Soil-Derived Dust PM₁₀ and PM_2.5 Fractions in Southern Xinjiang, China, Using an Artificial Neural Network Model

Abstract

Soil-derived dust emissions have been widely associated with health and environmental problems and should therefore be accurately and reliably estimated and assessed. Of these emissions, the inhalable PM₁₀ and PM_2.5 are difficult to estimate. Consequently, to calculate PM₁₀ and PM_2.5 emissions from soil erosion, an approach based on an artificial neural network (ANN) model which provides a multilayered, fully connected framework that relates input parameters and outcomes was proposed in this study. Owing to the difficulty in obtaining the actual emissions of soil-derived PM₁₀ and PM_2.5 over a broad area, the PM₁₀ and PM_2.5 simulated results of the ANN model were compared with the published results simulated by the widely used wind erosion prediction system (WEPS) model. The PM₁₀ and PM_2.5 emission results, based on the WEPS, agreed well with the field data, with R² values of 0.93 and 0.97, respectively, indicating the potential for using the WEPS results as a reference for training the ANN model. The calculated r, RMSE and MAE for the results simulated by the WEPS and ANN were 0.78, 3.37 and 2.31 for PM₁₀ and 0.79, 1.40 and 0.91 for PM_2.5, respectively, throughout Southern Xinjiang. The uncertainty of the soil-derived PM₁₀ and PM_2.5 emissions at a 95% CI was (−66–106%) and (−75–108%), respectively, in 2016. The results indicated that by using parameters that affect soil erodibility, including the soil pH, soil cation exchange capacity, soil organic content, soil calcium carbonate, wind speed, precipitation and elevation as input factors, the ANN model could simulate soil-derived particle emissions in Southern Xinjiang. The results showed that when the study domain was reduced from the entire Southern Xinjiang region to its five administrative divisions, the performance of the ANN improved, producing average correlation coefficients of 0.88 and 0.87, respectively, for PM₁₀ and PM_2.5. The performances of the ANN differed by study period, with the best result obtained during the sand period (March to May) followed by the nonheating (June to October) and heating periods (November to February). Wind speed, precipitation and soil calcium carbonate were the predominant input factors affecting particle emissions from wind erosion sources. The results of this study can be used as a reference for the wind erosion prevention and soil conservation plans in Southern Xinjiang.

1. Introduction

Wind erosion has been recognized as a major environmental problem in arid and semiarid regions [1,2], with soil erosion from wind threatening air quality by emitting dust into the atmosphere. Globally, annual soil-derived, wind-blown atmospheric dust is estimated to be as high as 3000 million tons [3]. Soil-derived dust that travels long distances can be a major source of atmospheric particles [4,5]. The PM₁₀ and PM_2.5 fractions associated with dust emissions are of particular concern because they can be easily inhaled and thus cause adverse health effects such as respiratory, cardiovascular and dermatological disorders, and even death [6,7,8]. It was found that dust-related PM₁₀ emissions from the Saharan Desert were significantly correlated with hospital and daily mortality in 13 Southern European cities [9]. Researchers measured long-term (2012–2020) PM₁₀ concentrations in a city in the Sistan Basin, Eastern Iran, which is one of the dustiest arid environments globally, and found that the quality of air for 25% of the days was hazardous to people’s health [10].

Several studies have shown that soil-derived particle emissions from wind erosion sources are influenced by various meteorological factors [11,12,13], with wind and precipitation being the dominant drivers [13,14]. Dust particles, such as PM₁₀ and PM_2.5, are usually produced by a mechanism known as the saltation process, which requires the wind regime to reach a critical wind velocity (threshold wind speed) sufficient to dislodge particles from the soil surface [15,16]. Higher soil moisture from increased precipitation allows capillary forces to develop between soil grains, reinforcing soil cohesion and inhibiting wind erosion [17]. Soil-derived particulate emissions are also associated with soil surface properties [18,19]. For example, calcium carbonate (CaCO₃) and organic content (OC) can affect the amount of dust generated from the soil surface by changing the soil cohesion [20,21,22].

Various methods have been used to study the wind erosion and emission properties of particles, namely field investigations, wind tunnel experiments and model simulations (Table S1). Compared with field measurements and wind tunnel experiments, wind erosion modeling, based on field measurements and geospatial data, is a more efficient and effective method to estimate dust emissions from soil wind erosion [23]. The wind erosion prediction system (WEPS), developed by the US Department of Agriculture, is one of the most widely used models for estimating soil erosion and particle production from soil surfaces and has been extensively applied throughout the United States [24,25,26,27,28] and in countries worldwide, such as Canada [29], Argentina [30] and China [31,32,33]. In numerical studies, good agreement (i.e., R² ranging from 0.87 to 0.98) was obtained between the measured and the WEPS-based simulated erosion results [34,35]. However, this model requires detailed information related to the PM₁₀ emission process, including meteorology, hydrology, land management and soil surface erodibility data. The collection and input of such large datasets are time and cost consuming and may restrict the use of comprehensive simulation models [31,36].

With the development of big data and artificial intelligence technology, different machine learning models such as LSTM-INFO, ELM-SAMOA, RVM-IMRFO and RVFL-QANA were used in the field of soil erosion [37,38,39]. As a commonly used model, artificial neural networks (ANNs) are universal and can learn complex nonlinear relationships between important input and output factors by approximating a large class of functions with a high degree of accuracy [40,41]. It has been successfully used in the field to explain soil erosion phenomena and does not require the restrictive assumptions of the WEPS model [42,43]. In a complex system of soil erosion, a full understanding of the erosion regime and transport process is challenging, although the difficulties in simulating such complex phenomena could be significantly reduced through the application of ANN models [2]. However, as shown in Table S2, studies related to ANN-based soil erosion have mainly focused on simulating the total soil loss [44,45,46] with few studies using ANN models for soil-derived dust emissions related to PM₁₀ and PM_2.5 fractions. This is probably due to the difficulty in obtaining, through field measurements, data on actual emissions of this particulate matter at a regional scale. Because of this challenge, an approach was proposed for estimating soil-derived particle emissions by using an ANN model and comparing its results with the WEPS results obtained from our prior work [33]; this was verified by the PM emissions data collected from field measurements undertaken during this study. The use of ANNs in estimating soil-derived PM₁₀ and PM_2.5 emissions will benefit governments in the implementation of appropriate control measures to improve regional air quality and protect public health in arid and semiarid areas.

The aim of this study was to select, optimize and apply an ANN method to estimate soil-derived particles (PM₁₀ and PM_2.5) from wind erosion sources. The ANN-estimated emissions from the wind-erosion sources were compared with the results obtained from the WEPS model. A further aim was to specify the importance of factors affecting soil erosion via a sensitivity analysis technique (ANN and Monte Carlo combined model). The significance of our study demonstrated the successful use of an ANN combined with the WEPS model to simulate soil-derived PM₁₀ and PM_2.5 when the selected input parameters are available; this avoids the substantial effort required to obtain particle emissions data at a regional scale through field measurements.

2. Materials and Methods

2.1. Study Area

The study area, Southern Xinjiang (73.4–88.2° E, 34.2–42.3° N), is approximately 1.06 million km² and comprises five administrative divisions (Kashi, Kezhou, Hetian, Bazhou and Akesu; Figure 1). A semiarid or desert climate prevails in Southern Xinjiang, which has substantial seasonal differences in temperature, with hot summers and cold winters. The annual average temperature is 12.4 °C, with mean temperatures in summer and winter of 25.1 °C and −3.3 °C, respectively. The annual average precipitation is 59.1 mm. The average wind speed at a 10 m height is 1.8 m/s; the highest wind speed (2.15 m/s) is in spring, followed by summer (2.14 m/s), autumn (1.44 m/s) and winter (1.29 m/s). The Taklamakan Desert, which is unique and serves as the main wind erosion source of particulate matter, is situated in the middle of Southern Xinjiang and is the second largest drifting desert globally. In 2016, daily concentrations of PM₁₀ (PM_2.5) averaged 435 μg/m³ (158 μg/m³), 211 μg/m³ (77 μg/m³), 285 μg/m³ (106 μg/m³), 103 μg/m³ (42 μg/m³) and 243 μg/m³ (92 μg/m³) for Kashi, Kezhou, Hetian, Bazhou and Akesu, respectively (China National Environmental Monitoring Center, http://www.cnemc.cn/, accessed on 3 October 2023). These average daily concentrations of PM₁₀ and PM_2.5 in Southern Xinjiang exceeded the National Ambient Air Quality Standard of China (GB3095-2012), (https://www.transportpolicy.net/standard/china-air-quality-standards/, accessed on 3 October 2023), (daily average of 70 μg/m³ for PM₁₀ and 35 μg/m³ for PM_2.5) by 3.65 and 2.71 times, respectively.

2.2. Data Collection and Soil Properties Analysis

A total of 264 representative soil samples were collected from the study area. As shown in Figure 1, the samples were evenly distributed around the Taklamakan Desert because obtaining samples from the desert hinterland was difficult. The samples were collected from the top 10 cm of the soil surface by cutting rings, which were then transported to the laboratory for analysis of their characteristics, including their pH, CaCO₃, cation exchange capacity (CEC) and OC. These factors were selected because they have a strong influence on wind erosion [47,48].

Two meteorological parameters including wind speed (WS) at a 10 m height and precipitation (Pre) were collected from the American National Center for Atmospheric Research (https://ncar.ucar.edu/, accessed on 3 October 2023), which were then downscaled by the Weather Research and Forecasting model resampled to a resolution of 10 km. An elevation (Ele) map of the study area was generated by using a digital elevation model from the Geospatial Data Cloud (http://www.gscloud.cn/, accessed on 3 October 2023).

2.3. Artificial Neural Network Model

The structure of the ANN model used in this study is shown in Figure 2. The total study area was gridded into 17,576 data cells with a resolution of 10 km × 10 km. The cells were divided randomly into two subsets: 70% of data cells for training and the remaining 30% for testing. First, the training datasets were used to train the model to obtain the optimal weight and network structure. Next, the determined optimal ANN model structure and input factors in the testing datasets were used to simulate the soil-derived particles in areas without erosion (the output in the testing datasets). To evaluate the performance of the ANN in simulating soil-derived particles from wind erosion in areas with unknown erosion amounts, the ANN’s simulated values were compared with the corresponding reference values obtained by the WEPS published in our previous study [33].

In addition, the study year was divided into three periods: the sand period (March–May), nonheating period (June–October) and heating period (November–February) [49,50]. Data from the last month in each period were used as the testing dataset, with data from the other months being used as the training dataset. The values predicted by the ANN in May, October and February were further compared with their corresponding reference values, obtained through the WEPS, in terms of evaluating the model performance of the ANN in predicting soil-derived particles from wind erosion.

As shown in Figure 2, the network structure was designed in the form of 7-5-1: one input layer with seven neurons, one hidden layer with five neurons and one output layer that represented the soil-derived dust emissions related to PM₁₀ or PM_2.5. The seven neurons were the soil pH, soil CaCO₃, soil CEC, soil OC, WS, Pre and elevation. Previous studies have reported that an ANN with one hidden layer can estimate any nonlinear problem with high accuracy [51,52]; therefore, one hidden layer was set for our model. The number of neurons in the hidden layer was determined based on the two-thirds rule [53]; i.e., two-thirds of the number of neurons in the input layer (7 × 2/3 ≈ 5). The normalized input factors were applied to two commonly used activation functions, namely the logarithmic sigmoid function and hyperbolic tangent function, to introduce nonlinearity into the output. Detailed information is shown in the section headed “Principal of artificial neural network model” in the Supplementary Materials. The ANN model analysis was conducted by using the “neuralnet” package in R software RoxygenNote 6.1.0 (https://cran.r-project.org/web/packages/neuralnet/neuralnet.pdf, accessed on 3 October 2023).

2.4. Field Experiment

Because the results obtained by the WEPS were used as a reference for training the ANN model, these WEPS results had to be evaluated by using field measurement data from the selected study area. To measure the wind erosion flux of the particles between October 2016 and March 2017, field experiments were conducted by using a Sensit (Redlands, CA, USA) H14-LIN wind erosion mass flux sensor at a national field observation site (80°43′45″ E, 37°00′57″ N) in the Hetian administrative division. The erosion sensor was installed at a height of 2.5 m above the ground, and daily average particle concentrations with sizes ranging from 50 to 70 μm were measured by the instrument. The emissions of PM₁₀ and PM_2.5 were calculated based on the following equations:

$$E_{PM10}=\alpha E_{PM50-70}$$

(1)

$$E_{PM2.5}=\beta E_{PM50-70}$$

(2)

where E_PM10, E_PM2.5 and E_PM50–70 represent the amount of soil-derived particles with sizes less than or equal to 10 μm; 2.5 μm; and the amount of eroded particles collected by the sensor, with particle sizes ranging from 50 to 70 μm, respectively. $\alpha$ and $\beta$ are constant factors, which were set as 0.11 and 0.03, respectively.

The constant factors were calculated based on the information in Table S3, which was obtained from a study which analyzed the particle size distribution from sand dust storms in the Taklamakan Desert [54]. The monthly average values of the PM emissions were calculated and compared with the WEPS-simulated monthly values in the cell where the observation site was located.

2.5. Modeling Validation

The predictive ability of the ANN model for estimating soil-derived particles was examined based on several statistical indexes, including the determination coefficient (R²), Pearson correlation coefficient (r), root mean square error (RMSE) and mean absolute error (MAE). The equations of these indexes are shown in the section “Model validation” in the Supplementary Materials. The closer the RMSE and MAE values are to zero, the more accurate the simulations are. When the absolute value of r is in the range of 0.7 to 1, a strong correlation is obtained.

2.6. Uncertainty Analysis and Sensitivity Analysis

The uncertainty analysis of the amount of soil-derived PM₁₀ and PM_2.5, obtained by the ANN, was conducted by using the Monte Carlo method [55,56]. The distributions and values of the parameters of these input factors are listed in

Table 1. Distributions for input factors, Spearman correlation coefficients between input factors and PM₁₀ emissions and PM₁₀ ranges using the optimum ANN model.

Input Factors	Distribution	Values of Parameters	Correlation with PM10 Emissions per Unit Area (R2)	ΔPM10 (t/km2)
pH	Normal	Mean = 7.82; Sd = 0.27	−0.04 **	8.08
CaCO3	Normal	Mean = 0.048; Sd = 0.038	−0.23 **	11.13
CEC	Log normal	Geometric Mean = 16.67; Geometric Sd = 1.15	−0.09 **	6.15
OC	Log normal	Geometric Mean = 0.0099; Geometric Sd = 1.16	−0.04 **	5.25
WS	Weibull	Shape = 6.29; Scale = 2.21	0.36 **	14.86
Pre	Exponential	Rate = 0.0016	−0.20 **	42.03
Ele	Generalized extreme value	Location = 9.14; Scale = 3.04; Shape = −0.04	0.04 **	1.06

**: p < 0.01.

. A total of 100,000 samples, without replacement for each input factor, were resampled from the fitted probability distribution. Seven blocks of 100,000 input vectors were used to generate 100,000 values of PM₁₀ (or PM_2.5) emissions (t/km²) by using the optimal structure of the ANN. The uncertainty propagation was calculated by using the following equation:

$$\mathrm{U}\mathrm{n}\mathrm{c}\mathrm{e}\mathrm{r}\mathrm{t}\mathrm{a}\mathrm{i}\mathrm{n}\mathrm{t}\mathrm{y}=(\frac{{\mathrm{V}}_{2.5\mathrm{\%}}-\mathrm{M}\mathrm{e}\mathrm{a}\mathrm{n}}{\mathrm{M}\mathrm{e}\mathrm{a}\mathrm{n}}\times 100\mathrm{\%}~\frac{{\mathrm{V}}_{97.5\mathrm{\%}}-\mathrm{M}\mathrm{e}\mathrm{a}\mathrm{n}}{\mathrm{M}\mathrm{e}\mathrm{a}\mathrm{n}}\times 100\mathrm{\%})$$

(3)

where V_2.5% and V_97.5% are the 2.5th and 97.5th percentile value, respectively, of the generated PM amount. The narrower the 95% confidence interval (CI) obtained by the above equation, the greater the accuracy of the prediction.

The sensitivity analysis was performed to probe the response of the soil-derived particle emissions to independent changes in the soil parameters and meteorological data [57]. Detailed information on the sensitivity analysis is shown in the “Sensitivity analysis” section in the Supplementary Materials. In addition, a Spearman analysis was used to verify the effect of each factor on soil-derived particle emissions.

3. Results

3.1. Verifying PM Emissions of WEPS Based on Field Experiments

During the study, the monthly average values of the PM emissions per unit area were calculated. As shown in Figure 3, the PM₁₀ and PM_2.5 emission results based on the WEPS agreed well with the field data, with R² values of 0.93 and 0.97, respectively. The absolute values of the RMSE were 0.34 and 0.18 for PM₁₀ and PM_2.5, respectively. Based on this comparison, the emission results obtained by the WEPS were deemed reliable and acceptable. Thus, using the simulation data obtained by the WEPS as a reference for training the ANN model is possible.

3.2. PM Emissions from Wind Erosion Sources Based on ANN

After network training and optimization, it was found that the model performance with the logarithmic sigmoid function could provide better results than that with the hyperbolic tangent function for both PM₁₀ and PM_2.5 (

Table 2. Performance of ANN model using 7-5-1 network structure with different activation functions to predict soil-derived particles from wind erosion in Southern Xinjiang.

District	Active Function	Structure	PM10				PM2.5
			R2	r	RMSE	MAE	R2	r	RMSE	MAE
Kashi	Logarithmic sigmoid	7-5-1	0.75	0.87	0.17	0.09	0.71	0.84	0.11	0.05
	Tangent sigmoid	7-5-1	0.65	0.80	0.24	0.12	0.63	0.79	0.11	0.06
Kezhou	Logarithmic sigmoid	7-5-1	0.92	0.96	2.88	1.73	0.93	0.97	0.51	0.32
	Tangent sigmoid	7-5-1	0.87	0.93	3.92	2.74	0.94	0.97	0.43	0.24
Hetian	Logarithmic sigmoid	7-5-1	0.80	0. 90	2.89	1.98	0.80	0.89	1.21	0.83
	Tangent sigmoid	7-5-1	0.83	0.91	2.78	1.89	0.76	0.87	1.35	0.91
Bazhou	Logarithmic sigmoid	7-5-1	0.74	0.86	1.58	1.09	0.64	0. 80	0.49	0.25
	Tangent sigmoid	7-5-1	0.72	0.85	1.62	1.11	0.56	0.70	0.62	0.29
Akesu	Logarithmic sigmoid	7-5-1	0.71	0.84	3.00	2.06	0.72	0.85	1.31	0.91
	Tangent sigmoid	7-5-1	0.66	0.81	3.28	2.04	0.76	0.87	1.18	0.79

). Thus, the optimized network structure, which is 7-5-1 with a logarithmic sigmoid function as the activation function, was applied to different administrative divisions and study periods. Test data with unknown erosion amounts were selected randomly by the ANN (30% of the total data) to test the model performance of the ANN in simulating the dust emissions for the study area. Figure 4 shows the violin chart, which compares the results obtained by the ANN by using the optimized network structure and WEPS. It was found that the distribution of the soil-derived dust emissions simulated by the WEPS is compatible with the distribution obtained by the ANN regardless of particle size. As shown in

Table 3. Performance of ANN model using optimal network architecture (7-5-1 network structure with logarithmic sigmoid function) to predict soil-derived particles from wind erosion in area with unknown emissions in Southern Xinjiang during different periods.

District	Period	PM10				PM2.5
		R2	r	RMSE	MAE	R2	r	RMSE	MAE
Kashi	Sand	0.68	0.83	0.17	0.10	0.74	0.86	0.07	0.05
	Heating	0.69	0.83	0.09	0.06	0.68	0.82	0.07	0.04
	Nonheating	0.74	0.86	0.15	0.11	0.74	0.86	0.10	0.06
	Whole year	0.75	0.87	0.17	0.09	0.71	0.84	0.11	0.05
Kezhou	Sand	0.79	0.89	2.33	1.18	0. 90	0.95	0.26	0.16
	Heating	0.77	0.88	0.90	0.53	0.82	0.91	0.12	0.07
	Nonheating	0.82	0.91	1.95	0.99	0.88	0.94	0.25	0.14
	Whole year	0.92	0.96	2.88	1.73	0.93	0.97	0.51	0.32
Hetian	Sand	0.81	0.90	1.49	1.03	0.79	0.89	0.63	0.43
	Heating	0.55	0.74	0.74	0.52	0.57	0.76	0.26	0.20
	Nonheating	0.78	0.88	1.45	1.00	0.76	0.87	0.62	0.40
	Whole year	0.80	0.90	2.89	1.98	0.80	0.89	1.21	0.83
Bazhou	Sand	0.62	0.79	0.99	0.53	0.65	0.81	0.21	0.12
	Heating	0.74	0.86	0.61	0.43	0.73	0.86	0.14	0.11
	Nonheating	0.70	0.84	0.59	0.41	0.67	0.82	0.15	0.09
	Whole year	0.74	0.86	1.58	1.09	0.64	0.80	0.49	0.25
Akesu	Sand	0.64	0.80	1.87	1.18	0.66	0.81	0.82	0.57
	Heating	0.71	0.84	0.69	0.47	0.67	0.82	0.33	0.23
	Nonheating	0.67	0.82	0.90	0.61	0.72	0.85	0.38	0.25
	Whole year	0.71	0.84	3.00	2.06	0.72	0.85	1.31	0.91
Southern Xinjiang	Sand	0.53	0.73	1.86	1.32	0.62	0.79	0.76	0.49
	Heating	0.59	0.77	0.77	0.53	0.57	0.75	0.27	0.18
	Nonheating	0.66	0.81	1.29	0.82	0.66	0.81	0.52	0.30
	Whole year	0.61	0.78	3.37	2.31	0.62	0.79	1.40	0.91

, the values of R² ranged from 0.53 to 0.66 for PM₁₀ and 0.57 to 0.66 for PM_2.5 throughout Southern Xinjiang, indicating that the model explains 53–66% (57–66%) of the variation in the soil-derived PM₁₀ (PM_2.5) around its mean. The values of R were approximately equivalent to 0.8 regardless of the particle size (see Figure S1 in the Supplementary Materials). It was also found that R² increased when the study domain was reduced. In Kezhou, the maximum R² reached 0.92 and 0.93 for PM₁₀ and PM_2.5, respectively, during the study period. The value of the correlation coefficient ranged from 0.73 to 0.97, indicating a strong positive linear relationship between the results simulated by the ANN and WEPS. In summary, the results indicated that the performance of the ANN model in estimating PM emissions for the reduced study area were more reliable than those for the entire area of Southern Xinjiang.

Moreover,

Table 4. Performance of ANN model by using optimal network architecture (7-5-1 network structure with logarithmic sigmoid function) to predict soil-derived particles from wind erosion in the last month during each period in Southern Xinjiang.

District	Period	PM10				PM2.5
		R2	r	RMSE	MAE	R2	r	RMSE	MAE
Kashi	Sand	0.57	0.76	1.67	1.66	0.64	0.80	0.80	0.79
	Heating	0.53	0.73	0.06	0.04	0.55	0.74	0.03	0.02
	Nonheating	0.55	0.74	0.08	0.03	0.58	−0.76	0.11	0.09
Kezhou	Sand	0.95	0.97	0.92	0.81	0.91	0.95	0.16	0.14
	Heating	0.68	0.82	0.06	0.04	0.64	0.80	0.05	0.03
	Nonheating	0.90	0.95	1.18	0.73	0.88	0.94	0.20	0.13
Hetian	Sand	0.76	0.87	0.72	0.59	0.74	0.86	0.30	0.25
	Heating	0.61	0.78	0.17	0.10	0.52	0.72	0.08	0.06
	Nonheating	0.68	0.83	0.65	0.25	0.59	0.76	0.38	0.25
Bazhou	Sand	0.79	0.89	0.75	0.70	0.69	0.72	0.07	0.05
	Heating	0.68	0.82	0.28	0.20	0.52	0.83	0.19	0.17
	Nonheating	0.72	0.84	0.52	0.30	0.47	0.68	0.13	0.08
Akesu	Sand	0.71	0.84	0.90	0.79	0.74	0.86	0.41	0.36
	Heating	0.56	0.75	0.43	0.31	0.57	0.75	0.20	0.14
	Nonheating	0.63	0.79	0.59	0.43	0.69	0.83	0.27	0.19

shows the performance of the ANN model by using the optimal network architecture to predict soil-derived dust emissions from wind erosion sources in the last month for a period. The results showed that the predictivity of PM emissions by the ANN in the sand and nonheating periods was better than that in the heating period, because a higher R² and r and lower RMSE and MAE were obtained for the sand and nonheating periods than the values obtained for the heating period.

3.3. Uncertainty Analysis

The uncertainty analysis of the simulated PM emissions from wind erosion sources in Southern Xinjiang was conducted by estimating the 95% CI of the simulated results. As shown in

Table 5. Uncertainty analysis of PM emissions from wind erosion sources in Southern Xinjiang.

Particles	Period	Mean	2.5%	97.5%	Uncertainty
PM10	Sand	12.14	3.66	20.46	(−70%, 68%)
	Heating	3.34	0.98	7.88	(−71%, 136%)
	Nonheating	3.72	1.03	7.99	(−72%, 115%)
	Whole year	15.15	5.20	31.27	(−66%, 106%)
PM2.5	Sand	1.67	0.41	3.82	(−76%, 127%)
	Heating	0.62	0.16	1.34	(−74%, 115%)
	Nonheating	0.99	0.28	2.29	(−72%, 132%)
	Whole year	1.83	0.45	3.82	(−75%, 108%)

, the uncertainty of the soil-derived PM₁₀ emissions at a 95% CI was (−66–106%) in 2016. The uncertainty of PM₁₀ during different periods was (−70–68%), (−71–136%) and (−72–115%) for the sand, heating and nonheating periods, respectively. The uncertainties for PM₁₀ were lower than the uncertainties for PM_2.5, indicating that the ANN model was more accurate in simulating the emission of PM₁₀ than that of PM_2.5. Figure 5 shows the fitted probability distribution function for the PM₁₀ emission inventory, which indicates that the PM₁₀ emissions from wind erosion sources followed a log normal distribution. All the PM₁₀ values were within the 95% CI regardless of the study period, confirming the robustness of the ANN model [57].

3.4. Sensitivity Analysis

Because of the high uncertainty of PM_2.5, a sensitivity analysis was only conducted on PM₁₀. The obtained PM₁₀ emission range and the results of the Spearman analysis are presented in Table 1. The value of ΔPM₁₀ was the highest (42.03 t/km²) for Pre, followed by WS (14.86 t/km²) and CaCO₃ (11.13 t/km²), indicating that the Pre, WS and CaCO₃ of the soil were the most important factors for soil-derived PM₁₀ emissions in Southern Xinjiang. The contributions of the pH, CEC and OC of the soil were relatively low, with the contribution of elevation being the lowest. The relationship between PM₁₀ emissions and the influencing factors was analyzed by estimating the Spearman correlation coefficients between them. The negative value of R² (R² = −0.20) showed that with increasing precipitation, PM₁₀ emissions decreased accordingly. Simulated PM emissions increased with an increase in the WS (R² = 0.36). The growth of the soil CaCO₃ and OC resulted in a significant decrease in PM₁₀ emissions.

4. Discussion

A selected ANN model was used to simulate PM₁₀ and PM_2.5 from a wind erosion source by using relevant factors, including soil properties and meteorological parameters as input variables. The results indicated that an ANN can be used as an alternative tool to simulate soil-derived particle emissions provoked by wind erosion in areas with similar meteorological conditions and soil properties. The determined optimal network structure was used to simulate the emissions of PM₁₀ and PM_2.5 in an area with unknown emission amounts. The high correlation coefficients and low errors in five administrative divisions in Southern Xinjiang indicated that the model was able to simulate PM emissions for places without erosion data. The effectiveness of machine learning models in terms of solving the soil erosion problem was also reported by a study which evaluated dust emissions in the Jazmurian Basin in Southeastern Iran by using six algorithms [58]. All the models yielded a high performance with R > 0.9 and RMSE < 20%. It is also found that for regions such as the entire area of Southern Xinjiang, the performance of the ANN requires further refinement. This phenomenon can be explained by similar meteorological conditions and similar soil properties in a relatively small study area. The contribution of each input parameter to soil-derived dust emissions from wind erosion sources differs by administrative division [59,60]. An ANN was also used to predict the soil-derived particles for the last month of each study period. The ANN model performed better in the sand and nonheating periods compared with that of the heating period. This is probably because the soil samples were collected during the sand and nonheating periods, so the properties of the soil used in the modeling process represented the soil conditions in the sand and nonheating periods.

The Monte Carlo technique was used to examine the uncertainty of the ANN model in simulating PM emissions from soil erosion. The uncertainty analysis also confirmed that the ANN model could properly simulate PM₁₀ emissions from soil in Southern Xinjiang because all the PM₁₀ values fell within the 95% confidence interval. It is found that the ANN model was more accurate in simulating the emission of PM₁₀ than that of PM_2.5, probably because the PM_2.5 emission inventory was obtained by multiplying the PM₁₀ emissions by factors obtained from the wind tunnel experiments. This process may introduce additional uncertainties to the ANN model. To the best of our knowledge, limited studies were conducted on estimating soil-derived particles (PM₁₀ and PM_2.5) by using an ANN. Most studies reported that an ANN can be applied to the modeling of soil erosion rates, which measure the amount of soil mass lost from a unit area. For example, Kouchami-Sardoo et al. studied the performance of an ANN on predicting the soil erosion rate in Kerman province, Southeast Iran, by using factors such as the bulk density, soil organic matter and soil moisture [46]; they found a high correlation (R² = 0.94, MSE = 0.04) between the erosion rates measured in the wind tunnel and the values simulated by the ANN. Kim et al. studied the soil erosion rate at a farm in Lincoln, USA, by using an ANN model [52]. By comparing the field measurement values, R² and RMSE were reported as 0.62 and 0.093, respectively. Egbueri et al. evaluated the performance of the ANN for forecasting soil erodibility potentials in Southern Anambra, and they found that the R² for an ANN model ranged from 0.971 to 0.998 and the sum of residual errors ranged from 0.000 to 0.024 [61]. Thus, given the fact that ANNs can decrease the estimated time and costs, the use of this model in simulating soil erosion is recommended [62].

Although the accuracy of ANN-based simulation results is high, the ANN’s inability to interpret the physical influences of the input parameters on soil erosion is one major disadvantage of ANN-based studies. The combined use of an ANN and Monte Carlo simulation was thus conducted to analyze the sensitivity of each variable to PM emissions, allowing us to assess the importance of each component [46]. Statistical analyses showed that the Pre, WS and CaCO₃ of the soil were the three major factors affecting soil erosion. Similarly, studies reported that rainfall was the most important factor in the soil-erosion process [3,58]. The study year had the highest recorded rainfall since 1961 at 247.6 mm. The increased soil moisture due to rainfall creates substantial interparticle forces and increases aggregation between soil particles, which inhibits soil erosion and dust emission [3]. It was found that the influence of moisture on PM₁₀ emissions decreases logarithmically [63]. Wind speed is the most direct power source and a leading factor in soil erosion [64]. Numerous studies have found that PM₁₀ emissions from soil increased significantly with an increase in WS, regardless of soil type [65,66,67]. Coupled with the special topography of the Tarim Basin in Southern Xinjiang, considerable wind shear could be generated because of the Kunlun Mountain block along the southern edge of the Xinjiang region, producing advantageous conditions for high soil-derived dust emissions [33]. A high WS contributes to soil erosion and increases the emission of particles, which can explain why the emissions of PM₁₀ and PM_2.5 were the highest during the sand period (throughout Southern Xinjiang, the sand period had the highest average WS of 2.37 m/s). The calcium carbonate content and organic matter of soil are typical cementing materials for aggregation in soils, which limit the emission of particles by wind erosion [68]. Soil aggregation produces a high demand for water, which can fill the intra-aggregate pores and bond the dust particles, inhibiting PM emissions [69]. It was found that calcium carbonate in soil could increase the content of aggregates, which decreases the potential for soil erosion by wind [70]. It was also reported that PM₁₀ emissions were low because of high OC in soils [63].

The limitation of this study is the use of simulation data obtained by the WEPS for training the ANN model. In the absence of validated observational field measurement data, the data produced by globally available WEPS model simulations were used as the data for comparing with the ANN model. The error of the results simulated by the WEPS was not considered. Although field data are the most representative and best data to use to validate the performance of an ANN model, dust emission data with a high spatial resolution were seldom obtained through field experiments in Southern Xinjiang. To mitigate the limitation, a field measurement was carried out at a single sampling site to verify the accuracy of the results simulated by the WEPS. Further field experiments at a regional level, which require a relatively high spatial resolution, are needed to measure actual soil-derived dust emissions related to PM₁₀ and PM_2.5. Thus, the efficiency of an ANN in simulating PM emissions from wind erosion sources can be verified directly by measuring actual field data.

5. Conclusions

Our study evaluated the performance of an ANN model in simulating soil-derived dust emissions related to PM₁₀ and PM_2.5 fractions in Southern Xinjiang, which includes the second largest drifting desert worldwide. By comparing its results with well-established modeling data obtained by the WEPS model, it was concluded that the ANN model provided acceptable results for highly nonlinear problems such as a soil erodibility analysis, which is complex and has many influencing factors. The calculated r, RMSE and MAE for the results simulated by the WEPS and ANN were 0.78, 3.37 and 2.31 for PM₁₀, respectively. The corresponding values were 0.79, 1.40 and 0.91 for PM_2.5, respectively. The results also indicated that when the study domain was reduced from the entire Southern Xinjiang region to its five administrative divisions, the performance of the ANN improved, producing average correlation coefficients for the comparison of 0.88 and 0.87, respectively, for PM₁₀ and PM_2.5. This phenomenon is probably due to the strength of each input parameter on PM₁₀ emissions being different in different administrative divisions. The best performance of the ANN model in simulating PM emissions was achieved during the sand period; this was followed by the nonheating and heating periods. The Monte Carlo uncertainty analysis indicated the robustness of the ANN model because all the PM₁₀ emission values fell within the 95% CI regardless of the study period or particle size. Using an ANN to estimate PM₁₀ and PM_2.5 emissions due to soil erosion has several advantages; for example, the model requires less input parameters to estimate soil-derived particles as it is flexible and can be updated in terms of different structures to improve the estimation results. A sensitivity analysis using an ANN and Monte Carlo combined model revealed that the WS, Pre and CaCO₃ in the soil were the most important factors influencing soil-derived particle emissions. The limitation of this study was that we only compared the soil-derived dust emission results obtained by the WEPS with data collected at one available field measurement site. For future related investigations, it is recommended to increase the number of locations where measurement data are collected for use in verifying the results obtained by an ANN. The method provided in this study can be used as a tool for the prevention of wind erosion and for soil conservation plans.