Aerosol Information Retrieval from GF-5B DPC Data over North China Using the Dark Dense Vegetation Algorithm

Abstract

A directional polarimetric camera (DPC) is a key payload on board China’s Gaofen 5B (hereafter denoted as GF-5B) satellite, a hyperspectral observation instrument for monitoring aerosols. On the basis of the dark dense vegetation (DDV) algorithm, this study applied DDV algorithm to DPC measurements. First, the reflectance of vegetation in three channels (0.443, 0.49, and 0.675 μm) was analyzed, and inversion channels were identified. Subsequently, the decrease in normalized difference vegetation index associated with various view angles was simulated, and the optimal view angle for extracting dark pixels was determined. Finally, the top-of-atmosphere reflectance at different view angles was simulated to determine the optimal view angle for aerosol inversion. The inversion experiments were conducted by using DPC data collected over North China from November 2021 to January 2022. The results revealed that DDV algorithm could monitor pollution from 30 December 2021 to 4 January 2022, and the inversion results were strongly correlated with Moderate Resolution Imaging Spectroradiometer (MODIS) aerosol product and AERONET station data (R > 0.85).

1. Introduction

On 7 September 2021, China launched a hyperspectral observation satellite named Gaofen 5B (hereafter denoted as GF-5B) from the Taiyuan Satellite Launch Center. Serving as a successor to the Gaofen-5 (GF-5) satellite, GF-5B is also equipped with three payloads: an environmental trace gas monitoring instrument, a directional polarimetric camera (DPC), and a greenhouse gas monitoring instrument. The DPC receives signals in eight channels from the visible to the near-infrared spectral region, and three channels, namely 490, 660, and 865 nm, are used for polarization observations. The spatial resolution at nadir is 3.3 km. Compared with that of the camera onboard the GF-5 satellite (13), the maximum view angle of the DPC on board GF-5B has increased to 17 (

Table 1. Main parameters of DPC/GF-5B, DPC/GF-5, and POLDER-3.

Characteristics	DPC/GF-5B	DPC/GF-5	POLDER3
Channel (P is polarization)	443 nm, 490 nm (P), 565 nm, 670 nm (P), 763 nm, 765 nm, 865 nm (P), 910 nm	443 nm, 490 nm (P), 565 nm, 670 nm (P), 763 nm, 765 nm, 865 nm (P), 910 nm	443 nm, 490 nm (P), 565 nm, 670 nm (P), 763 nm, 765 nm, 865 nm (P), 910 nm, 1020 nm
Polarized angle	0°, 60°, 120°	0°, 60°, 120°	0°, 60°, 120°
Number of directions	≤17	≤13	≤17
resolution at nadir	≤3.3 km	≤3.3 km	6 × 7 km
FOV (Field of View)	≥50°	≥50°	≥51°

). Moreover, compared with that of French Polarization and Directionality of the Earth’s Reflectances (POLDER-3) instrument on board the Parasol microsatellite (6 km), the spatial resolution of the DPC on board GF-5B has increased to 3.3 km. GF-5B is also equipped with particulate observing scanning polarimeter (POSP) and atomic absorption spectroscopy payloads; these payloads, in conjunction with the DPC, considerably enhance China’s ability to monitor atmospheric environmental quality through remote sensing and provide an additional remote sensing data source for large-scale monitoring of aerosols and particulate matter in China.

Satellite remote sensing technology can quickly obtain atmospheric information such as aerosols in a large range. Many scholars have proposed a variety of algorithms to remove the influence of the surface and retrieve the terrestrial aerosol. There are mainly the following:

(1)

Dark target method. The reflectance of thick vegetation is low in the visible band such as red light and blue light, and the linear relationship between the short-wave infrared band and red light and blue light is assumed. The dark target method has been well applied in the data of MODIS, AVHRR/NOAA, MERIS/ENVISAT, VIIRS/S-NPP, WFV/GF-1, MERSI/FY-3D, and so on [1,2,3,4,5,6,7,8,9,10,11];

(2)

Deep blue method. According to the principle that the reflection of ground objects in the blue band is weak and the aerosol optical thickness makes a significant contribution to the satellite observation signal, the deep blue algorithm can be well applied to MODIS and Sea WIFS data [12], and also successfully applied to LANDSAT, NOAA, GF-1, INSAT-3D and other satellites [13,14,15,16,17]. DB method has also been applied on VIIRS instrument and delivery NRT product [18,19];

(3)

Multiple Angle method. Assuming that the surface reflectance ratio of the same surface object in multiple observation directions does not change with wavelength, the land aerosol was obtained by removing the surface influence from the two Angle data of ATSR-2, MISR and AATSR [20,21,22,23];

(4)

Multi-angle polarization algorithm. Since polarization signals are sensitive to aerosol physical properties and the surface polarization reflection is small and stable, the surface reflection can be better removed for retrieving aerosol modal and AOD information [24]. The PODLER series was the first satellite payload to use polarization observation for aerosol monitoring [25], which realized the inversion of AOD, complex refractive index, single scattering albedo and other parameters [26,27]. The algorithms applied to POLDER aerosol microphysical retrieval include GRASP, SPON, statistical optimization inversion algorithm and so on [28,29,30,31,32,33,34,35]. The DPC camera mounted on China’s Gaofen-5 satellite can also use polarization observation for aerosol inversion [36,37]. Recently, some research on DPC has been applied to retrieve the content and optical properties of terrestrial aerosol components, retrieve high-precision and high-resolution aerosol products [33,34,35,38,39].

Currently, the DDV algorithm is the most commonly used method for terrestrial aerosol inversion. The DPC not only has three polarization observation channels (i.e., 0.49, 0.67, and 0.865 μm) but also obtains scalar data at multiple angles in nine channels ranging from 0.443 to 0.910 μm. It is a suitable data source for the detection and retrieval of aerosol information through the DDV algorithm.

This study analyzed aerosol measurements at different angles and applied DDV algorithm to DPC data. The DDV algorithm was applied to DPC data obtained in North China, and the MODIS aerosol product and AERONET data were compared with the DPC results.

2. Materials and Methods

For plane-parallel atmospheres, if gaseous absorption is ignored, the top-of-atmosphere (TOA) reflectance ${\rho }_{TOA}$ can be expressed as follows [40]:

$${\rho }_{\mathrm{TOA}}={\rho }_0({\theta }_{\mathrm{s}},{\theta }_{\mathrm{v}},\phi )+\frac{T({\theta }_{\mathrm{s}})T({\theta }_{\mathrm{v}}){\rho }_{surface}}{[1-{\rho }_{surface}S]}$$

(1)

where ${\rho }_0$ represents the atmospheric path reflectance, T represents atmospheric transmission, S represents the atmospheric backscatter ratio, and ${\rho }_{surface}$ represents the surface reflectance. $\varphi$ denotes the relative azimuth angle between the sun and the sensor, and θ denotes the zenith angle, where s is the sun and v is the sensor view angle.

In Equation (1), ${\rho }_0$, T, and S can be determined using information on atmospheric molecules and aerosols. Generally, the optical depth of atmospheric molecules ${\tau }_{R,\lambda }$ can be calculated using the altitude Z and wavelength λ [41].

$${\tau }_{R,\lambda }=0.00877{\lambda }^{-4.05}exp\left(-\frac{Z}{8.5}\right)$$

(2)

Over land surfaces, the AOD and aerosol type vary because of factors such as differences in aerosol sources and atmospheric transport. For aerosol inversion, the aerosol type is usually determined using ground measurements. In the paper, aerosol type was determined by three aerosol components, and the proportions of the three aerosol components are 18% dust-like, 80% water-soluble, and 2% soot [42]. For accelerating retrieval speed, atmospheric parameters were calculated by 6SV 1.0 [43] and stored in lookup table (LUT). In the LUT, the AOD was from 0 to 2 with the interval of 0.2, the solar and sensor zenith angles were from 0.0 to 66.0 degree with the interval of 6 degree, and relative azimuth angles were from 0 to 180 degree with the interval of 12 degree.

The DDV algorithm is based on the assumption that over dark pixels (dense vegetation), the surface reflectance exhibits an empirical relationship between the red and blue channels [44] or between the visible (red and blue) and short-wave infrared channels [1].

When the DDV algorithm was applied to DPC data, dark pixels were selected using the normalized difference vegetation index (NDVI), and the surface reflectance relationship between the blue (0.443 and 0.49 μm) and red channels was examined using ground spectral measurements.

$$NDVI=\frac{{\rho }_{nir}-{\rho }_{\mathrm{red}}}{{\rho }_{nir}+{\rho }_{\mathrm{red}}}$$

(3)

$${\rho }_{\mathrm{red}}=a\times {\rho }_{\mathrm{blue}}+b$$

(4)

where ${\rho }_{nir}$ is the reflectance in the near-infrared channel, ${\rho }_{\mathrm{red}}$ is the reflectance in the red channel, and ${\rho }_{\mathrm{blue}}$ is the reflectance in the blue channel.

Before applying the DDV algorithm to DPC data, we masked cloud, selected the appropriate channel for aerosol information retrieval, determined the optimal NDVI for identifying dark pixels, and determined the optimal measurement angle.

(1)

Masking cloud. Cloud were masked through the threshold. It is shown that, in the red channel, cloud-free reflectance is lower than 0.18 [45]. In the paper, when apparent reflectance was greater than 0.18, the pixel was masked as cloud.

(2)

Selecting appropriate channel. The DPC receives signals in two blue channels: 0.443 and 0.490 μm. In 2008, we measured the spectra of several areas of vegetation by using an analytical spectral device (ASD) in Beijing in October. By integrating ASD data with the DPC filter response function, we could obtain 324 sets of surface reflectance measurements in the 0.443-, 0.49-, 0.675-, and 0.865-μm channels of the DPC. The scale for the red and blue channels was approximately 2, and the scale for the 0.443-μm channel was greater than that for the 0.49-μm channel (Figure 1). Moreover, the relationship between the 0.49- and 0.675-μm channels was stronger than that between the 0.443- and 0.675-μm channels. Accordingly, this study used the 0.490 μm channel for retrieving aerosol measurements collected in the 0.675 μm channel. As indicated in (4), the scale $a$ was set to 1.8875, and the offset $b$ was set to −0.0161.

(3)

Selecting dark pixels. This study used the NDVI to identify dark pixels. On the basis of (3), the NDVI values for different vegetation areas were calculated using surface reflectance measurements captured in the 0.675 and 0.865 μm channels. As displayed in Figure 2, the NDVI values for most vegetation areas were between 0.5 and 0.9. The data applied for NDVI calculation were from ASD measurements above.

Because of aerosol extinction, the NDVI decreases as the AOD increases. The DPC has more than 10 view angles. Using 6SV 1.0 [43], we simulated the decrease in NDVI associated with different view angles. Usually, view angles do not change too much in different region and seasons. However, sun angles are different in different regions and seasons. The paper was focused on the aerosol in northern China. In December, sun zenith angle is largest, while it is smallest in June. We chose two viewing geometry in December and June to simulate the satellite signals. The viewing geometry for 17 angles is presented in

Table 2. Viewing geometry for simulation in December.

No.	Sun Zenith Angle	View Zenith Angle	Relative Azimuth Angle	Scattering Angle	Choose for Show
1	63.48	54.23	153.13	155.31	Yes
2	63.48	50.05	153.22	154.04	No
3	63.46	45.25	153.61	152.09	No
4	63.44	39.75	154.22	149.18	No
5	63.44	33.44	154.78	144.99	Yes
6	63.42	26.39	155.88	139.78	No
7	63.41	18.44	158.52	133.46	No
8	63.4	10.05	165.01	126.27	No
9	63.4	2.32	134.48	118.21	Yes
10	63.38	8.53	43.74	110.32	No
11	63.36	16.99	36.2	102.59	No
12	63.35	24.92	33.24	95.27	Yes
13	63.35	32.19	31.99	88.61	No
14	63.34	38.65	31.33	82.74	No
15	63.32	44.23	30.81	77.67	No
16	63.31	49.13	30.37	73.20	Yes
17	63.29	53.43	30.18	69.37	No

and

Table 3. Viewing geometry for simulation in June.

No.	Sun Zenith Angle	View Zenith Angle	Relative Azimuth Angle	Scattering Angle	Choose for Show
1	21.45	54.98	123.43	134.36	Yes
2	21.42	50.85	123.75	138.17	No
3	21.4	46.11	124.07	142.45	No
4	21.35	40.75	124.73	147.24	No
5	21.33	34.58	125.33	152.41	Yes
6	21.29	27.62	126.49	157.77	No
7	21.26	19.77	128.89	162.56	No
8	21.22	11.45	134.15	164.51	No
9	21.19	3.28	189.33	162.04	Yes
10	21.16	7.22	76.02	156.10	No
11	21.12	15.65	65.73	149.13	No
12	21.1	23.82	62.42	141.83	Yes
13	21.07	31.24	61.04	135.04	No
14	21.04	37.79	60.19	128.92	No
15	21	43.48	59.6	123.55	No
16	20.96	48.53	59.18	118.76	Yes
17	20.94	52.89	58.75	114.56	No

; only 5 angles are included in this table for ease of display. Because the simulation results changed slowly with angles, the observation results of adjacent angles were relatively similar. The selected five angles were with the interval of about 20 degrees. As illustrated in Figure 3, the NDVI decreased as the AOD increased. In addition, as the zenith angle increased, the atmospheric scattering increased and the NDVI decreased. Therefore, the NDVI associated with the smallest zenith angle was used for the identification of dark pixels. For angle 9 (the smallest zenith angle), as AOD increased to 2, the NDVI decreased to about 0.2 in December and 0.4 in June. Therefore, the NDVI threshold was set to different values in different seasons. We set the threshold of NDVI according to the reflectance of the smallest zenith angle for AOD of 2. In winter, the threshold was 0.2, while that was 0.38 in summer. In the future, we will simulate the decreased NDVI of different solar zenith angles, and set the threshold with the solar zenith angle.

(4)

Selecting measurement angle. Because the DPC on board GF-5B observes its target from 17 angles, we selected the most aerosol-sensitive angle for aerosol measurement. We simulated TOA reflectance in the blue (0.49 μm) and red (0.675 μm) channels, as presented in Figure 4 and Figure 5, respectively. The viewing geometry is also shown in Table 2. Increases in TOA reflectance were greatest from angle 1. In the blue channel, as the AOD increased from 0 to 2, the TOA reflectance from angle 1 increased from 0.22 to 0.50 in December, while that increased from 0.11 to 0.24 in June. In the red channel, the TOA reflectance increased from 0.12 to 0.47 in December, while that increased from 0.07 to 0.17 in June. This was because angle 1 was associated with the largest view zenith angle and the longest sunlight transmission path; therefore, the angle with the largest view zenith angle was selected as the measurement angle.

3. Results and Validation

We retrieved AOD images of North China captured by the DPC on board GF-5B between 1 November 2021, and 31 January 2022.

(1)

Retrieving AOD images during pollution event. In the beginning of January 2022, an aerosol pollution event occurred in North China. We monitored the contamination process using DPC images. As depicted in Figure 6, on 1 January 2022, the overall AOD in North China began to increase, with obvious AOD growth over the border of Shandong Province and Hebei Province and the eastern part of Hebei Province. On 3 January 2022, obvious pollution was observed in the south of North China, and the overall AOD in Shandong was >1. On 4 January 2022, the pollution event continued, and the AOD also began to increase in Beijing and western Hebei; however, most of the polluted areas in the south could not be monitored because of cloud cover.

(2)

Comparing of inversion results with MODIS aerosol product data. We resampled the aerosol measurements from GF-5B to the same spatial resolution (10 km) as the MODIS aerosol product (MOD04 Collection 6.1) [46,47,48]. Subsequently, the GF-5B aerosol measurements obtained on 6 December 2021, were compared with the MODIS aerosol product data. As presented in Figure 7, the aerosol measurements obtained from the DPC were consistent with distribution of the MODIS aerosol product data, and a high AOD was detected in eastern and southern Shandong. However, the aerosol levels detected using our DPC data were lower than those indicated by the MODIS aerosol product. In addition, in northeastern Shandong, some cloud regions were not successfully identified, resulting in numerous regions with high pixel values. By comparing the valid data from MODIS and the DPC inversion results, 4553 effective pixel results were obtained in Figure 8. The mean value, maximum value, minimum value, and standard deviation of the absolute difference between the MODIS aerosol product data and the DPC inversion results were −0.06, 0.94, −1.48, and 0.13, respectively. The mean value, maximum value, minimum value, and standard deviation of the relative difference between the MODIS aerosol product data and the DPC inversion results were 6%, 5714%, −86%, and 44%, respectively. The correlation coefficient between the DPC measurements and MODIS aerosol product was 0.87, the absolute error was 0.2, and the relative accuracy was 49%.

(3)

Validating the data by using AERONET data. The GF-5B aerosol measurements obtained from 1 December 2021 to 25 January 2022, were compared with those from of the AERONET Beijing-RADI station. The inversion results were obtained in 34 days. The mean value, maximum value, minimum value, and standard deviation of the absolute difference between the measurements were −0.06, 0.03, −0.35, and 0.07, respectively. The mean value, maximum value, minimum value, and standard deviation of the relative difference between the measurements were −32%, 27%, −86%, and 26%, respectively. As shown in Figure 9, the correlation coefficient between the DPC aerosol inversion results and the AERONET measurements was 0.97, the absolute error was 0.06, and the relative accuracy was 32%. Our results underestimated because the scale was 0.7878. The reason may be the lack of vegetation in winter. Although the mean AOD diff. (DPC-MODIS) and (DPC-AERONET) was −0.06, our results were closed to AERONET because of the higher R.

4. Conclusions

This study used DPC data to develop a terrestrial aerosol inversion algorithm. The main procedures and findings of the study are outlined as follows:

(1)

The TOA reflectance of typical vegetation in different channels and different AODs were analyzed, and the DDV algorithm was applied to DPC data.

(2)

The algorithm was tested on data collected in North China, and the results reveal that the process of air pollution can be accurately measured using DPC data.

(3)

Comparing the results obtained using our algorithm with the MODIS aerosol product and AERONET ground observation data showed that the results obtained using our algorithm were strongly correlated with the MODIS aerosol product and AERONET ground observation data, although the overall results obtained using our algorithm were lower than the MODIS aerosol product and AERONET ground observation data.

Although our aerosol inversion algorithm can help measure the AOD, it still has room for improvement. Possible areas of improvement that can be considered in future research are outlined as follows:

(1)

More view angles should be introduced to more effectively remove surface effects and improve the accuracy of AOD inversion.

(2)

The DDV and Deep Blue algorithms should be combined to expand the application possibilities of DPC data.

(3)

Polarization observation data should be introduced for simultaneous inversion of AOD and modal aerosol data.