Next Article in Journal
Quality Analysis of Small Maritime Target Detection by Means of Passive Radar Reflectors in Different Sea States
Previous Article in Journal
Longitudinal Evolution of Storm-Enhanced Densities: A Case Study
 
 
Search for Articles:
Title / Keyword
Author / Affiliation / Email
Journal
Article Type
 
 
Advanced Search
 
Section
Special Issue
Volume
Issue
Number
Page
 
Journals
Remote Sensing
Volume 14
Issue 24
10.3390/rs14246341
remotesensing-logo
Submit to this Journal Review for this Journal Propose a Special Issue
► Article Menu

Article Menu

share Share announcement Help format_quote Cite question_answer Discuss in SciProfiles thumb_up
...
Endorse textsms
...
Comment
first_page
settings
Order Article Reprints
Font Type:
Arial Georgia Verdana
Font Size:
Aa Aa Aa
Line Spacing:
Column Width:
Background:
Article

Retrieval of Aerosol Single-Scattering Albedo from MODIS Data Using an Artificial Neural Network

by
Lin Qi
1,2,
Ronggao Liu
1,* and
Yang Liu
1
1
State Key Laboratory of Resources and Environmental Information System, Institute of Geographic Sciences and Natural Resources Research, Chinese Academy of Sciences, Beijing 100101, China
2
University of Chinese Academy of Sciences, Beijing 100049, China
*
Author to whom correspondence should be addressed.
Remote Sens. 2022, 14(24), 6341; https://doi.org/10.3390/rs14246341
Submission received: 19 October 2022 / Revised: 24 November 2022 / Accepted: 13 December 2022 / Published: 14 December 2022
(This article belongs to the Section Atmospheric Remote Sensing)
Browse Figures
Versions Notes

Abstract

:
Aerosol single-scattering albedo (SSA) is one of the largest sources of uncertainty in the evaluation of the aerosol radiative forcing effect. The SSA signal, coupled with aerosol optical depth (AOD) and surface reflectance in satellite images, is difficult to retrieve by the look-up table approach. In this study, we proposed an artificial neural network- (ANN) based approach that retrieves SSA over land based on MODIS (moderate resolution imaging spectroradiometer) visible (red band) reflectance variations among nearby pixels that have different surface reflectivities. Using the training dataset generated by the radiative transfer model, the ANN model was trained to establish the relationship among SSA, surface reflectance, and top of atmosphere (TOA) reflectance. Then, based on the trained ANN model, SSA can be retrieved using the surface and apparent reflectance of several heterogeneous pixels. According to sensitivity analysis, this method works well on nonuniform land surfaces with high AODs. The root mean square error (RMSE) of retrieved and measured SSA (from 28 sites of AErosol RObotic NETwork, AERONET) was 0.042, of which the results with an error less than 0.03 accounted for 51%. In addition, the SSA retrieval method was applied to several thick aerosol layer events over different areas (South Asia, South America, and North China Plain) and compared with the ozone monitoring instrument near-UV aerosol data product (OMAERUV). The comparison results of the images show that the retrieval method of visible wavelength proposed in this study has similar outcomes to those from the ultraviolet wavelengths in these regions. The retrieval algorithm we propose provides an effective way to produce an SSA product in visible wavelength and might help to better estimate the aerosol radiative and optical properties over high heterogeneous areas, which is important for the aerosol radiative impact estimate at a regional scale.

1. Introduction

Atmospheric aerosols can directly or indirectly affect the radiation budget of the Earth’s climate through the scattering and absorption of radiation by particles or the influence on the radiation characteristics of cloud properties [1,2,3]. Aerosols participate in the radiative transfer process by absorbing and scattering the solar downward radiation and the surface upward-reflected radiation. The scattering of aerosols can reduce solar radiation incidents on the Earth’s surface and reduce the temperature of the Earth–atmosphere system [4,5,6,7,8]. On the other hand, the absorption of radiation by aerosol particles increases the radiation received by the Earth–atmosphere system [9,10].
The aerosol single-scattering albedo (SSA, ω) determines the ratio of radiation scattering and extinction, which is the ratio of the aerosol scattering coefficient to the total aerosol scattering plus absorption [11,12]. Therefore, the separation of aerosol absorption and the scattering ratio in remote sensing signals is the core idea of retrieving aerosol SSA from remote sensing data. The effect of aerosol absorption is much smaller than that of scattering, making it difficult to detect the absorption effect from the aerosol contribution in satellite signals. In the past 20 years, some researchers have developed retrieval algorithms [13,14,15,16,17,18,19,20,21]. Among them, amplifying the absorption of aerosols on the surface reflection signal by method can make the retrieval of absorption and SSA easier [14,18].
During the last 20 years, several studies have been conducted to develop SSA retrieval algorithms through better captured absorption of aerosols on the surface reflectance signals using the ultraviolet (UV) band or the solar glint [13,14,15,16,17,18,19,20,21]. Specifically, due to the strong Rayleigh scattering in the UV band, its optical path is longer and the lower surface contribution of the UV band makes the UV band more sensitive to the absorption of aerosols. Therefore, the absorption ratio in the aerosol radiation effect can be better separated by the UV band to realize the retrieval of SSA [18]. Based on the above principles, an SSA product with a satisfactory resolution (13 km × 24 km) for total ozone mapping spectrometer (TOMS) and ozone monitoring instrument (OMI) sensors at UV wavelengths was produced [19,20,21]; this SSA product was also verified to be in good agreement with the inversion product of SKYNET sites when aerosols are thick [22]. The Earth polychromatic imaging camera (EPIC) has 10 channels (317, 325, 340, 388, 443, 551, 680, 688, 764, and 779 nm) covering UV and visible wavelength and can continuously observe the Earth’s surface. Using the well-estimated surface reflectance by the multi-angle implementation of atmospheric correction (MAIAC) algorithm, the absorption of aerosol and SSA values can be retrieved from the UV band [16]. Similar to the characteristics of the UV band, the solar glint over the high-reflectivity ocean also makes aerosol absorption more significant, which is conducive to aerosol SSA retrieval. However, both UV band and solar glint techniques showed limitations in SSA retrieval. Kaufman et al. (2002) reported that, because it is limited by the areas of light spot, the solar glint is difficult to widely use in the production of regional SSA products [14]. In addition, although there are many algorithms and products for retrieving SSA in the UV band, due to the low energy in the UV band, it is impossible to obtain high spatial resolution images [19,20,21], and the SSA values in the UV band can only represent the radiation characteristics of the UV band with less solar radiation energy. Therefore, it is more important to retrieve SSA in the visible band, which accounts for half of the solar radiation.
In addition to directly highlighting the aerosol absorption effect, parameterizing critical values as a function of SSA is an alternative way to retrieve SSA. This approach assumed that there existed critical surface albedo (CSA) or critical optical depth (COD), at which the top of atmosphere reflectance (TOA, directly measured by remote sensing instruments) is independent to the variation of surface reflectance (SR) or atmosphere optical depth (AOD) [17,23]. By making full use of the differences of satellite signals between different pixels or time series data, the corresponding critical values can be derived and thus the SSA can further be retrieved using the look-up table (LUT). Due to the high applicability, the critical value technique has been widely applied to derive SSA in visible bands across regions and even the globe [24,25,26,27,28]; however, this technique also showed several limitations in SSA retrieval. For one thing, it highly relies on the existence of the critical values. Previous studies [17,29] found both CSA and COD could not be obtained in some certain conditions. For instance, Seidel and Popp (2012) found the derivative of TOA reflectance with respect to AOD was always positive in low aerosol types (e.g., the stratospheric type), indicating no CSA values in such situations. For another, existing studies to retrieve SSA through critical value techniques were based on established LUT by simulating the radiative transfer model (RTM). However, the relationship between SSA and reflectance is nonlinear and, when surface conditions and the type of aerosol are complex, the LUT will be inefficient to look up the corresponding result [30].
To eliminate the limitations of the critical value technique, this study proposed a method to directly establish the relationship between reflectance and SSA through the difference in reflectance of adjacent surface pixels using MODIS (moderate resolution imaging spectroradiometer) data. The relationship between reflectance (including SR and TOA reflectance) and SSA needs to be established to retrieve SSA from satellite signals and surface signals. Instead of establishing LUT, the artificial neural network (ANN) technique was applied to simulate the complex nonlinear relationships. ANN has also been applied in remote sensing inversion in recent years [31,32,33,34]. Using the ANN technique, the relationship between reflectance and SSA can be more easily established to retrieve SSA value from the reflectivity of adjacent pixels.
In Section 2, the data and retrieval methods used in this research are described. Section 3.1 conducts a preliminary experiment on the principle of the retrieval method. Section 3.2 quantitatively analyzes the SSA retrieval results based on the method established in this study. In addition, three typical research cases (thick aerosol layer events over South Asia, South America, and North China) were used to further validate the retrieval method in Section 3.3.

2. Materials and Methods

2.1. MODIS Data

MODIS is an instrument mounted onboard the Terra satellite; its probe operates in 36 channels (spectral range from 0.41 to 15 μm) and offers daily coverage of the globe at resolutions of 250 (2 channels), 500 (5 channels), and 1000 m (29 channels). MODIS provides a variety of bands for long-term and high-resolution aerosol retrieval and can be widely used for studying the distribution and changes of aerosols due to its swath, spectral bands, and spatial resolution [35].
In this study, band 1 (red band, 620–670 nm, in 500 m resolution), in Level-1B data of the MODIS (MOD02) product with higher surface reflectance and greater sensitivity to SSA changes [36], was selected to retrieve the SSA value. The large reflectance difference of adjacent pixels in red bands is conducive to realizing the SSA retrieval method from different TOA-surface reflectance relationships.

2.2. OMAERUV Data

The ozone monitoring instrument (OMI) is one of the four sensors onboard the Aura satellite that was launched in 2004. The OMI near-UV aerosol data product (OMAERUV) uses 354 nm and 388 nm wavelengths to retrieve aerosol characteristics [18,21]. The OMAERUV product provides AOD, aerosol absorption optical depths (AAOD), SSA at three different wavelengths (354, 388, and 500 nm), aerosol index, and other ancillary and geolocation parameters in the OMI field of view (13 × 24 km).

2.3. AERONET Data

The AERONET (AErosol RObotic NETwork) project is a federation of ground-based remote sensing aerosol networks established by NASA and PHOTONS [37,38]. It provides a long-term, continuous, and readily accessible public domain database of aerosol optical, microphysical, and radiative properties for the validation of satellite retrieval. The retrieval results of the retrieval method are verified by AERONET version 3 level 1.5 all points inversion data [39]. When AOD is greater than 0.3, the theoretical uncertainty of SSA is 0.03 [40].
The SSA inversion model based on ANN was verified by the SSA data at 675 nm from AERONET sites with different aerosols in 2020. To ensure the data quality before comparison, the AERONET sites were filtered. Only the AERONET inversion data within one hour before and after the overpass time, cloud free within 10 km, and the zenith angle between 15° and 65° were selected. A total of 28 sites were collected and used to compare with the method’s retrieval results. The detailed information of the 28 AERONET sites is listed in Table 1.
During the validation process, the SSA values retrieved from the pixels within 10 km × 10 km around AERONET sites were compared with the inversion results of the AERONET sites.

2.4. Methods

TOA reflectance is determined by AOD, SSA, and surface reflectance. In Section 3.1, we conducted some theoretical experiments to verify the feasibility of decoupling SSA and AOD. Through the preliminary test in the simulation of 3.1.2, it is proved that using the surface reflectance and TOA reflectance of a pair of surface pixels with different reflectance can distinguish the contributions of AOD and SSA to TOA reflectance, which allows us to directly retrieve SSA value from the surface reflectance and TOA reflectance of pixel pairs without determining the AOD first. Figure 1 shows the flow chart of this retrieval method. In our SSA retrieval method, we take the surface pixels in the range of 10×10 as a retrieval unit based on the assumption that the properties of aerosols are homogeneous in a certain area, in which we will look for the pixels with reflectance differences to participate in the retrieval and to decouple SSA and AOD contribution using the difference in the responsiveness to aerosol absorption caused by the difference of surface reflectance. The red band (band 1 in MODIS, 620 nm–670 nm, with 500 m resolution) that is more sensitive to SSA changes than other visible bands is selected to retrieve SSA. Based on the non-linear characteristics of the relationship between reflectance and SSA and the need to simplify the retrieval method, the ANN technique is chosen to establish the relationship between reflectance and SSA values to realize the SSA retrieval in the visible waveband. The method includes six main steps: (1) preliminary test, (2) generating training datasets, (3) designing and training of neural network, (4) data preprocessing, (5) obtaining the retrieval SSA results, and (6) error analysis of the ANN model results. Details of each step are shown below.

2.4.1. Preliminary Test

TOA reflectance is affected by surface reflectance, aerosol AOD, and SSA. Before proposing the retrieval method of SSA, the 6S model was used to simulate the response of TOA reflectance to SSA and AOD under a variety of different aerosol conditions, and the response of TOA reflectance to SSA and AOD was studied with the change of ground surface reflectance. The aerosol conditions are set as 0.67~0.99 in SSA and 0.0~2.0 in AOD. Simultaneously, the TOA reflectance of two reflectance surfaces (0.1 and 0.3) was simulated under different combinations of AOD and SSA and the sensitivity of TOA reflectance to the AOD and SSA effects was analyzed. The results are shown in Section 3.1.

2.4.2. Generating Training Datasets

In order to make the training data to include as many observed geometries, aerosol conditions, and surface reflectance as possible, we use RTM to simulate the reflectance and SSA relationship in different situations to build the ANN training set. Here, we choose to use the 6S model [41,42] to generate the ANN training dataset. The input parameters of the 6S model include geometric conditions, a gaseous component atmospheric model, an aerosol model (type and concentration), a spectral band, and a surface reflectance (type and spectral variation). The atmospheric model is set as US Standard 62. SSA values are simulated by setting the combination of different aerosol particle components, which can simulate various aerosol types. Retrieval units are simulated from 3 input reflectance pixels with different reflectance. The generated training dataset includes a total of 419,904 lines of data for 4 solar zenith angle values, 4 sensor zenith angle values, 9 relative azimuth angle values, 6 AOD values, 18 SSA values, and 3 × 3 × 3 surface reflectance pairs’ values, covering different combinations of geometric conditions, aerosol conditions, and surface reflectance. It is worth noting that the 3 × 3 × 3 surface reflectance pairs’ values refer to the mutual combination of reflectance values from 3 groups. To enhance the training efficiency, the 9 surface reflectances in this study were first ordered and divided into 3 groups (low, medium, and high groups) with 3 values in each group. We then randomly selected one reflectance value from each group, resulting in 3 × 3 × 3 surface reflectance pairs’ values. See Table 2 and Table 3 for the value list of the input parameters and the percentage of each aerosol particle component in the simulation of SSA.

2.4.3. Design and Training of Neural Network

A complete neural network includes input layer, hidden layer, and output layer. The input layer of the neural network model is set as 9 neurons, which means that 9 input variables are required. The 9 input variables are the cosine of the three angles referring to the geometric observations (solar zenith angle, sensor zenith angle, and relative azimuth angle between solar and sensor), the reflectance of the darkest, middle, and brightest surfaces (3 input variables of surface reflectance), and their corresponding TOA reflectance (3 input variables of TOA reflectance) in one retrieval unit. The hidden layer is set to contain three layers, which are 500, 200, and 50 neurons with a ReLU activation function [43]. The output layer contains one neuron, which represents the SSA value. This neural network is trained using the datasets introduced in the previous section generated by the 6S model.

2.4.4. Data Preprocessing

Before inputting the data into the ANN model for retrieval, it is necessary to preprocess the input data. Figure 2 shows the details of the data preprocessing and the SSA retrieval algorithm.
Cloud mask data from MOD35 is used to remove cloudy pixels before inputting parameters into the ANN model. Each retrieval of each retrieval unit requires 9 input parameters, namely the cosine of the three angles (solar zenith angle, sensor zenith angle, relative azimuth angle between solar and sensor), the reflectance of the darkest, middle, and brightest surfaces, and their corresponding TOA reflectance in one retrieval unit. For the three input parameters of angles, we take the value of the center pixel of the retrieval unit in the observation data as the input. To provide the surface reflectance information required for the retrieval input, an 8-day 500 m surface reflectance dataset containing seasonal variation was synthesized from 5 years (2014–2018) of MOD02 data using the minimum reflectivity technique [36]. The synthesized surface reflectance dataset was used to search for pixel sequences of different reflectances for SSA retrieval (as shown in the blue part of Figure 2). The pixels are sorted according to the surface reflectance of the red band in a retrieval unit. The No. 25, No. 50, and No. 75 pixels from minimum to maximum are taken into the SSA retrieval algorithm as a sequence of pixels representing different reflectances. Using the surface reflectance and TOA reflectance of the three selected pixels, we filled in the remaining 6 input parameters.

2.4.5. Performing Retrieval

After the model training and the input data preprocessing are completed, we can input the input data into the trained ANN model and obtain the SSA results from the retrieval of the ANN model.

2.4.6. Error Analysis of ANN Model Results

The 6S model was used to simulate 537,600 datasets (Table 4, input parameter list) to perform error analysis on the established ANN model. In order to explore the effect of low aerosol load and surface reflectance differences on the accuracy of the retrieval ANN model, the input parameters (AOD and surface reflectance) of 6S model were expanded to cover more cases. Based on the ANN model-calculated SSA and the corresponding 6S model-calculated SSA, error analysis of the ANN model was performed. Then, we analyzed the correlation between the error of the ANN model and the AOD, the surface reflectance difference, and the surface reflectivity estimation error to show the ANN model’s error response to each parameter. In the analysis, except for the parameter to be studied, other parameters remained the same to ensure a single variable.
In the error analysis, since all input parameters except surface reflectance could be obtained from satellite observations, random noise was added to the input surface reflectance to quantify the influence of the difference between the input value of the surface reflectance and the true value of the model retrieval. The root mean square error (RMSE) was used to accurately evaluate the model result.

3. Results

3.1. Preliminary Test of Decoupling AOD and SSA from TOA Reflectance

3.1.1. Simulation of TOA-SR vs. AOD for Different SSA Settings

TOA reflectance is the response of surface reflectance and atmosphere condition. Figure 3 shows the relationship between TOA-SR (the difference of TOA reflectance and surface reflectance) and AOD under different aerosol SSA conditions. When there is no aerosol above surface (AOD equal to 0), TOA reflectance is all contributed by the Rayleigh scattering of atmospheric molecules, resulting in a higher value of TOA reflectance than surface reflectance. When aerosol exists, the relationship between TOA-SR and AOD is highly determined by the values of SSA (Figure 3). For higher SSA (SSA no less than 0.77), the TOA-SR increased with AOD. The increasing magnitude was more significant when SSA was larger. On the contrary, for lower SSA, the TOA-SR decreased with AOD. The strong absorption effect of aerosol even led to the decrease of TOA, making TOA-SR negative in the case of thick strong absorption aerosol. Figure 3 indicated that the changes of TOA-SR resulted from the coupling effect of SSA and AOD. Therefore, to retrieve the SSA, decoupling them from TOA reflectance is necessary.

3.1.2. Simulation of TOA Reflectance under Two Specific Conditions for Different AOD and SSA Setting Pairs

By assuming that there is a pair of surfaces with different reflectivity with the same aerosol coverage (both AOD and SSA), Figure 4 shows the simulation of the TOA reflectance under different combinations of AOD and SSA in two surface pixels with different reflectivity. For low-reflectivity surface (ref = 0.1), the TOA reflectance increased with AOD or SSA. Similarly, for the surface with higher reflectance (red = 0.3), the TOA reflectance still always increased with SSA regardless of the values of AOD. However, the changes of TOA reflectance showed an opposite trend, with AOD increasing under different SSA values with a cut-off point at approximately 0.94. When SSA took a relatively small value (0.82, 0.88, 0.92), TOA reflectance decreased with AOD. When the SSA was higher (0.96, 1.00), the aerosol absorption was lower, resulting in an increasing trend in TOA reflectance with AOD.
The line a (ref = 0.1) or line b (ref = 0.3) in Figure 4 are depicted to show that it was ill-posed to derive the unique combination of AOD and SSA for a certain value of TOA under single-surface reflectance due to the coupling contrition of AOD and SSA; while, the disjoint of all contour lines of AOD or SSA within the two-dimensional coordination indicated this coupling relationship could be disentangled with two different surface reflectances (Figure 4). For example, the point P derived from the intersection of line a and line b determined the exclusive combination of AOD and SSA.

3.2. Retrieval Error Analysis

The error distribution of the ANN retrieval model on the validation dataset is shown in Figure 5a. The error of the ANN model conforms to a normal distribution and the RMSE of the ANN retrieval model is 0.0319.

3.2.1. Relationship between Retrieval Error and AOD

Figure 5b shows the value of the ANN retrieval model error as the AOD changes under the same conditions (solar zenith angle = 20°, sensor zenith angle = 40°, relative azimuth angle = 60°). The retrieval results of the RMSE of the ANN model show that when the AOD increased from 0.1 to 0.5, the RMSE dropped sharply (from 0.085 to 0.012). After that, the AOD continued to increase but the RMSE remained stable, indicating that the model can provide high retrieval accuracy under high aerosol load but still has large uncertainty under low aerosol load

3.2.2. Relationship between Retrieval Error and Reflectance Difference

The retrieval results of the validation data were grouped according to the difference of reflectance, and the error of retrieval of each group was counted to determine the effect of the difference in reflectance on the model retrieval accuracy under the same conditions (solar zenith angle = 30°, sensor zenith angle = 50°, relative azimuth angle = 120°, AOD = 0.8) (Figure 6). When the average reflectance difference was less than 0.06, the RMSE was 0.015 and the RMSE gradually decreased as the reflectance difference increased. When the average reflectance difference was greater than 0.10, the RMSE was reduced to 0.01. Larger differences between surface reflectances can bring more available information for decoupling AOD and SSA and can make the retrieval of SSA more effective.

3.2.3. Influence of Surface Reflectance Noise on the Accuracy of Model Retrieval

Figure 7 shows the error distribution histogram of the ANN model under four different reflection noise settings and the corresponding RMSE. When the surface reflection noise increased to 0.05, the ANN model error increased by two times compared with the no noise setting, and the corresponding RMSE increased from 0.032 to 0.077. Smaller surface reflectivity noise settings will not have a major impact on the inversion results of the ANN model. When the noise was between 0.01 and 0.02, the increase in RMSE was not large compared with the noise-free setting (gradually increasing from 0.032 to 0.048).

3.2.4. Validation Result of ANN Model Simulation Results Based on AERONET Data

The aerosol data of 28 AERONET sites (the details are listed in Table 1) were selected to validate the simulation results of the ANN model. The overall RMSE is 0.042. From the comparison between the AERONET SSA at these in situ sites and the SSA based on the neural network model inversion (Figure 8), 75% of the model inversion error is less than 0.05, and 51% of the model inversion error is less than 0.03. There were several sites where moderate absorption aerosols will occur. The Skukuza site is located in a sub-tropical rural agricultural area in South Africa that is impacted by biomass burning or smoke aerosols during the dry season. It is also impacted by other airmasses associated with anthropogenic activities (biomass burning and urban/industrial aerosols) and natural sources (sea salt and desert dust) [44]. Belsk is an AERONET site located in the center of Poland. Affected by local anthropogenic emission sources and transportation of marine aerosols and biomass burning aerosols in surrounding areas, multiple aerosol sources mix with each other, resulting in variable aerosol types [45]. The mixing of aerosols with different absorbing and scattering effect means the aerosol with moderate absorption may be caused by a variety of different combinations, which makes the composition of aerosol particles with SSA in the range of 0.85–0.90 more complex. For some kinds of moderate absorption aerosols, the model simulation results lead to some underestimates of 0.05. Apart from these aerosol types, the overall retrieval error of the model can be controlled within 0.05.

3.3. Example Analysis of ANN Model Retrieval Results

3.3.1. A Thick Aerosol Case of South Asia

Due to the monsoon climate, an abundance of aerosol sources (including anthropogenic aerosols produced by human activities and dust aerosol from deserts), and the high Himalayan Mountains in the north, South Asia has become one of the regions with the thickest aerosols in the world. There was a thick aerosol episode that occurred in south Asia on 25 June 2019. Figure 9a,b show the true color composite map and AOD map of South Asia from MODIS. The AERONET sun/sky radiometer in Kanpur measured an AOD value of 1.24 and an SSA value of 0.98. It can be seen that nearly 40% of this area is covered by cloud. Between the cloud-covered areas of northern India, these areas are covered by thick aerosols. We applied our method to the MOD02 image. Figure 9c,d show the SSA value distribution of our retrieval results and the OMAERUV product in the corresponding areas, respectively. These SSA values of the OMAERUV product are between 0.95 and 1.00, which is consistent with the measurement results of AERONET, indicating strongly scattering aerosols. Our results have different features in different regions. The SSA values vary from 0.95 to 1.00 in the northwestern of the zone, which is consistent with the values of the OMAERUV product. In central and eastern India, the SSA values range from 0.75 to 0.90 and the corresponding OMAERUV results range from 0.95 to 1.0. Considering that the closer the wavelengths are, the more similar the retrieved SSA values would be; therefore, compared with the difference of SSA distribution, the distinction between SSA values due to narrow band differences (500 nm for OMAERUV and 650 nm for the proposed method) is acceptable. The differences of SSA distribution might be the large retrieval uncertainty caused by the low aerosol loading.

3.3.2. An Aerosol Diffusion Case of South America

An aerosol diffusion episode from Sao Paulo to Santa Cruz occurred on 19 September 2019. Figure 10a,b show the true color composite map and the AOD map of South America from MODIS. There was an area with AOD over 1.0 spreading northwest from Sao Paulo. There are two AERONET sites in this area. The sun/sky radiometer measurements of the SANTA_CRUZ_UTEPSA and Sao Paulo sites are 0.95 and 0.96, respectively. The OMAERUV SSA results shown in Figure 10d range from 0.90 to 1.00 in most areas. In the areas covered by thick aerosols, our retrieval results (shown in Figure 10c) range from 0.90 to 1.00 and are consistent with the measurements from the AERONET sites. There are differences between the two results in some regions covered by aerosols, which may be caused by the errors in the estimation of surface reflectance. The aerosol diffused from Sao Paulo to Santa Cruz shows good continuity in both retrieval results’ values, but our results show a higher resolution compared with the OMAERUV results.

3.3.3. A Thick Aerosol Case of North China

North China is one of the regions where thick aerosols occur most frequently in the world. The North China Plain is in the coastal monsoon region of eastern China, carrying a quarter of China’s population and economic activities and providing a large amount of aerosol sources. The north and west sides of the North China Plain relate to mountains, which hinder the diffusion of aerosols. These factors have led to the frequent occurrence of heavy haze weather in the North China Plain. Figure 11a,b show the true color composite map and AOD map of a thick aerosol covering the North China Plain taken by MODIS on 1 January 2019. Most of the areas of the North China Plain are covered by thick aerosols with AOD larger than 1.0. In the OMAERUV results shown in Figure 11d, the retrieved SSA values of the area covered with aerosols range from 0.85 to 1.00, which is consistent with the measurements of the AERONET sites. Affected by the MODIS cloud detection algorithm, the range of SSA we can retrieved is limited. However, in the area covered by thick aerosols, our retrieval SSA results and distribution are consistent with those of the AERONET sites and the OMAERUV results.

3.3.4. The Statistics of Pixel-Based Difference

In order to evaluate the retrieval ability of our method more intuitively, we compared the difference between our retrieval results and the OMAERUV product’s results by pixel after cloud screening and aerosol screening in each image with aerosol case. Figure 12 shows the results of the histogram. The statistical difference generally conforms to the normal distribution with the mean value of 0. When the difference between two results is less than 0.03, we consider that the retrieval results of the two are consistent, which accounts for 37.5%, 45.9%, and 46.0%, respectively, in each image. Among all the pixels, the proportions of the pixels with a difference less than 0.05 were 64.4%, 65.2%, and 63.8%, respectively. The comparison results of the images show that the retrieval method proposed in this study has the same level of retrieval ability as the existing algorithms that use the ultraviolet wavelengths.

4. Discussion

In addition to AOD, SSA knowledge is essential for estimating the direct and semidirect radiative forcing of aerosols [46]. The influence of SSA and AOD on solar radiation is reflected in TOA reflectance and the effects of AOD and SSA are coupled with each other, making it difficult to distinguish the contribution of SSA from TOA reflectance. As shown in Figure 4, for two surface pixels with different reflectances, the TOA reflectance’s responses to SSA and AOD changes are distinguished.
SSA contributes to TOA reflectance and can be distinguished by pixels with different surface reflectivities. A method to retrieve aerosol SSA in the visible band from the MODIS data is constructed, which can reflect the aerosol characteristics in the key bands that account for half of the solar radiation. This retrieval method, based on surface data series, requires a large amount of data as input and the traditional LUT method cannot meet the needs of this complex retrieval method and linear interpolation will bring errors. Neural network technology can establish the relationship between reflectance and SSA, which can realize complex retrieval and fast calculation at the same time, making it suitable for the simulation of nonlinear relationships among multiple variables.
The established ANN model was evaluated and the error and the source of the retrieval model were studied. The error analysis results show that the retrieval accuracy of the ANN model is closely related to AOD, the reflectivity difference of the surface pixels participating in the retrieval, and the accuracy of the estimation of the surface reflectance. From the analysis of the reflectance difference, it can be seen that the heterogeneity of the surface reflectance has significant impact on the SSA retrieval capability of our proposed method. When the differences between surface reflectance is small (below 0.12 in this study), the retrieval accuracy of the method would be relatively lower (RMSE above 0.01). For different land cover, the heterogeneity of surface reflectivity is also different. Because of the demand for surface heterogeneity of this retrieval method, the heterogeneity of low reflectivity surfaces such as dense forests may not meet the demand, which will limit the applicability of the algorithm. Considering the relationship between the surface heterogeneity and the model performance, the applicability of this retrieval method for different heterogeneous real surfaces needs to be explored in future studies. Adding random noise to the input surface reflectance can simulate the estimation error of the surface reflectance. When the estimation error is under 0.01, the surface reflectance error has little effect on the results of the ANN model. However, when the estimation error is above 0.05, it will have a larger impact on the ANN model retrieval results. The uncertainty of the estimation red band surface reflectance using the minimum reflectivity technology is about 0.012 [47]. It is found that the average estimation uncertainty of the synthetic surface reflectance will not significantly affect the retrieval accuracy of the model. The above phenomenon puts forward a high requirement for the estimation method of surface reflectivity.
The ANN model is verified by the inversion result of the solar/sky photometer at the AERONET sites. When the optical depth retrieved by the satellite is greater than 0.5, the single-scatter albedo uncertainty retrieved by the ANN model is estimated to be ±0.012. As shown in Figure 9, when there is a moderate absorption aerosol, the model’s inversion value appears to be underestimated. The moderate absorption aerosols with SSA ranging from 0.85 to 0.90 contain a mixture of different aerosol components, showing a complex aerosol type. The relative lower simulation performance for some moderate absorption aerosols may be because the training data of this experiment are achieved through the mixing of aerosol particle components, which may lack specific aerosol combinations. The higher simulation uncertainty for this type of aerosol implies that the simulation of moderate absorption aerosols with different component combinations should be added to expand the applicability in the retrieval of moderate absorption aerosols in the future.
The ANN model is applied to three real aerosol events in South Asia, South America, and North China, and the retrieved SSA distribution map corresponding to the remote sensing image is obtained and compared with the results of OMAEROUV. The SSA value from OMAERUV is 500 nm while MODIS is 620–670 nm. The wavelength difference between OMAERUV and MODIS B1 is 150 nm, which leads to the difference of SSA values. Figure 13 shows the spectral dependence of SSA with four aerosol types. For the continental aerosol, the maritime aerosol and the urban aerosol, SSA decreases while wavelength increases. For the desert aerosol, SSA increases with the increase in wavelength. The absolute difference and relative difference between OMAERUV and MODIS are 0.0203, 0.0124, 0.0002 and 0.0148 and 2.13%, 1.38%, 0.02%, and 2.13%, respectively. For strong scattering aerosols, such as the maritime aerosol, SSA changes little with wavelength and is basically considered consistent. For absorption aerosols, the change rate of SSA increases with the increase in absorption. The absolute error and relative error of urban aerosols, which are strong absorbing aerosols, reach 0.0148 and 2.13%, which are less than the uncertainty of the model itself. For desert aerosols, the absolute error and relative error are 0.0203 and 2.13%, which are relatively large but also acceptable for comparison. Comparing the retrieval results corresponding to the actual case and OMAEROUV, the ANN-based model can display the distribution of SSA in the area with higher resolution. However, in areas with no aerosol coverage, the ANN model will obtain invalid retrieval values. In the comparison of two results, we only used the MODIS data from Terra to compare with the OMAERUV data from OMI. Due to the difference in their overpass times, the distribution of aerosols may change over time. This may lead to mismatches during comparisons. However, SSA represents the inherent properties of aerosols, which are relatively stable. It will not change dramatically during this time gap and can still be used for comparison.
In Figure 9, Figure 10 and Figure 11, we can see some error data marked as black, which refer to the result that the retrieval value is greater than 1.0. It is not difficult to find out from the error information and distribution areas that this problem is caused by the cloud residue of the cloud detection algorithm, which indicates that the error of the cloud detection algorithm will bring great uncertainty to aerosol SSA retrieval. An accurate cloud detection algorithm is essential to improve the retrieval ability of aerosol SSA. In Section 3.3, we can also see that there are some areas covered by aerosol that should be suitable for retrieval methods, but the actual retrieval results are obviously different from the surrounding results. Careful observation shows that these areas usually show different colors with adjacent areas. This difference in reflectivity indicates that the land cover of these areas may have changed in the past time, leading to the errors in the estimation surface reflectance by the minimum reflectivity technique, resulting in large errors in the retrieval SSA results.
The applicability of this retrieval method and the influence of different conditions on retrieval need further research in the future. In addition, the uncertainty of residual cloud in cloud detection and surface reflectance estimation will also affect the retrieval capability of this method; these uncertainties need further research. In the future, strict aerosol screening can be used to avoid retrieval in areas without aerosol coverage. In addition, a better estimation method of surface reflectance can also reduce the uncertainty caused by estimation error. The ANN-based SSA retrieval model provides a new idea for aerosol SSA retrieval on visible bands using surface pixel sets and is of great significance for the study of the radiative forcing of aerosol visible light band SSA datasets. To overcome the existing problems of the ANN-based retrieval method, the next step may require emerging deep learning network models and high-precision cloud detection algorithms to further improve the ability to retrieve SSA models.

5. Conclusions

In this study, a retrieval model of SSA products based on neural network technology is proposed, which is suitable for the SSA retrieval of MODIS visible band satellite products.
Compared with the in situ AERONET site sun/sky photometric inversion measurement results, the SSA retrieval method proposed in this study can reach an error of 0.042 under a typical uncertainty of satellite-derived optical depth larger than 0.15.
The ANN model is used to retrieve the satellite data of three thick aerosol layer events over South Asia, South America, and North China. The retrieval results show that the SSA retrieval method proposed in this study can well retrieve the SSA distribution in an aerosol-covered region under specific highly heterogeneous regions. A high-resolution SSA product retrieval model supported by this research might help to better understand the distribution and changes of aerosol radiation effects in visible wavelengths over high heterogeneous areas.

Author Contributions

Conceptualization, R.L.; methodology, L.Q and R.L.; software, L.Q.; validation, L.Q. and Y.L.; formal analysis, L.Q.; investigation, L.Q.; resources, L.Q.; data curation, L.Q.; writing—original draft preparation, L.Q.; writing—review and editing, R.L. and Y.L.; visualization, L.Q. and Y.L.; supervision, R.L.; project administration, R.L.; funding acquisition, R.L. All authors have read and agreed to the published version of the manuscript.

Funding

This research was funded by the Strategic Priority Research Program of the Chinese Academy Sciences under Grant XDA19080303.

Data Availability Statement

Not applicable.

Conflicts of Interest

The authors declare no conflict of interest.

References

  1. Charlson, R.J.; Schwartz, S.E.; Hales, J.M.; Cess, R.D.; Coakley, J.A.; Hansen, J.E.; Hofmann, D.J. Climate forcing by anthropogenic aerosols. Science 1992, 255, 423–430. [Google Scholar] [CrossRef]
  2. Sokolik, I.N.; Toon, O.B. Direct radiative forcing by anthropogenic airborne mineral aerosols. Nature 1996, 381, 681–683. [Google Scholar] [CrossRef]
  3. Lacagnina, C.; Hasekamp, O.P.; Bian, H.S.; Curci, G.; Myhre, G.; van Noije, T.; Schulz, M.; Skeie, R.B.; Takemura, T.; Zhang, K. Aerosol single-scattering albedo over the global oceans: Comparing PARASOL retrievals with AERONET, OMI, and AeroCom models estimates. J. Geophys. Res.-Atmos. 2015, 120, 9814–9836. [Google Scholar] [CrossRef] [Green Version]
  4. Boucher, O.; Randall, D.; Artaxo, P.; Bretherton, C.; Feingold, G.; Forster, P.; Kerminen, V.-M.; Kondo, Y.; Liao, H.; Lohmann, U.; et al. Clouds and Aerosols. In Climate Change 2013: The Physical Science Basis. Contribution of Working Group I to the Fifth Assessment Report of the Intergovernmental Panel on Climate Change; Stocker, T.F., Qin, D., Plattner, G.-K., Tignor, M., Allen, S.K., Boschung, J., Nauels, A., Xia, Y., Bex, V., Midgley, P.M., Eds.; Cambridge University Press: Cambridge, UK; New York, NY, USA, 2013; pp. 571–658. [Google Scholar]
  5. Chung, C.E.; Ramanathan, V.; Decremer, D. Observationally constrained estimates of carbonaceous aerosol radiative forcing. Proc. Natl. Acad. Sci. USA 2012, 109, 11624–11629. [Google Scholar] [CrossRef] [PubMed] [Green Version]
  6. Li, Z.Q.; Chen, H.; Cribb, M.; Dickerson, R.; Holben, B.; Li, C.; Lu, D.; Luo, Y.; Maring, H.; Shi, G.; et al. Preface to special section on east Asian studies of tropospheric aerosols: An international regional experiment (EAST-AIRE). J. Geophys. Res.-Atmos. 2007, 112. [Google Scholar] [CrossRef] [Green Version]
  7. Li, Z.Q.; Niu, F.; Fan, J.W.; Liu, Y.G.; Rosenfeld, D.; Ding, Y.N. Long-term impacts of aerosols on the vertical development of clouds and precipitation. Nat. Geosci. 2011, 4, 888–894. [Google Scholar] [CrossRef]
  8. Ramanathan, V.; Crutzen, P.J.; Kiehl, J.T.; Rosenfeld, D. Atmosphere—Aerosols, climate, and the hydrological cycle. Science 2001, 294, 2119–2124. [Google Scholar] [CrossRef] [PubMed] [Green Version]
  9. Kaufman, Y.J.; Koren, I.; Remer, L.A.; Rosenfeld, D.; Rudich, Y. The effect of smoke, dust, and pollution aerosol on shallow cloud development over the Atlantic Ocean. Proc. Natl. Acad. Sci. USA 2005, 102, 11207–11212. [Google Scholar] [CrossRef] [PubMed] [Green Version]
  10. Rosenfeld, D. Suppression of rain and snow by urban and industrial air pollution. Science 2000, 287, 1793–1796. [Google Scholar] [CrossRef]
  11. Russell, P.B.; Redemann, J.; Schmid, B.; Bergstrom, R.W.; Livingston, J.M.; McIntosh, D.M.; Ramirez, S.A.; Hartley, S.; Hobbs, P.V.; Quinn, P.K.; et al. Comparison of aerosol single scattering albedos derived by diverse techniques in two North Atlantic experiments. J. Atmos. Sci. 2002, 59, 609–619. [Google Scholar] [CrossRef]
  12. Myhre, G.; Berglen, T.F.; Johnsrud, M.; Hoyle, C.R.; Berntsen, T.K.; Christopher, S.A.; Fahey, D.W.; Isaksen, I.S.A.; Jones, T.A.; Kahn, R.A.; et al. Modelled radiative forcing of the direct aerosol effect with multi-observation evaluation. Atmos. Chem. Phys. 2009, 9, 1365–1392. [Google Scholar] [CrossRef] [Green Version]
  13. Devi, A.; Satheesh, S.K. Global maps of aerosol single scattering albedo using combined CERES-MODIS retrieval. Atmos. Chem. Phys. Discuss. 2022, 22, 5365–5376. [Google Scholar] [CrossRef]
  14. Kaufman, Y.J.; Martins, J.V.; Remer, L.A.; Schoeberl, M.R.; Yamasoe, M.A. Satellite retrieval of aerosol absorption over the oceans using sunglint. Geophys. Res. Lett. 2002, 29, 34-1–34-4. [Google Scholar] [CrossRef] [Green Version]
  15. Li, L.; Dubovik, O.; Derimian, Y.; Schuster, G.L.; Lapyonok, T.; Litvinov, P.; Ducos, F.; Fuertes, D.; Chen, C.; Li, Z.; et al. Retrieval of aerosol components directly from satellite and ground-based measurements. Atmos. Chem. Phys. 2019, 19, 13409–13443. [Google Scholar] [CrossRef] [Green Version]
  16. Lyapustin, A.; Go, S.; Korkin, S.; Wang, Y.; Torres, O.; Jethva, H.; Marshak, A. Retrievals of Aerosol Optical Depth and Spectral Absorption From DSCOVR EPIC. Front. Remote Sens. 2021, 2, 645794. [Google Scholar] [CrossRef]
  17. Satheesh, S.K.; Srinivasan, J. A method to infer short wave absorption due to aerosols using satellite remote sensing. Geophys. Res. Lett. 2005, 32, 13814. [Google Scholar] [CrossRef]
  18. Torres, O.; Bhartia, P.K.; Herman, J.R.; Ahmad, Z.; Gleason, J. Derivation of aerosol properties from satellite measurements of backscattered ultraviolet radiation: Theoretical basis. J. Geophys. Res.-Atmos. 1998, 103, 17099–17110. [Google Scholar] [CrossRef]
  19. Torres, O.; Bhartia, P.K.; Herman, J.R.; Sinyuk, A.; Ginoux, P.; Holben, B. A long-term record of aerosol optical depth from TOMS observations and comparison to AERONET measurements. J. Atmos. Sci. 2002, 59, 398–413. [Google Scholar] [CrossRef]
  20. Torres, O.; Bhartia, P.K.; Sinyuk, A.; Welton, E.J.; Holben, B. Total Ozone Mapping Spectrometer measurements of aerosol absorption from space: Comparison to SAFARI 2000 ground-based observations. J. Geophys. Res.-Atmos. 2005, 110, 18. [Google Scholar] [CrossRef] [Green Version]
  21. Torres, O.; Tanskanen, A.; Veihelmann, B.; Ahn, C.; Braak, R.; Bhartia, P.K.; Veefkind, P.; Levelt, P. Aerosols and surface UV products from Ozone Monitoring Instrument observations: An overview. J. Geophys. Res.-Atmos. 2007, 112. [Google Scholar] [CrossRef]
  22. Jethva, H.; Torres, O. A comparative evaluation of Aura-OMI and SKYNET near-UV single-scattering albedo products. Atmos. Meas. Tech. 2019, 12, 6489–6503. [Google Scholar] [CrossRef] [Green Version]
  23. de Almeida Castanho, A.D.; Vanderlei Martins, J.; Artaxo, P. MODIS Aerosol Optical Depth Retrievals with high spatial resolution over an Urban Area using the Critical Reflectance. J. Geophys. Res. Atmos. 2008, 113, 2201. [Google Scholar] [CrossRef]
  24. Kaufman, Y.J. Satellite sensing of aerosol absorption. J. Geophys. Res.-Atmos. 1987, 92, 4307–4317. [Google Scholar] [CrossRef]
  25. Kaufman, Y.J.; Fraser, R.S.; Ferrare, R.A. Satellite measurements of large-scale air-pollution—Methods. J. Geophys. Res.-Atmos. 1990, 95, 9895–9909. [Google Scholar] [CrossRef]
  26. Kaufman, Y.J.; Tanre, D.; Dubovik, O.; Karnieli, A.; Remer, L.A. Absorption of sunlight by dust as inferred from satellite and ground-based remote sensing. Geophys. Res. Lett. 2001, 28, 1479–1482. [Google Scholar] [CrossRef] [Green Version]
  27. Zhu, L.; Martins, J.V.; Remer, L.A. Biomass burning aerosol absorption measurements with MODIS using the critical reflectance method. J. Geophys. Res.-Atmos. 2011, 116, 7202. [Google Scholar] [CrossRef] [Green Version]
  28. Wells, K.C.; Martins, J.V.; Remer, L.A.; Kreidenweis, S.M.; Stephens, G.L. Critical reflectance derived from MODIS: Application for the retrieval of aerosol absorption over desert regions. J. Geophys. Res.-Atmos. 2012, 117, 3202. [Google Scholar] [CrossRef] [Green Version]
  29. Seidel, F.C.; Popp, C. Critical surface albedo and its implications to aerosol remote sensing. Atmos. Meas. Tech. 2012, 5, 1653–1665. [Google Scholar] [CrossRef] [Green Version]
  30. Huttunen, J.; Kokkola, H.; Mielonen, T.; Mononen, M.E.J.; Lipponen, A.; Reunanen, J.; Lindfors, A.V.; Mikkonen, S.; Lehtinen, K.E.J.; Kouremeti, N.; et al. Retrieval of aerosol optical depth from surface solar radiation measurements using machine learning algorithms, non-linear regression and a radiative transfer-based look-up table. Atmos. Chem. Phys. 2016, 16, 8181–8191. [Google Scholar] [CrossRef] [Green Version]
  31. Kolios, S.; Hatzianastassiou, N. Quantitative Aerosol Optical Depth Detection during Dust Outbreaks from Meteosat Imagery Using an Artificial Neural Network Model. Remote Sens. 2019, 11, 1022. [Google Scholar] [CrossRef]
  32. Lary, D.J.; Remer, L.A.; MacNeill, D.; Roscoe, B.; Paradise, S. Machine Learning and Bias Correction of MODIS Aerosol Optical Depth. IEEE Geosci. Remote Sens. Lett. 2009, 6, 694–698. [Google Scholar] [CrossRef] [Green Version]
  33. Okada, Y.; Mukai, S.; Sano, I. Neural network approach for aerosol retrieval. In Proceedings of the IGARSS 2001: Scanning the Present and Resolving the Future, Sydney, Australia, 9–13 July 2001; Volume 1–7, pp. 1716–1718. [Google Scholar]
  34. Schroeder, T.; Behnert, I.; Schaale, M.; Fischer, J.; Doerffer, R. Atmospheric correction algorithm for MERIS above case-2 waters. Int. J. Remote Sens. 2007, 28, 1469–1486. [Google Scholar] [CrossRef]
  35. Levy, R.C.; Mattoo, S.; Munchak, L.A.; Remer, L.A.; Sayer, A.M.; Patadia, F.; Hsu, N.C. The Collection 6 MODIS aerosol products over land and ocean. Atmos. Meas. Tech. 2013, 6, 2989–3034. [Google Scholar] [CrossRef] [Green Version]
  36. Hsu, N.C.; Tsay, S.C.; King, M.D.; Herman, J.R. Aerosol properties over bright-reflecting source regions. IEEE Trans. Geosci. Remote Sens. 2004, 42, 557–569. [Google Scholar] [CrossRef]
  37. Holben, B.N.; Eck, T.F.; Slutsker, I.; Tanre, D.; Buis, J.P.; Setzer, A.; Vermote, E.; Reagan, J.A.; Kaufman, Y.J.; Nakajima, T.; et al. AERONET—A federated instrument network and data archive for aerosol characterization. Remote Sens. Environ. 1998, 66, 1–16. [Google Scholar] [CrossRef]
  38. Dubovik, O.; Smirnov, A.; Holben, B.N.; King, M.D.; Kaufman, Y.J.; Eck, T.F.; Slutsker, I. Accuracy assessments of aerosol optical properties retrieved from Aerosol Robotic Network (AERONET) Sun and sky radiance measurements. J. Geophys. Res. 2000, 105, 9791–9806. [Google Scholar] [CrossRef] [Green Version]
  39. Giles, D.M.; Sinyuk, A.; Sorokin, M.G.; Schafer, J.S.; Smirnov, A.; Slutsker, I.; Eck, T.F.; Holben, B.N.; Lewis, J.R.; Campbell, J.R.; et al. Advancements in the Aerosol Robotic Network (AERONET) Version 3 database—Automated near-real-time quality control algorithm with improved cloud screening for Sun photometer aerosol optical depth (AOD) measurements. Atmos. Meas. Tech. 2019, 12, 169–209. [Google Scholar] [CrossRef] [Green Version]
  40. Dubovik, O.; Holben, B.; Eck, T.F.; Smirnov, A.; Kaufman, Y.J.; King, M.D.; Tanre, D.; Slutsker, I. Variability of absorption and optical properties of key aerosol types observed in worldwide locations. J. Atmos. Sci. 2002, 59, 590–608. [Google Scholar] [CrossRef]
  41. Kotchenova, S.Y.; Vermote, E.F.; Levy, R.; Lyapustin, A. Radiative transfer codes for atmospheric correction and aerosol retrieval: Intercomparison study. Appl. Opt. 2008, 47, 2215–2226. [Google Scholar] [CrossRef] [Green Version]
  42. Vermote, E.F.; Tanre, D.; Deuze, J.L.; Herman, M.; Morcrette, J.J. Second Simulation of the Satellite Signal in the Solar Spectrum, 6S: An overview. IEEE Trans. Geosci. Remote Sens. 1997, 35, 675–686. [Google Scholar] [CrossRef]
  43. Nair, V.; Hinton, G.E. Rectified Linear Units Improve Restricted Boltzmann Machines. In Proceedings of the 27th international conference on machine learning, Haifa, Israel, 21–24 June 2010; pp. 807–814. [Google Scholar]
  44. Kumar, K.R.; Sivakumar, V.; Reddy, R.R.; Gopal, K.R.; Adesina, A.J. Inferring wavelength dependence of AOD and Angstrom exponent over a sub-tropical station in South Africa using AERONET data: Influence of meteorology, long-range transport and curvature effect. Sci. Total Environ. 2013, 461, 397–408. [Google Scholar] [CrossRef] [PubMed]
  45. Markowicz, K.M.; Stachlewska, I.S.; Zawadzka-Manko, O.; Wang, D.X.; Kumala, W.; Chilinski, M.T.; Makuch, P.; Markuszewski, P.; Rozwadowska, A.K.; Petelski, T.; et al. A Decade of Poland-AOD Aerosol Research Network Observations. Atmosphere 2021, 12, 1583. [Google Scholar] [CrossRef]
  46. Mallet, M.; Dulac, F.; Formenti, P.; Nabat, P.; Sciare, J.; Roberts, G.; Pelon, J.; Ancellet, G.; Tanre, D.; Parol, F.; et al. Overview of the Chemistry-Aerosol Mediterranean Experiment/Aerosol Direct Radiative Forcing on the Mediterranean Climate (ChArMEx/ADRIMED) summer 2013 campaign. Atmos. Chem. Phys. 2016, 16, 455–504. [Google Scholar] [CrossRef] [Green Version]
  47. Hsu, N.C.; Jeong, M.J.; Bettenhausen, C.; Sayer, A.M.; Hansell, R.; Seftor, C.S.; Huang, J.; Tsay, S.C. Enhanced Deep Blue aerosol retrieval algorithm: The second generation. J. Geophys. Res.-Atmos. 2013, 118, 9296–9315. [Google Scholar] [CrossRef]
Remotesensing 14 06341 g001 550
Figure 1. Flow chart of SSA retrieval method.
Figure 1. Flow chart of SSA retrieval method.
Remotesensing 14 06341 g001
Remotesensing 14 06341 g002 550
Figure 2. Flow chart of data preprocessing and SSA retrieval algorithm details.
Figure 2. Flow chart of data preprocessing and SSA retrieval algorithm details.
Remotesensing 14 06341 g002
Remotesensing 14 06341 g003 550
Figure 3. Relationship between the TOA-surface reflectance (the difference between TOA reflectance and surface reflectance) and AOD under different SSA conditions (surface reflectance = 0.1, solar zenith angle = 0°, sensor zenith angle = 0°, relative azimuth angle = 160°).
Figure 3. Relationship between the TOA-surface reflectance (the difference between TOA reflectance and surface reflectance) and AOD under different SSA conditions (surface reflectance = 0.1, solar zenith angle = 0°, sensor zenith angle = 0°, relative azimuth angle = 160°).
Remotesensing 14 06341 g003
Remotesensing 14 06341 g004 550
Figure 4. Simulation of TOA reflectance (surface reflectance = 0.3) vs. TOA reflectance (surface reflectance = 0.1) for different SSA and AOD setting pairs.
Figure 4. Simulation of TOA reflectance (surface reflectance = 0.3) vs. TOA reflectance (surface reflectance = 0.1) for different SSA and AOD setting pairs.
Remotesensing 14 06341 g004
Remotesensing 14 06341 g005 550
Figure 5. (a) The error distribution histogram of the ANN model. (b) Relationship between RMSE and AOD (environmental parameter settings: solar zenith angle = 20°, sensor zenith angle = 40°, relative azimuth angle = 60°).
Figure 5. (a) The error distribution histogram of the ANN model. (b) Relationship between RMSE and AOD (environmental parameter settings: solar zenith angle = 20°, sensor zenith angle = 40°, relative azimuth angle = 60°).
Remotesensing 14 06341 g005
Remotesensing 14 06341 g006 550
Figure 6. RMSE for different surface reflectance differences. SRlow, SRmid, and SRhigh are the lowest, middle, and highest reflectance among the participating pixels (environmental parameter setting: solar zenith angle = 30°, sensor zenith angle = 50°, relative azimuth angle = 120°, AOD = 0.8).
Figure 6. RMSE for different surface reflectance differences. SRlow, SRmid, and SRhigh are the lowest, middle, and highest reflectance among the participating pixels (environmental parameter setting: solar zenith angle = 30°, sensor zenith angle = 50°, relative azimuth angle = 120°, AOD = 0.8).
Remotesensing 14 06341 g006
Remotesensing 14 06341 g007 550
Figure 7. Analysis of inversion error of the ANN model for different surface reflectance noise. Four input surface reflectance noises were randomly assigned: (a) noise = 0.00; (b) noise = 0.01; (c) noise = 0.02; (d) noise = 0.05.
Figure 7. Analysis of inversion error of the ANN model for different surface reflectance noise. Four input surface reflectance noises were randomly assigned: (a) noise = 0.00; (b) noise = 0.01; (c) noise = 0.02; (d) noise = 0.05.
Remotesensing 14 06341 g007
Remotesensing 14 06341 g008 550
Figure 8. Comparison of estimated SSA from the ANN model and SSA from AERONET. Note: each dot represents a line of data where the AERONET observations spatiotemporally matched MODIS data. The observations from Belsk and Skukuza sites were highlighted using square and triangle marks, respectively.
Figure 8. Comparison of estimated SSA from the ANN model and SSA from AERONET. Note: each dot represents a line of data where the AERONET observations spatiotemporally matched MODIS data. The observations from Belsk and Skukuza sites were highlighted using square and triangle marks, respectively.
Remotesensing 14 06341 g008
Remotesensing 14 06341 g009 550
Figure 9. (a) True color composite map from MODIS (data came from Terra in South Asia on 25 June 2019), (b) AOD map of MOD04, (c) SSA retrieval value of our method, (d) SSA retrieval value (500 nm) of OMAERUV. The location and inversion SSA values of Kanpur site are marked in figures.
Figure 9. (a) True color composite map from MODIS (data came from Terra in South Asia on 25 June 2019), (b) AOD map of MOD04, (c) SSA retrieval value of our method, (d) SSA retrieval value (500 nm) of OMAERUV. The location and inversion SSA values of Kanpur site are marked in figures.
Remotesensing 14 06341 g009
Remotesensing 14 06341 g010 550
Figure 10. (a) True color composite map from MODIS (data came from Terra in South America on 19 September 2019), (b) AOD map of MOD04, (c) SSA retrieval value of our method, (d) SSA retrieval value of OMAERUV. The locations and inversion SSA values of SANTA_CRUZ_UTEPSA and Sao Paulo sites are marked in figures.
Figure 10. (a) True color composite map from MODIS (data came from Terra in South America on 19 September 2019), (b) AOD map of MOD04, (c) SSA retrieval value of our method, (d) SSA retrieval value of OMAERUV. The locations and inversion SSA values of SANTA_CRUZ_UTEPSA and Sao Paulo sites are marked in figures.
Remotesensing 14 06341 g010
Remotesensing 14 06341 g011 550
Figure 11. (a) True color composite map from MODIS (data came from Terra in North China Plain on 1 January 2019), (b) AOD map of MOD04, (c) SSA retrieval value of our method, (d) SSA retrieval value of OMAERUV. The locations and inversion SSA values of Beijing and XuZhou-CUMT sites are marked in figures.
Figure 11. (a) True color composite map from MODIS (data came from Terra in North China Plain on 1 January 2019), (b) AOD map of MOD04, (c) SSA retrieval value of our method, (d) SSA retrieval value of OMAERUV. The locations and inversion SSA values of Beijing and XuZhou-CUMT sites are marked in figures.
Remotesensing 14 06341 g011
Remotesensing 14 06341 g012 550
Figure 12. Difference histogram between our results and the OMAERUV product; (a) South Asia on 25 June 2019, (b) South America on 19 September 2019, (c) North China Plain on 1 January 2019.
Figure 12. Difference histogram between our results and the OMAERUV product; (a) South Asia on 25 June 2019, (b) South America on 19 September 2019, (c) North China Plain on 1 January 2019.
Remotesensing 14 06341 g012
Remotesensing 14 06341 g013 550
Figure 13. SSA values versus wavelength with different aerosol type.
Figure 13. SSA values versus wavelength with different aerosol type.
Remotesensing 14 06341 g013
Table
Table 1. Geographic location and annual mean aerosol parameters of AERONET sites.
Table 1. Geographic location and annual mean aerosol parameters of AERONET sites.
SiteLongitudeLatitudeMean AODMean SSA
Banizoumbou2.665° E13.547° N0.3140.874
Belsk20.792° E51.837° N0.1920.770
BONDVILLE88.372° W40.053° N0.1510.769
Capo Verde22.935° W16.733° N0.3200.970
CARTEL71.931° W45.380° N0.3430.969
Churchill93.818 W58.736 N0.1750.961
El Arenosillo6.733 W37.105 N0.1490.982
Evora7.912 W38.568 N0.2180.972
GSFC76.840 W38.992 N0.1240.959
Halifax63.594 W44.638 N0.4230.968
La Parguera67.045 W17.970 N0.1170.985
Lampedusa12.632 E35.517 N0.2590.980
Lecce University18.111 E40.335 N0.1570.939
Lille3.142 E50.612 N0.2000.953
MD Science Center76.612 W39.281 N0.2230.885
Minsk27.601 E53.920 N0.1710.905
Missoula114.083 W46.917 N0.4670.942
Monterey121.855 W36.593 N0.3530.917
Moscow MSU MO37.522 E55.707 N0.1810.926
Rimrock116.992 W46.487 N0.1580.923
Rio Branco67.869 W9.957 S0.1390.844
SERC76.556 W38.889 N0.1250.984
Sevilleta106.885 W34.355 N0.2370.928
Sioux Falls96.626 W43.736 N0.4510.965
Skukuza31.587 E24.992 S0.1990.896
TABLE MOUNTAIN CA117.680 W34.380 N0.9700.970
Toravere26.467 E58.265 N0.2080.936
Wallops75.472 W37.933 N0.1060.901
Table
Table 2. List of input parameters of 6S model.
Table 2. List of input parameters of 6S model.
ParameterRange
Solar Zenith Angle10, 30, 50, 70
Sensor Zenith Angle10, 30, 50, 70
Relative Azimuth Angle0, 30, 60, 90, 120, 150, 160, 170, 180
AOD (550 nm)0.3, 0.5, 0.8, 1.0, 1.5, 2.0
SSA
(Band 1 of MODIS)		0.65, 0.67, 0.70, 0.71, 0.74, 0.75, 0.77, 0.79, 0.80, 0.82, 0.87, 0.89, 0.91, 0.92, 0.94, 0.96, 0.98, 1.00
Surface Reflectance
(Band 1 of MODIS)		0.06, 0.10, 0.12, 0.14, 0.15, 0.17, 0.18, 0.20, 0.23
Table
Table 3. List of percentage of each aerosol particle component in the simulation of SSA.
Table 3. List of percentage of each aerosol particle component in the simulation of SSA.
SSA
Value		Dust-like
Component		Water-Soluble
Component		Oceanic
Component		Soot
Component
0.650.20.600.2
0.670.10.20.60.1
0.700.20.30.40.1
0.710.40.40.10.1
0.740.10.40.40.1
0.750.20.50.20.1
0.770.20.60.10.1
0.790.10.70.10.1
0.800.10.800.1
0.8200.900.1
0.870.90.100
0.890.700.30
0.910.70.10.20
0.920.70.20.10
0.940.50.500
0.9600.50.50
0.9800.10.90
1.000010
Table
Table 4. List of input parameters of 6S model of error analysis dataset.
Table 4. List of input parameters of 6S model of error analysis dataset.
ParameterRange
Solar Zenith Angle10, 20, 30, 40, 50
Sensor Zenith Angle10, 20, 30, 40, 50
Relative Azimuth Angle0, 30, 60, 90, 120, 150, 180
AOD (550 nm)0.1, 0.2, 0.3, 0.5, 0.8, 1.0, 1.5, 2.0
SSA
(Band 1 of MODIS)		0.71, 0.75, 0.79, 0.82, 0.92, 1.00
Surface Reflectance
(Band 1 of MODIS)		0.04, 0.06, 0.08, 0.10, 0.12, 0.14, 0.16, 0.18, 0.20, 0.22, 0.24, 0.26
Publisher’s Note: MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Share and Cite

MDPI and ACS Style

Qi, L.; Liu, R.; Liu, Y. Retrieval of Aerosol Single-Scattering Albedo from MODIS Data Using an Artificial Neural Network. Remote Sens. 2022, 14, 6341. https://doi.org/10.3390/rs14246341

AMA Style

Qi L, Liu R, Liu Y. Retrieval of Aerosol Single-Scattering Albedo from MODIS Data Using an Artificial Neural Network. Remote Sensing. 2022; 14(24):6341. https://doi.org/10.3390/rs14246341

Chicago/Turabian Style

Qi, Lin, Ronggao Liu, and Yang Liu. 2022. "Retrieval of Aerosol Single-Scattering Albedo from MODIS Data Using an Artificial Neural Network" Remote Sensing 14, no. 24: 6341. https://doi.org/10.3390/rs14246341

APA Style

Qi, L., Liu, R., & Liu, Y. (2022). Retrieval of Aerosol Single-Scattering Albedo from MODIS Data Using an Artificial Neural Network. Remote Sensing, 14(24), 6341. https://doi.org/10.3390/rs14246341

Note that from the first issue of 2016, this journal uses article numbers instead of page numbers. See further details here.

Article Metrics

Back to TopTop