Application of Artificial Intelligence Models for Aeolian Dust Prediction at Different Temporal Scales: A Case with Limited Climatic Data

Abstract

Accurately predicting ambient dust plays a crucial role in air quality management and hazard mitigation. Dust emission is a complex, non-linear response to several climatic variables. This study explores the accuracy of Artificial Intelligence (AI) models: an adaptive-network-based fuzzy inference system (ANFIS) and a multi-layered perceptron artificial neural network (mlp-NN), over the Southwestern United States (SWUS), based on the observed dust data from IMPROVE stations. The ambient fine dust (PM2.5) and coarse dust (PM10) concentrations on monthly and seasonal timescales from 1990–2020 are modeled using average daily maximum wind speed (W), average precipitation (P), and average air temperature (T) available from the North American Regional Reanalysis (NARR) dataset. The model’s performance is measured using correlation (r), root mean square error (RMSE), and percentage bias (% BIAS). The ANFIS model generally performs better than the mlp-NN model in predicting regional dustiness over the SWUS region, with r = 0.77 and 0.83 for monthly and seasonal fine dust, respectively. AI models perform better in predicting regional dustiness on a seasonal timescale than the monthly timescale for both fine dust and coarse dust. AI models better predict fine dust than coarse dust on both monthly and seasonal timescales. Compared to precipitation, air temperature is the more important predictor of regional dustiness on both monthly and seasonal timescales. The relative importance of air temperature is higher on the monthly timescale than the seasonal timescale for PM2.5 and vice versa for PM10. The findings of this study demonstrate that the AI models can predict monthly and seasonal fine and coarse dust, based on the limited climatic data, with good accuracy and with potential implications for research in data sparse regions.

1. Introduction

Aeolian dust storms degrade visibility and cause several hazards and human health problems including traffic accidents, degradation in industrial machinery, cardiovascular diseases, and lung cancers [1,2,3,4]. Fine dust (soil dust with particle matter ≤ 2.5 µm) contributes about 20–50% of total fine particulate matter (PM2.5) in the Western United States (WUS) [5]. Dust particles also play an important role in the global climate system, and the regional and local climate and environment, primarily by absorbing and scattering both solar and terrestrial radiation [6,7,8,9]). For example, aeolian dust on snow increases the snow albedo and accelerates snowmelt in the Colorado River [10]. Dust also alters the North American monsoon by heating the lower troposphere [11].

The dust emission, transport, and deposition are influenced by complex land–atmospheric interactions, mainly high wind speed, land erodibility and bareness, and humidity, among other influencing factors [12]. Studies show that ambient dustiness over dust-prone regions heavily depends on the regional drought [13,14,15,16]. Nabavi et al. [15] demonstrated that the 2007–2008 winter and spring (October to May) dust in Western Asia was associated with a severe precipitation deficit. The intermodal spread in CMIP6-simulated dust emission is explained by the differences in the models’ drought sensitivity to dust emission [17]. Earth system models heavily underestimate coarse dust [18].

Few previous studies have analyzed the predictability of western US dust. For example, Pu et al. [19] used the linear regression technique to predict the spring season’s high concentration coarse dust event frequency, based on wind speed, precipitation, and surface bareness. Aryal, Y [20] showed that machine learning models better predict fine dust than coarse dust on a monthly timescale. Temperature strongly depends on the monthly timescale, while precipitation is important to explain the long-term variability of ambient dust [20,21]. Dust emission is a complex, non-linear function of several atmospheric and surface variables [22]. Due to the limited data availability, it is often difficult to calibrate and predict the ambient dustiness based on earth system models. Therefore, understanding the predictability of regional dustiness based on the regional hydroclimate is an interesting research topic. However, much less effort has been made to compare the predictability of fine and coarse dust on different temporal scales based on the basic climatic data.

The specific objectives of this work are (1) to predict the aeolian dust using adaptive neuro-fuzzy inference systems (ANFIS) and multi-layered perceptron Artificial Neural Networks (mlp-ANN) models, and (2) to compare the models’ performance for fine dust (particle diameter ≤ 2.5 µm; PM2.5) and coarse dust (particle diameter 2.5–10 µm; PM10) on monthly and seasonal timescales.

2. Materials and Methods

2.1. Study Area and Data

The Southwestern US is a major dust source in the USA [23]. The Chihuahuan Desert, the Colorado River, and the High Plains are well-known dust sources in the region. The Interagency Monitoring of Protected Visual Environments (IMPROVE) network provides the observed near-surface dust concentration over the region [24]; available online: https://views.cira.colostate.edu/fed/Express/ImproveData.aspx (accessed on 10 April 2022) from 1988 to the present. The locations of the observation stations over the study region are shown in Figure 1. The stations that have at least 90% complete record months are chosen. Any month is considered complete if it has at least 75% of the daily observations. The total precipitation (P), average 2 m air temperature (T), and monthly average daily maximum 10 m wind speed (W) were taken from the North American Regional Reanalysis (NARR) dataset at 0.3 deg resolution [25], available online: https://psl.noaa.gov/data/gridded/data.narr.monolevel.html (accessed on 10 April 2022)), at 0.3° × 0.3° resolution. The analysis is done on a regional scale (Figure 1). Data from 1988 to 2009 are used as training data and those from 2010 to 2020 as test data. The model performance results are shown for the test data.

2.2. Artificial Intelligence (AI) Models

The performance of two widely-used AI models: an adaptive-network-based fuzzy inference system (ANFIS) and a multi-layered perceptron artificial neural network (mlp-NN) are examined. All computations are done in R [26].

2.2.1. Adaptive-Network-Based Neuro-Fuzzy Inference System (ANFIS)

The ANFIS model, proposed by Jang et al. [27], combines fuzzy logic and neural networks. The layered structure (Figure 2) in the ANFIS model adds fuzzy logic to the artificial neural networks.

The fuzzy inference system uses a hybrid learning algorithm to identify the system parameters and teach the model [28].

For three inputs (x, y, z), three rules and one output (f), the fuzzy if-then rules from Takagi and Sugeno [29] are as follows

$$\begin{gathered} 1.ifxis{A}_1,yis{B}_1,{\mathrm{and}\mathrm{z}\mathrm{is}\mathrm{C}}_{1}then{f}_1=a_{1}x+b_{1}y+c_{1}z+r_1 \\ 2.ifxis{A}_2,yis{B}_2,{\mathrm{and}\mathrm{z}\mathrm{is}\mathrm{C}}_{2}then{f}_2=a_{2}x+b_{2}y+c_{2}z+r_2 \\ 3.ifxis{A}_3,yis{B}_3,{\mathrm{and}\mathrm{z}\mathrm{is}\mathrm{C}}_{3}then{f}_3=a_{3}x+b_{3}y+c_{3}z+r_3 \end{gathered}$$

The five-layer architecture of the ANFIS [30,31,32], shown in Figure 2, is as follows:

Layer 1: This is the input layer. In this fuzzy layer, each node in the layer is the degree of membership function $\left(MFs, {\mu }_{Ai}\left(x\right)\right)$ from the input. The output of the first layer is the membership values of each input for specific MFs. The shape of the MFs can be any appropriate functions that are continuous and piecewise differentiable such as Gaussian, generalized bell-shaped, trapezoidal-shaped, and triangular-shaped functions [33]. In this study, the Gaussian MF is used, which is defined as:

$$G\left(x,c,\sigma \right)=e^{-\frac{1}{2}{\left(\frac{x-c}{\sigma }\right)}^2}$$

Layer 2: This layer multiplies the input values from the incoming signal. The AND fuzzy operator is applied to get the weight (firing strength). For example, for the first node,

$$w_i={\mu }_{Ai}\left(x\right)\ast {\mu }_{Bi}\left(x\right)\ast {\mu }_{ci}\left(x\right), i=1,2,3$$

where

$w_i$ is the firing strength of the ith rule

${\mu }_{Ai}\left(x\right)$ is the degree of membership function fuzzy sets $A_i$

${\mu }_{Bi}\left(x\right)$ is the degree of membership function fuzzy sets $B_i$

${\mu }_{ci}\left(x\right)$ is the degree of membership function fuzzy sets $C_i$

Layer 3: The circle nodes in this layer normalize the firing strength. The normalized weight is the ratio of the firing strength of ith rule to the sum of all firing strength

$$\overline{w_1}=\frac{w_1}{{{\displaystyle \sum }}_{i=1}^n{w}_i},i=1,2,\dots ,n$$

where n is the number of nodes in each layer.

Layer 4: This is a de-fuzzy layer and each node is called an adaptive node. In this layer, terms are the results of operation on input signals:

$$Z_i=\overline{w_1}{f}_1=\overline{w_1}\left(p_{1}x+q_{1}y+{\mathrm{t}}_{1}z+r_1\right)$$

where $p_1, q_1, \mathrm{and} r_1$ are the consequent parameters

Layer 5: The overall output calculated in this output layer as the summation of all incoming signals from the previous layers

$$Overalloutput={\displaystyle \sum }_{i=1}^n{\overline{w}}_i{f}_i$$

The ANFIS model is built using the R package ”FRBS” [34]. The Wang–Mendel algorithm (“WM”) is used as the learning method with “MIN”, “MAX”, and “ZADEH” to be the types of t-norm, s-norm, and implicator operators, respectively. The number of membership functions is chosen which gives the highest correlation in the test data set.

2.2.2. Multilayer Perceptron Neural Network (MLP-NN)

The mlp-ANN is the most widely used feed-forward neural network. The basic structure of the mlp-NN with two hidden layers is shown in Figure 3. A detailed explanation of the mlp-NN is given in Schalkoff [35] and Hanoon et al. [36].

The neurons in the input layer work as a buffer for the distribution of input signals to the hidden layers. Any hidden neuron $\left(j\right)$ sums up the signals from the input layers $\left(x_i\right)$ based on the strength of connection $\left(w{j}_i\right)$. Then the output $\left(yj\right)$ of any neuron $\left(j\right)$ is calculated as the function of sums up as follows:

$$yj=F\left({\displaystyle \sum }w{j}_i\ast x_i\right)$$

where function F can be a radial bias function (RBF), a hyperbolic tangent, a sigmoidal or a simple threshold function. The backpropagation and gradient descent are the most common training algorithms in a multilayer perceptron. The change in connection weights between neurons i and j are:

$$\Delta w{j}_i=\eta \delta j{x}_i$$

where $\eta$ is known as the learning rate and $\delta$ depends on whether $j$ is an input or hidden neuron.

For hidden neurons:

$$\delta j=\left(\frac{\partial f}{\partial netj}\right){\displaystyle \sum }wqj\ast {\delta }_q$$

A variation of a desired and factual output of neurons in hidden layers j which are replaced via weighted sums of δq term previously achieved for neuron q connecting to the output of j.

For the output neuron:

$$\delta j=\left(\frac{\partial f}{\partial netj}\right)\left(y{j}^t-y_j\right)$$

$netj$ is an overall weighted total of signals in the input layer. $y{j}^t$ is the goal output for neurons j.

The RSNNS R package [37] is used to fit the mlp-NN model. The model parameters are: hidden units = 10, number of training iterations = 100, initialization function = “Randomize Weights”, and learning function = “Std_Backpropagation”.

2.3. Uncertainty Analysis

The prediction uncertainty of the AI models was quantified using the d-facto, r as in Mohsenzadeh Karimi et al. [38]:

$$d-factor=\frac{d\overline{x}}{\sigma x}$$

where σ is the standard deviation and $d\overline{x}$ is the average distance between the upper and lower bands

$$d\overline{x}=\frac{1}{n}{\displaystyle \sum }_1^{n}XU-XL$$

3. Results and Discussion

The performance of the ANFIS and mlp-NN models in predicting fine dust and coarse dust on monthly and seasonal timescales is shown in Figure 4, Figure 5, Figure 6 and Figure 7 and

Table 1. Performance of ANFIS model to predict fine dust (PM2.5) for the test data (2010–2020). m = monthly timescale, s = seasonal timescale, and D = Dust.

Inputs	Output	RMSE (µg/m3)	% BIAS	Change in % BIAS
(a) P(m),T(m),W(m)	Dm	0.455	40.640
P(m),W(m)	Dm	0.561	78.90	−38.26
T(m),W(m)	Dm	0.448	55.50	−14.86	−23.40
(b) P(s), T(s), W(s)	Ds	0.308	28.304
P(s),W(s)	Ds	0.401	44.45	−16.14
T(s),W(s)	Ds	0.310	29.60	−1.30	−14.85

and

Table 2. Performance of ANFIS model to predict coarse dust (PM10) for test data (2010–2020). m = monthly timescale, s = seasonal timescale, and D = Dust.

Inputs	Output	RMSE (µg/m3)	% BIAS	Change in % BIAS
(a) P(m),T(m),W(m)	Dm	2.57	42.08
P(m),W(m)	Dm	2.62	50.57	−8.49
T(m),W(m)	Dm	2.59	42.22	−0.14	−8.35
(b) P(s), T(s), W(s)	Ds	1.96	27.82
P(s),W(s)	Ds	2.95	50.73	−22.92
T(s),W(s)	Ds	1.98	28.14	−0.32	−22.59

. The results from this case study show that the ANFIS model performs better than the mlp-NN model to predict aeolian dust on both monthly and seasonal timescales. The correlation (r), root mean square error (RMSE), and percentage bias (% BIAS) for monthly fine dust prediction (ANFIS model) is 0.7, 0.45 µg/m³, and 40.64 % respectively. The models’ performance is generally better in predicting seasonal dust than monthly dust. At seasonal timescale r, RMSE, and % BIAS (ANFIS model, fine dust) are 0.83, 0.31 µg/m³, and 28.30%, respectively. Hanoon et al. [36] also showed that AI models better predict the temperature and relative humidity on (longer) monthly timescales than (shorter) daily timescales. This might be attributed to the nonstationary spurious internal variability of the climate system that is difficult to capture by the models [39]. The results demonstrate that AI models based on the climatic data predict ambient dust with good skill. Long-term basic climatic data used as predictors in this study are more readily available than long-term land surface conditions data.

As shown in Figure 4, Figure 5, Figure 6 and Figure 7 and Table 1 and Table 2, the same month/season prediction skills are better than the month/season lead prediction skills. This highlights that ambient dustiness over the region depends more on the same month/season climate than the previous month/season climate. Pu et al. [19] also noted that spring season dust in the Western US depends more on the spring climatic and surface conditions than the winter climate and land conditions.

Comparing Figure 1 and Figure 2 with Figure 3 and Figure 4, the correlation for fine dust is better than the correlation for coarse dust on both monthly and seasonal timescales [20]. The IMPROVE stations are located on federal lands and national parks at some distance from the dust source [24]. The coarse dust has a short transport distance and air retention time. Observation stations are more likely to miss coarse dust than fine dust. Therefore, the regional climate has difficulty explaining coarse dust variability.

However, if we look at the % BIAS (comparing Table 1 and Table 2), the results of the accuracy of predicting fine dust and coarse dust are not consistent. This highlights that different statistical indices should be examined to confirm the effectiveness of the examined models [36].

The relative importance of precipitation and temperature to predict dust over the region is further examined. Previous research showed that temperature is a more important predictor, followed by precipitation, to predict monthly dust over the SWUS region [20]. This study compares the relative importance of temperature and precipitation predicting dust on monthly and seasonal timescales. Table 1 and Table 2 shows the results from the ANFIS model. The % BIAS is lower for the prediction, based on T and W, than the model, based on P and W, for predicting fine dust and coarse dust on both monthly and seasonal timescales, implying the relative importance of temperature in predicting dust over the SWUS. This is consistent with the previous study [20], on a monthly timescale. On the shorter timescale, (i.e., monthly vs. seasonal timescale), dust emission quickly responds to the warming due to the soil moisture variability in the surface soil layer. On the other hand, the impacts of precipitation on dust emissions are strongly detected on a longer timescale (i.e., annual to decadal) due to changes in vegetation cover [21]. Comparing the relative importance of precipitation and temperature, the difference between the relative importance of temperature and precipitation is larger on the monthly timescale for fine dust prediction (Table 1a, the difference in % BIAS; −23.4 vs. −14.85). For coarse dust prediction, the difference between the relative importance of temperature and precipitation is larger on a seasonal timescale (Table 2b, the difference in % BIAS; −8.35 vs. −22.59).

The results from the uncertainty analysis are given in

Table 3. d-factor for ANFIS mode.

	d-Factor
PM2.5
Monthly	0.002
Seasonal	0.00035
PM10
Monthly	0.1
Seasonal	0.067

. The d-factor of both fine and coarse dustiness is low, implying the reasonable accuracy of the model for predicting regional dustiness.

There are a few caveats in this study. First, all the results are based on the regional dustiness that does not separate locally-emitted dust and transported dust. The impacts of anthropogenic land disturbances are not accounted for. This study shows the general accuracy of the two widely used AI methods, ANFIS and mlp-NN, for aeolian dust. Future studies should focus on the use of other AI models, comprehensively analyzing the effect of parameter selection such as initialization function, number of layers, and learning algorithms, among others.

4. Conclusions

In this study, the performance of artificial intelligence (AI) models: an adaptive-network-based fuzzy inference system (ANFIS) and a multi-layered perceptron neural network (mlp-NN) are investigated to predict aeolian dust over the Southwestern US. The ambient dust data was taken from the Interagency Monitoring of Protected Visual Environments (IMPROVE) network, while the regional meteorology data (precipitation, temperature, wind speed) were retrieved from the North American Regional Reanalysis (NARR) dataset. The models’ performances for fine dust and coarse dust on monthly and seasonal timescales are compared.

AI models better predict regional dustiness on a seasonal timescale than a monthly timescale.

The ANFIS model works better than the mlp-NN model in predicting regional dustiness on both the monthly and seasonal timescales.

The ambient dustiness over the region is better predicted by the same month (or season) climatic condition than by using the previous month (or season) climatic conditions.

Compared to precipitation, the temperature is the more important predictor of regional dustiness on both the monthly and seasonal timescales. However, compared to the seasonal timescale, the difference between the relative importance of temperature and precipitation is larger on the monthly timescale for fine dust prediction and vice versa for coarse dust.