A Novel Aerosol Optical Depth Retrieval Method Based on SDAE from Himawari-8/AHI Next-Generation Geostationary Satellite in Hubei Province

Abstract

Atmospheric aerosols play an important role in the ecological environment, climate change, and human health. Aerosol optical depth (AOD) is the main measurement of aerosols. The next-generation geostationary satellite Himawari-8, loaded with the Advanced Himawari Imager (AHI), provides observation-based estimates of the hourly AOD. However, a highly accurate evaluation of AOD using AHI is still limited. In this paper, we establish a Stacked Denoising AutoEncoder (SDAE) model to retrieve highly accurate AOD using AHI. We explore the SDAE to retrieve AOD by taking the ground-observed AOD as the output and taking the AHI image, the month, hour, latitude, and longitude as the input data. This approach was tested in the Hubei province of China. Traditional machine learning methods such as Extreme Learning Machines (ELMs), BackPropagation Neural Networks (BPNNs), and Support Vector Machines (SVMs) are also used to evaluate model performance. The results show that the proposed method has the highest accuracy. We also compare the proposed method with ground-observed AOD measurements at the Wuhan University site, showing good consistency between the satellite-retrieved AOD and the ground-observed value. The study of the spatiotemporal change pattern of the hourly AOD in the Hubei province shows that the algorithm has good stability in the face of changes in the angle and intensity of sunlight.

1. Introduction

Atmospheric aerosols are tiny particles suspended in the air, with different particle sizes, shapes, and compositions. These attributes often change over time [1,2]. The harmful particles in aerosols are closely related to cardiovascular diseases, respiratory diseases, allergies, and even premature births [3,4,5], causing severe harm to humans, animals, and plants [6,7]. In addition, aerosols have a great influence on climate change [8]. Specifically, aerosols act as the condensation nuclei of clouds, resulting in much larger raindrops [9,10], further causing heavy rainstorms. Aerosol’s radiative effects are the direct driving factors of global average precipitation changes and can also cause changes in large-scale precipitation patterns [11]. Therefore, monitoring aerosols is particularly important [6,12].

Aerosol optical depth (AOD) is a metric for aerosols, which describes the light-reducing effect of aerosols; it is one of the most important parameters of aerosols and a key physical quantity to characterize the degree of atmospheric turbidity. The AOD value can reflect the content and distribution of aerosols in the atmosphere, which is an important factor in determining the climate effect of aerosols, and also plays an important role in remote sensing atmospheric correction and fine particulate matter pollution assessment. A high AOD value indicates an increase in the longitudinal accumulation of aerosols, which may lead to a decrease in atmospheric visibility, thus reflecting the degree of air pollution in the region, to a certain extent.

Combining satellite images and ground data is an effective and rapid method for retrieving AOD products [13]. Substantial research on AOD retrieval has been carried out since the AVHRR was first combined with ground data to yield AOD products in 1992 [14,15,16,17]. These studies can be divided into three stages: low-resolution AOD products (above 1 km, 100–1 km, and below 100 m), medium- to low-resolution AOD products, and high-resolution AOD products. In the first stage, SeaWiFS [15], POLDER [18], TOMS [19], OMI [20], ATSR [21], ATSR-2, AATSR, SEVIRI [22] images, and ground data were combined to produce AOD products. In the second stage, MODIS [13,23,24], MERIS [25], MISR [26,27], and VIIRS [28] ground data were combined to retrieve the AOD. In the last stage, high-resolution satellites (HJ [29,30], Landsat [31], and GF-1 [32,33]) were used for urban AOD inversion. These satellites were polar-orbiting satellites, providing relatively high spatial resolution for AOD products. However, their temporal resolution was very low, and high-temporal-resolution AOD products are not available. To address this issue, geostationary satellites such as GMS [34], MTSAT [22,23], MSG [35], FY2D [36], GOCI [24,28,37], GEOS [38], and AHI [39,40,41] were launched in succession and utilized to retrieve high-temporal-resolution AOD products.

The inversion methods used in the aforementioned satellite studies can be divided into the following three categories.

(1)

The dark Target (DT) algorithm assumes that dark pixel areas exist in the remote sensing image, that the surface exhibits Lambertian reflectance, and that the atmospheric properties are uniform. The linear relationship between the red light and near-infrared bands in the dark pixel areas is used to obtain the true surface reflectance of the red light band. Then, the real surface reflectivity and remote sensing reflectivity are substituted into the appropriate atmospheric radiation transfer model to obtain relevant parameters, and the AOD results can be obtained through calculation. However, this method is not effective in areas with bright surfaces.

(2)

The Deep-Blue (DB) algorithm is based on the fact that surface reflection is weak in the blue light band but the atmospheric reflection is strong. First, real long-term surface reflectivity data must be obtained, and then the remote sensing reflectivity data of the blue light band and the corresponding real surface reflectivity data of the blue light band are substituted into the appropriate atmospheric radiative transfer model to obtain the relevant parameters. However, when the surface reflectance in the blue light band is greater than 0.1, the error of this method increases.

(3)

Empirical methods can be used to analyze the statistical relationship between remote sensing images and AOD and then establish a mathematical statistical inversion model of remote sensing images and AOD. These methods are strongly affected by geographical, weather, and external interference factors, are limited to single remote sensing images, and cannot be applied at a large scale.

Although DT and DB algorithms can accurately invert AOD under specific conditions, they are strictly limited by geographical and weather conditions. The DT method relies on “dark pixels” such as vegetation and water bodies with low surface reflectivity, making them difficult to apply for complex surface types such as cities and deserts. Although the DB algorithm has expanded its scope of application to a certain extent, it is still limited by factors such as cloud cover and surface reflectance changes. The empirical method establishes the inversion model by analyzing the statistical relationship between remote sensing images and AOD; however, its accuracy and reliability are highly dependent on the selection and representativeness of sample data, and it is easily affected by external interference factors. In addition, empirical methods are usually limited to the application of a single remote sensing image, which is difficult to generalize and validate on a large scale [42,43].

Section 2 introduces the study area and data. The model setup and experimental design are presented in Section 3. The estimation validation and hourly patterns of the AOD distribution are expressed in Section 4. Finally, the conclusions are presented in Section 5.

2. Study Area and Data

2.1. Study Area

The study area is in the Hubei province, which has a spatial extent of 29°01′53″ to 33°16′47″N latitude and 108°21′42″ to 116°07′50″E longitude (Figure 1). Hubei is known as “the province with a thousand lakes” and is located in the middle reaches of the Yangtze River and north of Dongting Lake in Central China. Hubei covers an area of over 185,900 km², with 15 jurisdiction cities (including the capital city, Wuhan), 1 autonomous prefecture, and 1 forest area, and has a permanent population of approximately 59 million. There are mountains, hills, hillocks, and plains in the Hubei province. The westernmost part of the province is mountainous with high altitudes (above 3105 m), and the middle and southern parts of the province consist of plains at low altitudes (below 35 m). The Hubei province is located in the subtropical zone, and most of the province has a humid subtropical monsoon climate, with rain and heat in the same season and great vertical changes. This geographical location and climate characteristics make the Hubei province an important area for aerosol generation, transmission, and deposition [44].

2.2. Experimental Data

2.2.1. AHI Data

The Advanced Himawari Imager (AHI) is an advanced imager installed on the Himawari-8 geostationary satellite, which was launched in Japan on 7 October 2014. The AHI’s coverage area extends from 60°S to 60°N and 80°E to 160°W. The AHI is a multispectral imager with 16 channels, with a wide wavelength range of 0.47–13.3 µm, encompassing 3 Visible (VIS), 3 Near-Infrared (NIR), and 10 Infrared (IR) bands (https://www.eorc.jaxa.jp/ptree/userguide.html (accessed on 1 January 2024)). The AHI has a high temporal resolution of approximately 10 min across the entire observation area. The spatial resolution of the AHI is 0.5 km (band 3), 1 km (band 1, 2, and 4), and 2 km (band 5–16).

In this study, the AHI data used were Himawari L1 Gridded data, which included the albedo (reflectance × cos (solar zenith angle (SOZ)) of band 01–band 06), brightness temperature of band 07–band 16, satellite zenith angle, satellite azimuth angle, solar zenith angle, solar azimuth angle, and observation hours (UTC), and the spatial resolution was 2 km. The nearest neighbor interpolation method was applied to resample the original multi-resolution datasets, including band 3 (0.5 km spatial resolution) and bands 1, 2, and 4 (1 km spatial resolution), to a uniform grid of 2 km spatial resolution for subsequent analysis. We downloaded the Himari L1 Gridded data from the Japan Aerospace Exploration Agency (JAXA) (http://www.eorc.jaxa.jp/ptree/index.html (accessed on 1 January 2024)).

2.2.2. AERONET AOD Data

The Aerosol Robotic Network (AERONET) project is a federation of ground-based remote sensing aerosol networks established by NASA and PHOTONS based on the CIMEL Electronique (CE318) multiband sun photometer, which measures spectral sun irradiance and sky radiance (https://aeronet.gsfc.nasa.gov/ (accessed on 1 January 2024)). The network requires standardized instruments, calibration, processing, and distribution, and the AOD data from AERONET are always considered the ground truth for validating satellite products [27]. In this study, the AOD data from 61 ground-based sites from AERONET in the AHI observation area were selected as the experimental data. The details of the AOD data from the 61 ground-based sites are listed in

Table 1. Statistics of AOD in the AERONET from July 2015 to April 2017 (abbreviations for full names are in Appendix A).

Site	Number	Longitude	Latitude	Elevation (m)	AOD (Means and Standard Deviations)	Site	Number	Longitude	Latitude	Elevation (m)	AOD (Means and Standard Deviations)
ALI	20	120.81	23.51	2216	0.34 ± 0.20	KNI	181	126.64	37.57	25	0.40 ± 0.23
ANM	871	126.33	36.54	1450	0.32 ± 0.28	KOP	64	127.12	37.52	180	0.51 ± 0.30
BAE	202	124.63	37.97	380	0.27 ± 0.20	KSC	95	127.49	37.34	75	0.51 ± 0.38
BAN	95	107.61	−6.89	3	0.32 ± 0.15	KTH	83	127.31	37.31	65	0.41 ± 0.23
BCA	967	116.32	39.93	43	0.40 ± 0.40	KUU	268	129.19	35.58	3	0.36 ± 0.17
BJ	606	116.38	39.98	52	0.37 ± 0.35	LAG	1470	128.75	−16.11	1880	0.08 ± 0.07
BS	1540	139.35	−25.90	3	0.05 ± 0.03	LLF	295	121.71	−31.26	3	0.07 ± 0.03
CB	765	149.11	−35.27	560	0.05 ± 0.03	LNT	179	101.42	20.93	1030	0.43 ± 0.33
CKU	113	120.22	23.00	780	0.55 ± 0.39	LCD	555	146.39	−18.52	3	0.10 ± 0.04
CMM	597	98.97	18.77	1210	0.44 ± 0.30	LUL	27	120.87	23.47	1950	0.06 ± 0.07
CHI	298	120.50	23.50	230	0.61 ± 0.30	MAK	317	119.57	−5.00	3	0.23 ± 0.16
DAL	209	104.42	43.58	1040	0.08 ± 0.06	NGH	88	105.80	21.05	210	0.64 ± 0.29
DON	24	116.73	20.70	15	0.26 ± 0.23	NON	81	102.72	17.88	1560	0.60 ± 0.41
DOU	87	120.55	23.71	1650	0.63 ± 0.32	NOT	95	137.14	37.33	2800	0.21 ± 0.17
EPA	272	121.19	24.97	360	0.27 ± 0.17	OMK	610	98.43	17.80	1230	0.23 ± 0.17
FOW	1993	141.70	−31.09	3	0.04 ± 0.03	OSA	228	135.59	34.65	22	0.25 ± 0.21
FUK	202	130.48	33.52	12	0.32 ± 0.20	PAL	44	113.95	−2.23	3	0.45 ± 0.36
GOT	82	101.41	9.29	1890	0.24 ± 0.28	PON	30	109.19	0.08	3	0.90 ± 0.93
GAN	771	84.13	25.87	5200	0.83 ± 0.47	PUS	1053	129.08	35.24	35	0.23 ± 0.20
GAN	1374	128.87	37.77	220	0.20 ± 0.16	SEO	495	126.95	37.46	42	0.35 ± 0.29
GOS	228	126.16	33.29	3	0.28 ± 0.17	SHI	566	135.36	33.69	18	0.18 ± 0.14
HAN	721	127.27	37.34	38	0.31 ± 0.27	SIL	964	100.04	13.82	5	0.48 ± 0.21
HOK	264	141.34	43.08	30	0.23 ± 0.21	SIN	52	103.78	1.30	15	0.55 ± 0.38
HON	37	114.12	22.48	8	0.37 ± 0.21	SON	91	103.91	21.33	1120	0.78 ± 0.55
IRK	94	103.09	51.80	430	0.24 ± 0.22	SMS	54	100.61	7.18	80	0.31 ± 0.23
JAB	726	132.89	−12.66	3	0.13 ± 0.08	TPC	123	121.50	25.03	9	0.35 ± 0.21
KBK	75	127.57	37.41	90	0.45 ± 0.33	USM	287	100.30	5.36	3	0.35 ± 0.31
KDW	25	128.76	37.69	320	0.35 ± 0.16	UBO	283	104.87	15.25	170	0.31 ± 0.31
KIK	222	127.01	35.96	680	0.48 ± 0.23	XIA	874	116.96	39.75	40	0.38 ± 0.39
KKN	217	128.61	35.89	410	0.43 ± 0.20	YON	881	126.94	37.56	28	0.33 ± 0.26
KMN	289	126.44	34.91	15	0.31 ± 0.17

. The 61 ground-based sites were distributed mainly in the terrestrial and offshore areas of Asia and Oceania, with 49 land sites and 12 marine sites and a total of 24,413 measurements from 6:00 to 17:00 between 2015 and 2017. The overall mean and standard deviation of the AOD was 0.27 ± 0.29. The measured means and standard deviations at each site are shown in Table 1. The AODs of China and South Asia were relatively high, and the AODs in Australia and its surrounding areas were relatively low.

2.2.3. AOD Data in Hubei

In this study, ground-level AOD observations were conducted at Wuhan University (114°21′17.37″E, 30°31′46.26″N), located in the capital city of the Hubei province, from January to December 2016, with a CE318 sun photometer with a temporal resolution of 10 min, which is the same instrument used to obtain the AERONET data [44]. The mean and standard deviation of the AOD data from the Wuhan University observation site was 0.62 ± 0.36, which was similar to some of the 61 ground-based sites, including the CHI (Chiayi) site, DOU (Douliu) site, NGH (NGHIA_DO) site, and NON (Nong_Khai) site, with means and standard deviations of 0.61 ± 0.30, 0.63 ± 0.32, 0.64 ± 0.29, and 0.60 ± 0.41, respectively. The AOD statistics from all the AERONET and Wuhan University sites are presented in

Table 2. Statistics of AOD in the AERONET and at the Wuhan University site.

Site	Number of Data	Mean	Median	Std	Min	Max	Time Range (Year)
ALL AERONET	24,413	0.2703	0.1547	0.2989	0.0094	3.0357	2015–2017
Wuhan University Site	6945	0.6273	0.5250	0.3629	0.0850	2.9396	2016

.

2.2.4. JAXA AOD Data

This study incorporated the JAXA-official AHI Version 2.0 AOD product as a key validation reference. This product employed the dark target (DT) algorithm combined with a look-up table (LUT) for retrieval. The Level 2 AOD dataset, formatted in NetCDF, provided hourly retrievals at 5 km spatial resolution over the AHI full-disk observation area (85°E–165°W, 60°S–60°N), ensuring consistent spatiotemporal coverage with our primary satellite data source (http://www.eorc.jaxa.jp/ptree/index.html (accessed on 1 January 2024)).

2.2.5. MODIS AOD Data

This study employed the NASA-official MODIS Level 2 aerosol product (MOD04_3K/MYD04_3K) from the Terra/Aqua satellites as one of the validation datasets. This product employed the dark target (DT) algorithm combined with a look-up table (LUT) for retrieval. This 5 min swath product, with a 3 km spatial resolution, retrieved atmospheric aerosol optical properties and mass concentrations over global land and ocean surfaces. The algorithm utilized precomputed look-up tables to derive reflectance/transmittance fluxes from AOD and incorporated quality control flags (e.g., cloud mask, retrieval confidence) and auxiliary parameters (e.g., surface reflectance, Angström exponent). All data are publicly accessible via NASA’s Level-1 and Atmosphere Archive & Distribution System (LAADS) at https://ladsweb.modaps.eosdis.nasa.gov/ (accessed on 1 January 2024), stored in HDF4/5 format.

3. Methodology

A flowchart of our proposed approach is depicted in Figure 2.

(1)

Data preparation: We prepared AHI L1B remote sensing image data taken every ten minutes from 1 January 2016 to 31 December 2016 and AERONET AOD ground observation data in 2016. The AHI L1B data included albedo (reflectance × cos (solar zenith angle (SOZ)) of band 1–band 6), brightness temperature of band 7–band 16, satellite zenith angle, satellite azimuth angle, solar zenith angle, solar azimuth angle, and observation hours (UTC).

(2)

Preprocessing: Cloud and water in the remote sensing image data were removed through the cloud removal algorithm and the water removal algorithm. The outlier value judgment method was used to remove the outliers in the AOD ground observation data.

(3)

Data matching: We selected the remote sensing image data closest to the observation time of the ground observation data from the preprocessed remote sensing image data. The 3 × 3 remote sensing image grid corresponding to the ground observation data was selected through latitude and longitude, and the coefficient of variation in the grid was used to determine whether the remote sensing image grid met the requirements.

(4)

Model establishment: Each band of the 3 × 3 remote sensing image grid was averaged to obtain the training data, which were input into the SDAE model to establish an AOD regression model.

3.1. Prepossessing of the AHI L1B Data

In this study, the L1B remote sensing reflectance data from bands 3, 4, 7, 11, 14, and 15 of the Sunflower-8 geostationary satellite were used for cloud detection for a more accurate cloud mask [22]. The average detection accuracy of the algorithm exceeded 85%. The formula for the cloud mask was as follows, where when one of the conditions was met:

Band 3 > 0.3
Band 14 − Band 15 < −0.5
Band 7 − Band 11 > 10 and Band 4 > 0.1

(1)

Moreover, to remove the influence of water bodies, we used the normalized difference water index (NDWI) to extract the water bodies. When the NDWI was greater than 0.1, the area was judged to be a water body [22]. The NDWI formula is as follows:

$$\mathrm{N}\mathrm{D}\mathrm{W}\mathrm{I}={\displaystyle \frac{\mathrm{B}\mathrm{a}\mathrm{n}\mathrm{d} 2-\mathrm{B}\mathrm{a}\mathrm{n}\mathrm{d} 4}{\mathrm{B}\mathrm{a}\mathrm{n}\mathrm{d} 2+\mathrm{B}\mathrm{a}\mathrm{n}\mathrm{d} 4}}$$

(2)

3.2. AERONET AOD Outlier Removal

The AOD observation range of this network is usually between 0 and 5. Thus, when the AOD observation result was less than 0 or greater than 5, the AOD value was removed as an outlier.

3.3. Data Matching

We performed data matching to eliminate the influence of error from the mismatch between the longitude and latitude of a single pixel and the measured data on the inversion model. We applied a ±10 min time window centered on the satellite overpass times (e.g., Himawari-8 overpasses Wuhan at 03:10 and 03:20 UTC), retaining only ground-based observations with ≤10 min time differences from satellite acquisitions to minimize errors from aerosol temporal variability. We used the longitude, latitude, and time information from the AERONET AOD data as the basis and searched for the pixel row and column number of each AHI L1B image with the same or similar time as the AERONET AOD data in the AHI L1B database. Then, we constructed a 3 × 3 grid with this row and column number as the center and selected a 9-pixel grid range as the input data for inversion. If the coefficient of variation in each band of the L1B data within the 9-pixel cell was less than 0.15, the data were considered to meet the inversion requirements; otherwise, the data were removed. A $C_v$ threshold of 0.15 effectively balanced the trade-off between data retention and noise suppression in aerosol retrieval [8]. The coefficient of variation formula is as follows:

$$C_v={\displaystyle \frac{\sigma }{\mu }}$$

(3)

where $\sigma$ is the standard deviation of the 3 × 3 window of a given band of the AHI L1B remote sensing data, and $\mu$ is the average value of the 3 × 3 window of that band.

3.4. SDAE Model

Deep learning methods are widely used in the field of remote sensing [45,46]. In this paper, we investigated AOD remote sensing inversion via an SDAE algorithm based on a deep learning framework. The Stacked Denoising AutoEncoder (SDAE) was formed by stacking multiple denoising autoencoders [47]. Figure 3 shows a schematic diagram of a single denoising autoencoder. The denoising autoencoder was trained first by random mapping:$\tilde{x}~q_D\left(\tilde{x}\right|x)$. The corrupted input was then remapped to the hidden layers as in the basic autoencoder, represented as $y=f_{\theta }\left(\tilde{x}\right)$. Setting $z=g_{{\theta }^{\prime }}\left(y\right)$, $\theta ,$ and ${\theta }^{\prime }$ were learned through the minimization of the average error on the training dataset, which meant making $z$ as close as possible to the original input $x$.

3.5. AHI AOD Model Estimation via the SDAE

In this study, the AHI L1B images corresponding to the data of 24,413 AERONET point pairs were extracted in 3 × 3 windows (including 16 bands’ data, satellite zenith angle, satellite azimuth angle, solar zenith angle, solar azimuth angle, and observation hours (UTC)), the corresponding hour, month, latitude, and longitude data were recorded, and the data were finally merged into a dataset with 25 feature inputs.

Based on the input feature dataset, we established a deep learning network framework of 25-25-25-25-1 (25 nodes for the input layer and 3 layers for the self-encoder layer [47,48], 25 nodes per layer, and 1 node for the output layer), as shown in Figure 4. First, the established SDAE deep learning network was pretrained. After pretraining, the known AOD data were used for supervised fine-tuning.

4. Results

4.1. Model Evaluation

In this study, the 10-fold cross-validation method was utilized to test for overfitting and the predictive power of the model.

(a)

Correlation Coefficient ($\mathrm{R}$):

The correlation coefficient is a measure of the strength and direction of the linear relationship between two variables. Its value is between −1 and 1. When $\mathrm{R}$ is close to 1, it indicates that there is a strong positive correlation between the two variables. When $\mathrm{R}$ approaches −1, it indicates a strong negative correlation. When $\mathrm{R}$ is close to 0, it means that there is almost no linear relationship between the two variables.

$$\mathrm{R}={\displaystyle \frac{{\sum }_{i=1}^n(x_i-\overline{x})(y_i-\overline{y})}{\sqrt{{\sum }_{i=1}^n{(x_i-\overline{x})}^2{\sum }_{i=1}^n{(y_i-\overline{y})}^2}}}$$

(4)

where $x_i$ and $y_i$ are the observed values of the two variables, $\overline{x}$ and $\overline{y}$ are their mean, and n is the number of observations.

(b)

Mean Relative Error ($\mathrm{M}\mathrm{R}\mathrm{E}$)

The $\mathrm{M}\mathrm{R}\mathrm{E}$ is a measure of the relative difference between a predicted value and an actual value. It is calculated by dividing the error (the difference between the predicted and actual values) of each predicted value by the actual value and then taking the average of these relative errors.

$$\mathrm{M}\mathrm{R}\mathrm{E}={\displaystyle \frac{1}{n}}\sum _{i=1}^n\left|{\displaystyle \frac{y_i-\widehat{y_i}}{y_i}}\right|$$

(5)

where $y_i$ is the actual value, $\widehat{y_i}$ is the predicted value, and n is the number of observations.

(c)

Root Mean Square Error ($\mathrm{R}\mathrm{M}\mathrm{S}\mathrm{E}$)

The root mean square error is a measure of the difference between the predicted value and the actual value. It is calculated by squaring the errors of each predicted value, then taking the average of those squared errors, and finally taking the square root.

$$\mathrm{R}\mathrm{M}\mathrm{S}\mathrm{E}=\sqrt{{\displaystyle \frac{1}{n}}{\sum }_{i=1}^n{\left(y_i-\widehat{y_i}\right)}^2}$$

(6)

where $y_i$ is the actual value, $\widehat{y_i}$ is the predicted value, and n is the number of observations.

(d)

Within $EE$.

This metric is typically used to measure how much of a forecast falls within a predetermined margin of error (±5%).

$$\mathrm{w}\mathrm{i}\mathrm{t}\mathrm{h}\mathrm{i}\mathrm{n} \mathrm{E}\mathrm{E}={\displaystyle \frac{Number of \widehat{y_i} such that \left(1-EE/100\right)\le \frac{\widehat{y_i}}{y_i}\le \left(1+EE/100\right)}{n}}$$

(7)

where $y_i$ is the actual value, $\widehat{y_i}$ is the predicted value, $n$ is the number of observations, and $EE$ is the percentage of the error boundary.

The coefficient of correlation (R), $\mathrm{M}\mathrm{R}\mathrm{E}$, $\mathrm{R}\mathrm{M}\mathrm{S}\mathrm{E}$, and within expected error (within $\mathrm{E}\mathrm{E}$, Equation (7)) were applied as statistical indicators. In Equation (8), $\tau$ represents the AOD result at 500 nm, and $∆\tau$ represents the estimated uncertainty, which is the systematic error of AERONET. Within $EE$ means that the result falls within the allowable error range; in a similar theory, above $EE$ means that the result is above the allowable error range, and below $\mathrm{E}\mathrm{E}$ means that the result is below the allowable error range.

$$∆\tau =\pm 0.05\pm 0.15\tau$$

(8)

In selecting baseline models for comparison, we prioritized methods that (a) represented distinct categories of machine learning paradigms and (b) had demonstrated success in prior aerosol remote sensing studies. In this study, the models of Extreme Learning Machine (ELM), BackPropagation Neural Network (BPNN), Support Vector Machine (SVM), and Generalized Regression Neural Network (GRNN) were utilized to compare with SDAE.

ELM is a single-hidden-layer neural network with random fixed weights in the hidden layer. Output weights are solved analytically for fast training. It is robust to noise, ideal for real-time tasks.

BPNN uses gradient descent and error backpropagation to update weights. It handles complex nonlinear patterns but converges slowly and risks overfitting.

SVM finds optimal hyperplanes using kernels (e.g., RBF) for classification/regression. It is strong in small-sample, high-dimensional scenarios but computationally intensive.

GRNN implements nonparametric regression based on kernel density estimation. It provides instant training but is memory-heavy. It suits smooth continuous predictions and is sensitive to outliers.

Table 3. Performance of the various models.

Method	R	MRE	RMSE	Within EE
ELM	0.74	104%	0.21	43%
BPNN	0.87	76%	0.18	52%
SVM	0.91	49%	0.14	79%
GRNN	0.94	36%	0.09	85%
SDAE	0.98	26%	0.06	92%

shows the performance of the various models. In the cross-validation results of model fitting, the $\mathrm{R}$ values ranged from 0.74 to 0.98, RMSEs from 0.06 to 0.21, MREs from 26% to 104%, and EEs from 43% to 92%. The ELM model performed the worst, with a $\mathrm{R}$ value of 0.74. As one of the intelligent algorithms, the BPNN model could better represent the relative relationship, with $\mathrm{R}$ and $\mathrm{R}\mathrm{M}\mathrm{S}\mathrm{E}$ values of 0.87 and 0.18, respectively. Compared with the results of the BPNN model, R increased by 0.04 (from 0.87 to 0.91) and $\mathrm{R}\mathrm{M}\mathrm{S}\mathrm{E}$ decreased by 0.04 (from 0.18 to 0.14) for the cross-validation of the SVM model. The GRNN model performed excellently with $\mathrm{R}$ and $\mathrm{R}\mathrm{M}\mathrm{S}\mathrm{E}$ values of 0.94 and 0.09, respectively. The SDAE model performed the best, with $\mathrm{R}$ and $\mathrm{R}\mathrm{M}\mathrm{S}\mathrm{E}$ values of 0.98 and 0.06, respectively.

These findings suggest that the proposed SDAE model performed the best, followed by GRNN, and then SVM and BPNN, whereas the ELM model achieved the worst performance at full scale. In this study, SDAE could be applied to retrieve AOD from the AHI onboard satellite.

4.2. Validation of the Retrieved AOD

4.2.1. Comparison of JAXA AOD with Ground-Observed AOD

Compared with those of the other models, the AOD results retrieved via the SDAE yielded good results. Therefore, this model was used in this study. The ground-observed AOD at the Wuhan University site was utilized to verify the SDAE satellite-retrieved AOD. The satellite-retrieved AOD and ground-observed AOD data from January to December 2016, approximately 6000 data points, are shown in Figure 5a. Based on the validation results, the coefficient of correlation ($\mathrm{R}$) was relatively high (approximately 0.8); the MRE was 29.67%; the $\mathrm{R}\mathrm{M}\mathrm{S}\mathrm{E}$ was 0.26; the within $\mathrm{E}\mathrm{E}$ was 57%; the slope was on the verge of 1; and the intercept was only 0.02. The error distribution histogram between the ground-observed AOD and satellite-retrieved AOD is shown in Figure 5b. The average error was 0.13, and the standard deviation of the error was 0.20.

The accuracies of the satellite-retrieved AODs were verified hourly and monthly. Figure 5a shows the SDAE model validation results at the Wuhan University site. Figure 5b shows the JAXA AOD validation results at the Wuhan University site. Figure 6 shows the pixel-by-pixel validation of JAXA AOD using MODIS AOD at 6:30 (UTC) on 22 September 2016, in the Hubei province, as well as the validation of satellite-retrieved AOD using MODIS AOD.

Compared to previous work [30,49], the accuracy of the SDAE model in AOD inversion was significantly improved. This might have been due to the SDAE model’s stronger feature extraction and noise reduction capabilities, which could more effectively extract AOD information from satellite data. Compared with the SDAE model, there was a certain gap in the accuracy of JAXA AOD (JAXA AOD was inverted by the DT algorithm). This might have been due to the differences in data processing, feature extraction, etc., between different models. At the same time, it showed that the degree of consistency between the AOD inverted by the SDAE model in this paper and the MODIS AOD was high.

4.2.2. Hourly AOD Validation

The hourly daytime AOD validation results from January to December 2016 are presented in Figure 7. The R values of each hourly daytime AOD from 0:00 (UTC) to 8:00 (UTC) were greater than 0.8, the MRE values were less than approximately 25% (except at 2:00 (UTC) and 5:00 (UTC), when the MREs were 26.54% and 25.44%, respectively). At midday, solar radiation is strong and surface temperatures rise, which could lead to the vertical diffusion and redistribution of near-surface aerosols, which could affect the observed values of AOD. Therefore, the RMSE value was less than 0.3 (except at 10:00, when the RMSE was 0.3), and the within EE value was greater than 60% (except at 13:00, when the within EE was 58.08%). The slope was between 0.94 and 1.12, and the intercept was between −0.07 and 0.04.

Figure 8 illustrates the hourly variations in AOD during daytime. The ground-observed AOD and satellite-retrieved AOD exhibited a strong correlation, with R of 0.8, RMSE of 0.06, and MRE of 10.06%. The ground-observed AOD measurements generally fell within the error band of the satellite-retrieved AOD. As shown in Figure 8, both ground-observed and satellite-retrieved AOD values demonstrated relatively stable variations, indicating that neither temporal factors nor solar zenith angle significantly influenced the measurements. This consistency confirmed the validity and reliability of the AOD retrievals.

4.3. Hourly Patterns of AOD Distribution

In this study, the SDAE satellite-retrieved AOD was applied to study the hourly patterns of the AOD distribution in the Hubei province. From May to August, the Hubei province enters the rainy season, and from October to March of the following year, the north of this province is prone to being affected by wind and sand. Therefore, the sunny weather in September was chosen for this study. The hourly variation in AOD on 22 September 2016 was selected for analysis, as shown in Figure 9. The satellite-retrieved AOD changed smoothly in the Hubei province. As early as 00:00 (UTC), in the central cities of the Hubei province, such as Jingmen city, Yichang city, Qiangjiang city, and Jingzhou city, the AOD was high. The aerosol then spread to other cities with the wind. Therefore, the AOD increased in Shiyan city, Xiangyang city, and Enshi city, reaching a relatively high value at 02:00 (UTC). As the sun rose, the temperature gradually rose, the water vapor in the air gradually disappeared, and the AOD decreased by noon, especially in Shiyan city, Shenlongjia city, Enshi city, and Huanggang city. Afterward, the wind direction began to change, and an east wind emerged from the eastern Hubei province, while a northwest wind appeared from Xiangyang city and Enshi city, which was accompanied by the AOD increasing in the central city of the Hubei province again at 06:00 (UTC). Finally, the AOD was spread to the southern province with a northerly wind that emerged from Xiangyang city at 07:00 (UTC).

4.4. Discussion

The purpose of this study was to develop a novel AOD inversion method that could overcome the influence of geographical and external interference factors. In this study, we used a SDAE combined with the AHI data of the Himawari-8 satellite and successfully obtained high-precision AOD measurements in the Hubei province by inputting AHI images, months, hours, latitude, and longitude. Comparison with traditional machine learning methods shows that the SDAE method had the highest correlation coefficient and value within the minimum expected error in AOD inversion. At the same time, it exhibited the lowest MRE and RMSE, and the satellite-retrieved AOD showed high consistency with the ground-based AOD measurements at the Wuhan University site.

5. Conclusions

The Himawari-8 is equipped with the AHI, which has a 2 km spatial resolution and a temporal resolution of 10 min. This provided the possibility to monitor near-real-time changes in the atmospheric environment.

In the official JAXA product, although an AOD product can be obtained, the AOD cannot be retrieved from bright areas in an image due to the limitations of the DT algorithm used. The AOD from JAXA was compared with long-term Wuhan University site data, with an R of only 0.7 and a within EE of only 57%. To retrieve more accurate and effective AOD data from satellites, inversion data including L1B data from the AHI, imaging geometry, imaging time, imaging month, and pixel point latitude and longitude were used, and AERONET-measured site data were used for verification of the inversion result. The main method used in this study was the application of an SDAE, which was then compared with various other methods, namely, ELM, BPNN, GRNN, and SVM algorithms. The results revealed that the SDAE had a relatively high R (0.97) and within EE (67.84%) and low MRE (21.32%) and RMSE (0.2) values. Thus, the SDAE can be applied to retrieve AODs from the AHI onboard the Himawari-8 satellite.

To verify the applicability of the model in the Hubei province, estimation validation was performed. The accuracy of the satellite-retrieved AOD was verified on hourly, daily, and monthly data in this study. In the monthly AOD validation, the satellite-retrieved AOD was verified to be effective from January to December 2016, as the R ranged from 0.76 to 0.93, the MRE ranged from 15.89% to 25.37%, the RMSE ranged from 0.10 to 0.27, the within EE ranged from 56.27% to 81.96%, and the satellite-retrieved AOD and ground-observed AOD tended to be consistent. For the daily average AOD validation, the satellite-retrieved AOD and ground-observed AOD were compared from January to December 2016, and the R value was 0.91, the MRE value was 18.13%, the RMSE value was 0.21, and the within EE was 70.97%. For the hourly AOD validation, the satellite-retrieved AOD and ground-observed AOD from January to December 2016 from 8:00 to 16:00 were divided into eight groups, and the R value ranged from 0.82 to 0.91, the MRE ranged from 18.59% to 26.56%, the RMSE ranged from 0.16 to 0.30%, the EE ranged from 58.08% to 73.86%, and the average hourly AOD agreed. Finally, the hourly patterns of AOD distribution on 22 September 2016 in the Hubei province were analyzed to study the AOD changes in this region.

In summary, high-spatial-temporal resolution AOD data can continuously and intuitively reflect the characteristics of regional aerosol particles (such as diffusion and accumulation), which is highly important for assessing regional pollutants and air quality.

In this study, the AOD observed on the ground was unstable at small values due to the insufficient sensitivity of the measuring instrument. Nonetheless, this points us to the direction of future research. This method may demonstrate better applicability in the Hubei province, and further research may be required for its adaptation to other regions.

At the same time, although the AOD inverted using the deep learning algorithm framework SDAE has not been fully studied, this is the starting point of our exploration. In the future, we will continue to explore the application potential of other deep learning algorithms in AOD inversion to continuously improve the accuracy and reliability of AOD measurements.