1. Introduction
Dust storm is a natural disaster that causes enormous damage to the environment, transportation, and crops [
1]. It also has direct and indirect impacts on public health and human health [
2,
3]. Small volumes of dust particles can enter the respiratory tract and lungs, causing asthma, pneumonia, and other respiratory and cardiovascular diseases [
4,
5]. Studies have shown that exposure to dust particles in the air for hours to days can lead to other problems such as conjunctivitis, meningitis, and valley fever [
6,
7,
8,
9]. There are currently four major dust source areas and dust-prone zones in the world, located in northern Africa, eastern Asia, North America, and Australia [
10]. The Taklamakan and Gobi Deserts in northwestern China are the main dust source in East Asia and also the second largest dust source in the world [
10]. Dust storm episodes typically occur in the spring (March–May) [
11,
12] and cause great damage to the economy and people’s lives. Therefore, there are economic, social, and scientific benefits from monitoring dust storms and understanding the transport, spatial, and temporal distribution of dust [
13,
14].
Dust monitoring can be achieved by ground-based observations and satellite remote sensing [
15,
16]. Data from ground-based air quality monitoring stations are often used to study dust events [
17]. However, due to the limited number of ground stations, the dynamic monitoring of dust transport is challenging based on ground-based observations. Dust mainly occurs in desert areas where meteorological stations are not easy to set up and maintain [
18]. With the development of meteorological satellite technology, remote sensing images with a high observation frequency and wide spatial coverage play an essential role in dust detection and monitoring [
19]. Satellite remote sensing can not only determine the dust source, transport path, impact range, and intensity of dust events but also monitor the dynamics of dust formation, development, and spread [
20].
Theoretically, airborne dust can be detected with different optical properties in the ultraviolet (UV), visible (VIS), and infrared (IR) satellite images [
21]. Dust detection methods can be generally divided into three categories depending on the channels used. The first category is based on the absorbability of dust aerosols at shorter wavelengths [
22]. Typical examples are the absorbing aerosol index from Ozone Monitoring Instrument (OMI) [
22,
23] and the normalized difference dust index (NDDI) proposed by Qu et al. [
24]. Second, methods based on the brightness temperature (BT) in thermal infrared (TIR) channels are more widely used due to the capability to detect dust over bright surfaces and at night [
25,
26]. Among them, the brightness temperature difference (BTD) between two IR channels [
27,
28] and infrared difference dust index (IDDI) [
29] are the two classic dust detection methods. According to Ackerman [
27], the BTD between 8.5, 11, and 12 μm of Moderate Resolution Imaging Spectroradiometer (MODIS) data is capable of detecting dust aerosols. Negative BTD between 11 μm and 12 μm (BTD (11–12 μm)) typically indicate dust presence. IDDI is more often applied to geostationary satellites and has also been used in the operational dust monitoring system in China [
30]. Based on the above two types of methods, the combination of shortwave reflectance and infrared brightness temperatures to detect dust have also been developed [
31,
32,
33].
Table 1 lists several representative indices for dust detection based on the reflectivity and emissivity of airborne dust at multiple channels. However, the ability of these methods are influenced by various factors such as sensor spectral channel configurations and the target area’s surface reflection and thermal characteristics. In recent years, machine learning has been increasingly used to detect dust with satellite data [
34,
35,
36]. Due to the powerful learning capability of machine learning, it has shown high accuracy in regional dust monitoring. The machine learning approach, however, has some drawbacks for detecting dust such as difficulty for large area and near-real-time applications due to the computation requirements and easy over-fitting, especially in the presence of redundant features [
37].
Compared to polar-orbiting satellites, the geostationary satellites have a higher frequency of observations, making them more suitable for dust detection [
16,
43]. Geostationary satellites used to study dust monitoring for China include the Multifunctional Transport Satellite (MTSAT-2) and Himawari 8 launched by Japan [
12] and Feng Yun-2 (FY-2) [
30] and Feng Yun-4 (FY-4) [
44] launched by China. Himawari-8 and FY-4A, as the new generation of geostationary satellites, are superior to MTSAT-2 and FY-2 in terms of temporal resolution, spatial resolution, and the number of measurement channels. Compared to Himawari-8, FY-4A is more appropriate for dust storm monitoring over China, with a broad observation range covering the entirety of China. Jiang et al. [
44] conducted FY-4A-based dust monitoring based on machine learning and found that machine learning has the advantage in identifying dust over specific areas. Gao et al. [
45] analyzed three dust events that occurred in 2021 with FY-4A data. Additionally, the China Meteorological Administration (CMA) developed two dust products based on FY-4A data including the dust score product (DST) and dust intensity product (IDDI_BK). Wu et al. [
46] found that these two FY-4A L2 dust detection products in the northwest region agreed well with the MODIS AOD products. All of these studies demonstrate that FY-4A has good potential for dust detection.
So far, few studies have comprehensively and quantitatively evaluated the performance of different dust detection methods based on FY-4A for dust detection over China. Therefore, this study aimed to evaluate the ability of the BTD, NDDI, CMA IDDI, and DST algorithms/products to detect dust. Focusing on dust detection over China based on FY-4A data, this study:
- (1)
Qualitatively evaluated the performance of the four algorithms/products on the dust identification during two typical dust events;
- (2)
Quantitatively compared the BTD, NDDI, and two FY-4A dust products with the real-time ground-based PM10 (less than 10μm in aerodynamic diameter) concentration data and assess their accuracy in identifying dust in the spring of 2021.
5. Discussion
All POFDs for the four dust detection algorithms were high, ranging between 71.83% (DST was the lowest at 02:00 UTC) and 96.85% (NDDI was the highest at 06:00 UTC). However, effective dust detection is still a major challenge. The first reason for high POFDs for the four algorithms is that most of the PM10 stations are distributed in central and eastern China while in the dust-prone northwestern region, PM10 stations are relatively few. Consequently, low YY values lead to high POFD. Second, PM10 concentrations greater than 600 µg·m−3 were set as the threshold value for identifying dust, which may result in high NY values and therefore high POFD. Additionally, the high POFD for NDDI was due to its failure to distinguish between bare ground from dust over the Tibetan Plateau region and the Taklamakan Desert and to misclassify dark ground surfaces in the northeast and red vegetation in the southwest as dust. We also found that from 02:00 UTC to 08:00 UTC, the POCD for all algorithms increased before 06:00 UTC and then decreased after 06:00 UCT, with a generally higher POCD at 06:00 UTC, which may be related to the solar angle.
The POCDs for the four dust detection algorithms were relatively low partly due to discrepancies between the satellite dust identification results with the ground-based PM10 observations. First, the satellite seemed to detect severe dust events with a wide spatial coverage, while some high PM10 concentrations may not represent dust events. PM10 concentration in the urban environment is significantly affected by local sources. Local anthropogenic sources also affect the atmospheric PM10 concentration. For example, some studies have shown that traffic [
60,
61] and road dust resuspension [
62] are the main sources of PM10 in urban areas. Satellite, however, is much better at detecting dust events in the ambient areas of dust sources where dust concentrations are relatively high [
63]. In this study, most of the used sites are located in eastern China, which could increase the influence of anthropogenic dusts. Second, unavoidable mismatch could also lead to differences in dust identification using two approaches. This is because satellites are not able to detect dust under clouds from the top of atmosphere, and FY-4A, with a spatial resolution of 4 km, seems to be impossible to detect small local dust events.
Finally, we also tried to combine the BTD and IDDI algorithms for accuracy verification. The combined algorithms could reduce the misjudgment of bare land, similar to the Qinghai Tibet Plateau. As the result shows, the POFD for the fused algorithm was lower than those for the selected four single algorithms
In this study, based on the FY-4A imagery, the selected thresholds for different algorithms were close to those in previous studies. This is because there are various factors that would affect the setting of thresholds such as surface roughness, sensor wavelength ranges, and solar and viewing zenith angle [
64,
65]. Additionally, the atmospheric effects are also an important influencing factor [
58].
6. Conclusions
FY-4A/AGRI provides valuable data sources for studying dust over China, especially where ground-based observations are often not available. In this study, four dust detection algorithms applied to FY-4A/AGRI data were evaluated to examine the effectiveness of these algorithms in detecting the extent of dust over China. These algorithms have good agreement in detecting the spatial distribution of dust. BTD can detect dust events in the Chinese region effectively and accurately if the threshold is well set. NDDI still has a large area of misclassification, even though the thresholds were set according to different regions and times of the day. IDDI and DST had similar performances in dust detection during the daytime. Validation of the dust detection results for the four methods using PM10 concentrations showed that BTD had the best detection performance with an average POCD of 56.15%, followed by those for IDDI, DST, and NDDI, which were 48.22%, 46.75%, and 39.39%, respectively. Additionally, a low POFD for the combined algorithm of BTD and IDDI indicates its potential to accurately detect dust events. The validation of the dust detection algorithm remains a challenge. Future work will focus on two areas: first, to focus on dust detection at night to enable continuous dust detection during the day and night, and second, to investigate if the machine learning approach to dust detection can overcome the threshold limitations of empirical-based algorithms.