In addition, in the process of model development, model prediction, and model evaluation, RapidMiner Studio was used to predict the air pollution concentration. Moreover, 80% of the data from January 2002 to December 2017 were divided into two parts, which means 80% of the data were used for training and 20% for validation of the models. Then, the results of each model were compared to find the best-proposed model for predicting PM₁₀ concentration.

Table 2. General model for the next day’s prediction.

Prediction	Models
PM10,D+1	PM10,D+1 ~ DT (CO (D), PM10 (D), NO2 (D), SO2 (D), O3 (D), T (D), RH (D), WS(D),)
	PM10,D+1 ~ BRT (CO (D), PM10 (D), NO2 (D), SO2 (D), O3 (D), T (D), RH (D), WS(D),) PM10,D+1 ~ RF (CO (D), PM10 (D), NO2 (D), SO2 (D), O3 (D), T (D), RH (D), WS(D),)

shows the general model of DT, RF, and BRT in predicting next-day PM₁₀ concentration (PM_10,D+1). The dependent variable was next-day PM₁₀ concentration represented by PM_10,D+1, while the independent variables were daily maximum CO_, PM₁₀, NO₂, SO₂, O₃, T, RH, and WS represented by CO _(D), PM_{10 (D)}, NO_{2 (D)}, SO_{2 (D)}, O_{3 (D)}, T _(D), RH _(D), and WS_(D).

2.3.1. Decision Tree

The decision tree (DT) model is a nonparametric method that can be used to predict a variety of quantitative and qualitative variables. In the form of a tree structure and with a reciprocal classification of data, a DT model illustrates direct and indirect correlations of numerous independent variables with a target variable (dependent) [40]. Variables on the upper branches of the tree structure have a greater impact on the prediction of the related class. For the DT model, four split measures were tested using gain ratio, information gain, Gini index, and accuracy.

2.3.2. Random Forests

Random forests (RFs) are a collection of methods for assembling a collection (or forest) of decision trees [41]. According to [25], the random forest (RF) approach applies the bagging technique to an ensemble of DTs, through which it trains many trees concurrently and uses the majority judgment developed in DTs as the RF model’s final decision. After that, the outputs of each tree are combined to create an ensemble prediction of the target variable. The model also estimates the most relevant features by determining how much the prediction error increases when data for that variable are permuted, while the rest of the data are not [42]. For RF models, four split measures were tested using gain ratio, information gain, Gini index, and accuracy.

2.3.3. Boosted Regression Tree

Boosted regression tree (BRT) models are developed by integrating two algorithms: Decision trees [11] are used to fit a series of single models, and then boosting [26] is used to aggregate their outputs to obtain the total prediction. A thorough description of the approaches can be found in [6,43]. The learning rate (lr), which is the shrinkage parameter used in each iteration to reduce the contribution of the tree, the complexity of the tree (tc), which is the maximum tree depth of variable interactions, as well as the number of trees (nt), are all tuning parameters that must be controlled in the BRT model. In this study, default settings for lr (0.01), tc (5), and nt (1000) were used to fit BRT models in RapidMiner. To optimise the number of trees in BRT, the optimise parameters (grid) operator was used, which is a nested operator. The loss functions or distributions that were used for this study were multinomial and Bernoulli distributions.

The DT, BRT, and RF models were run using RapidMiner software with the abovementioned independent variables (CO_(D), PM_10(D), NO_2(D), SO_2(D), O_{3 (D)}, T_(D), RH_(D), WS_(D)) and dependent variable (PM_10,D+1), with two class labels.

2.4. Performance Measures

Several performance measures—namely, accuracy, sensitivity, specificity, and precision values—were used to evaluate each classification model in this study. The formula for the performance measures used in this study is shown in Equations (1)–(4).

$$\mathrm{Accuracy}=\frac{\mathrm{TP}+\mathrm{TN}}{(\mathrm{TP}+\mathrm{FP}+\mathrm{TN}+\mathrm{FN})}$$

(1)

$$\mathrm{Sensitivity}=\frac{\mathrm{TP}}{(\mathrm{TN}+\mathrm{FP})}$$

(2)

$$\mathrm{Specificity}=\frac{\mathrm{TN}}{(\mathrm{TN}+\mathrm{FP})}$$

(3)

$$\mathrm{Precision}=\frac{\mathrm{TP}}{(\mathrm{TP}+\mathrm{FP})}$$

(4)

where TP is the true-positive, TN is the true-negative, FP is the false-positive, and FN is the false-negative value based on a confusion matrix. To support the accuracy values, we used sensitivity, specificity, and precision values as suggested by [44]. The overview of the experiments performed for this study is shown in Figure 2.

3. Results and Discussion

3.1. Statistical Characteristics of PM₁₀

The descriptive statistics of all independent variables are shown in

Table 3. The descriptive statistics of independent variables (IVs).

IV	Minimum	Maximum	Mean	Standard Deviation	Skewness	Kurtosis
WS	0.01	360	9.50	8.04	31.08	1346.58
T	22.89	37.5	31.36	2.33	−0.63	0.21
RH	3.22	100.2	91.86	6.80	−5.92	59.73
SO2	0	71.4	0.90	1.54	19.54	836.04
NO2	0	63	15.15	6.23	1.22	4.16
CO	180	21712	926.37	475.30	16.44	689.80
O3	0.9	69	29.21	10.99	0.23	−0.05
PM10	12	198	48.73	18.11	1.72	6.22

. The mean of the PM₁₀ concentration was 48.73 µg/m³ below the specified MAAQG for the yearly average of 50 µg/m³ [45]. The distribution was highly skewed for PM₁₀ (1.72), CO (16.44), NO₂ (1.22), SO₂ (19.54), RH (−5.92), and WS (31.08) since the values were less than −1 or more than +1 [6]. During the 2002–2017 period, the data revealed the presence of an extreme level of concentration in Kota Bharu.

The box plot results in Figure 3 also showed that PM₁₀ concentrations had extreme values since they consisted of many outliers. The highest extreme value indicates the maximum daily reading in 16 years of PM₁₀ concentration. The maximum PM₁₀ concentration was reported at 198 g/m³.

3.2. Decision Tree (DT)

The performance values for each of the DT splitting criteria are shown in

Table 4. The best splitting criteria for the decision tree.

Splitting Criteria	Accuracy	Sensitivity	Specificity	Precision
Gain ratio	95.81	92.59	99.04	98.98
Information gain	96.52	95.78	97.25	97.21
Gini index	96.52	95.02	98.02	97.69
Accuracy	93.39	95.53	91.25	91.61

. Overall, the results indicate that all of the splitting criteria can be used to classify the air pollution dataset. On the other hand, the information gain produced the highest accuracy value in terms of total performance and recorded an impressive accuracy value of 96.52%, compared with the other splitting criteria, which was supported by sensitivity (95.78%), specificity (97.25%), and precision (97.21%).

Figure 4 shows the decision tree classifier with eight features included in the model. The most influential variable regarding PM₁₀ concentration levels was SO₂ for gain ratio (0.33) and accuracy (0.21), while the second-most influential parameter would be WS for gain ratio (0.31) and accuracy (0.20). For information gain and Gini indices, the most influential variable was WS.

3.3. Boosted Regression Tree (BRT)

The performance values for each of the BRT distributions are shown in

Table 5. The best distribution for the boosted regression tree.

Distribution	Accuracy	Sensitivity	Specificity	Precision
Multinomial	98.12	97.51	98.72	98.71
Bernouli	98.12	97.75	98.66	98.64

. Overall, the results indicate that both distributions can be used to classify air pollution data sets. However, the best distribution was the multinomial distribution, which had the same accuracy (98.12%) as the Bernoulli distribution but was supported by sensitivity (97.51%), specificity (98.72%), and precision (98.71%).

The boosted regression tree classifier, which has eight features, is shown in Figure 5. The most influential (PM₁₀) and second-most (WS) influential variables were the same for both distributions. The parameter with the least significant influence on PM₁₀ was ozone concentration, with only 811.43 (multinomial) and 958.16 (Bernoulli). After optimising the number of trees using the parameters (grid) operator, the best number of trees to be used in this study for multinomial was 500, and for Bernoulli, it was 450.

3.4. Random Forest (RF)

Table 6. The best splitting criteria for the random forest.

Splitting Criteria	Accuracy	Sensitivity	Specificity	Precision
Gain ratio	97.03	94.44	99.62	99.6
Information gain	98.37	97.19	99.55	99.54
Gini index	98.27	97	99.55	99.54
Accuracy	94.5	96.68	92.33	92.65

shows the performance numbers for each of the RF splitting criteria. Overall, the results show that the air pollution dataset can be classified using all of the splitting criteria. In comparison to the other splitting criteria, the information gain produced the highest accuracy value in terms of total performance, with an exceptional accuracy of 98.37%, which was confirmed by sensitivity (97.19%), specificity (99.55%), and precision (99.54%).

The random forest classifier, which has eight features, is shown in Figure 6. The most influential variable on PM₁₀ concentration was WS for all methods of splitting criteria in RF, with a gain ratio of 0.22, information gain of 0.22, Gini index of 0.24, and accuracy of 0.19. The temperature was the second-most influential factor.

3.5. Performance Comparison

According to the results of these performance error and accuracy measurements, the RF model outperformed BRT and DT models. As a result, in Kota Bharu, RF predicted PM₁₀ concentrations better than BRT and DT. The values of the performance measures based on accuracy, sensitivity, specificity, and precision are shown in

Table 7. Performance comparison for decision tree (DT), boosted regression tree (BRT), and random forest (RF).

Models	Accuracy	Sensitivity	Specificity	Precision
DT	96.52	95.78	97.25	97.21
RF	98.37	97.19	99.55	99.54
BRT	98.12	97.51	98.72	98.71

. Overall, the best model to use in predicting the next day’s PM₁₀ concentration classification was the random forest classifier, with the five best parameters of WS, T, RH, NO₂, and PM₁₀.

Table 8. Performance comparison and the results gained from other researchers.

Authors	Method	Accuracy	Sensitivity	Specificity	Precision
[31]	Multilayer Perceptron (MLP)	98.1	-	-	98
	Support vector machines (SVM)	92.5	-	-	92
	Naïve Bayes	91.25	-	-	90
[29]	Logistic Regression (LR)	98	97	98	97
	Naïve Bayes	97	98	97	93
	K-nearest Neighbour (KNN)	97	97	98	97

summarises the comparison results with other researchers. The data indicate that the results in this study are quite similar to those of other researchers. In addition, the values of the accuracy are larger for the RF model, compared with those found by other models. This shows that the RF model can be used to predict PM₁₀ concentrations since it improved the performance of the model.

4. Conclusions

Based on the results of this study, the random forest classifier performed the best in predicting for the next day’s PM₁₀ concentration classification, with an accuracy of 98.37, sensitivity of 97.19, specificity of 99.55, and precision of 99.54, but the results showed that the BRT performance was not substantially different from that of the RF model, with an accuracy of 98.12, sensitivity of 97.51, specificity of 98.72, and precision of 98.71.

Next, the wind speed was the most relevant feature to classifying the next day’s PM₁₀ concentration as indicative of a low or high level of air quality in RF and DT techniques, but for BRT, PM₁₀ was the most relevant feature.

Overall, this study’s findings show that machine learning algorithms can help classify the next day’s PM₁₀ concentration as indicative of a low or high level of air quality. The three classifiers applied in this research also covered the most relevant features in PM₁₀ concentration prediction. The best model (RF) is suitable to predict PM₁₀ concentration in Kota Bharu, Kelantan, for early warning systems and for local authorities to develop strategies to improve air quality.

The model can only be used when the sources and conditions of PM₁₀ concentration remain constant, which is a limitation of this study. As a result, it may not be appropriate for the other areas. Furthermore, if there is a sudden forest fire or storm in a specific area, the PM₁₀ concentration will be affected.