Retrieval of Aerosol Optical Properties over Land Using an Optimized Retrieval Algorithm Based on the Directional Polarimetric Camera

Abstract

The Directional Polarization Camera (DPC) onboard the Chinese Gaofen-5 satellite, launched in May 2018, has similar specifications as the POLDER-3 instrument. The SRON Remote Sensing of Trace gas and Aerosol Products (RemoTAP) full retrieval algorithm is applied to DPC measurements to retrieve aerosol properties including the total Aerosol Optical Depth (AOD), the fine/coarse mode AOD and the SSA (Single Scattering Albedo). Measurements of the global ground-based AERONET network between December 2019 and April 2020 have been used for the validation of the DPC retrievals. According to the average Fine Mode Fraction (FMF) of the selected AERONET stations, the stations are divided into urban stations (FMF ≥ 0.5) and dust stations (FMF < 0.5). For the total AOD validation, DPC retrievals show better performance over urban stations than over dust stations, with average biases of 0.055 and 0.106, and RMSEs of 0.151 and 0.228, respectively. Regarding the fine mode AOD, the retrieval also performs better over urban stations. Compared with the total AOD validation, both the relatively lower bias (0.021 and 0.065) and the higher Gfrac (Fraction of Good retrieval, 63.8% and 47.3%, respectively) further indicate that DPC performs better when fine mode aerosols dominate. For the land SSA validation, most of our SSA retrievals (~71%) show differences with AERONET SSA retrievals lower than 0.05. Case studies over fire spots and dust over northern China demonstrate the encouraging application potential of DPC aerosol products. The difference between fine and coarse AOD can provide more aerosol source information compared with the total AOD alone. Since the SSA retrievals are particularly sensitive to absorbing fine particles, they can be easily used in the tracking of biomass burning aerosol.

1. Introduction

Atmospheric aerosol distributions are complex, being influenced by both natural phenomena and human activity, with urban areas in particular being dominated by anthropogenic aerosols released from diesel-powered vehicles and industrial activity [1]. Increased emissions of aerosol particles can lead to increased concentrations of hazardous air pollutants. In particular, fine particles increase the mortality rate of patients suffering from heart and/or lung diseases since these fine particles penetrate deeper into the lungs than coarser inhalable particles can, and thus have a more severe impact on human health [2]. Alongside anthropogenic particles causing air pollution, better estimates of the perturbations to the Earth’s radiation budget require accurate optical and physical properties of both fine and coarse aerosol particles. For quantifying the radiative forcing due to aerosol–radiation interactions, accurate measurements of aerosol optical depth (AOD, ±0.02/5%) and single scattering albedo (SSA, ±0.03) are needed [3]. To quantify the Radiative Forcing due to aerosol–cloud interactions (RFaci), retrievals of Cloud Condensation Nuclei (CCN) are needed [4] which can be derived from the effective radius (reff, ±10%) and aerosol column number (±50%). Multi-spectral, multidirectional and polarized satellite observations performed by a Multi-Angle Polarimeter (MAP) contain the richest set of information of aerosols in our atmosphere from a passive remote sensing perspective [5,6,7].

The first MAP instruments are the Polarization and Directionality of Earth’s Reflectance (POLDER) series from 1996 to 2013 [8]. The POLDER-1 and POLDER-2 launched in 1996 and 2002, respectively, failed to obtain stable and effective long-term data due to the failure of solar panels within one year after the launch. The POLDER-3 instrument on the Polarization and Anisotropy of Reflections for Atmospheric Science coupled with Observations from a Lidar (PARASOL) satellite has provided a multi-year data record between 2004–2013 [9]. The Directional Polarimetric Camera (DPC) onboard the Gaofen-5 (GF-5) satellite in orbit between May 2018 and April 2020 is the first Chinese MAP designed for aerosol and cloud detection. It performs measurements in eight spectral channels from 443 nm through to 910 nm with polarization at 490 nm, 670 nm and 865 nm. DPC measures at up to 12 angles [10]. There are four more DPCs with a series of improvements that will be installed on GF-5 (02), CM-1 (Carbon dioxide Monitoring satellite), DQ-1 (DaQi-1, also named AEMS (Atmospheric Environment Monitoring Satellite)) and DQ-2 satellites to be launched successively from 2021 to 2022, respectively [11].

A number of improved MAP instruments are scheduled to be launched over the next 5 years. The Multi-viewing, Multi-channel, Multi-polarization Imaging (3MI) instrument being developed by the European Aeronautics and Space Administration (ESA) is expected to be used for global aerosol and cloud characterization [12,13]. The 3MI instrument has heritage from the POLDER and will provide a multi-spectral (from 410 to 2130 nm), multi-polarization (−60°, 0°, and +60°), and multi-angular (14 views) image of the Earth outgoing radiance at the top of the atmosphere. The SPEXone instrument onboard the NASA Phytoplankton Aerosol Cloud and ocean Ecosystems (PACE) mission to be launched in 2024, will measure radiance and polarization in the spectral range 385–770 nm (50 bands for polarization, 400 bands for radiance) at five viewing angles [14,15]. The spectral modulation technique is used to obtain high polarimetric accuracy (0.003 on DoLP (Degree of Linear Polarization)) [16]. The HARP-2 instrument, also on PACE, will combine data from multiple along track viewing angles (up to 60), four spectral bands in the visible and near-infrared ranges, and three angles of linear polarization.

There are a number of aerosol retrieval algorithms using multi-spectral, multi-angle and polarized measurements [17]. Compared with the algorithms based on pre-simulated lookup tables (LUTs) [18,19,20], full retrieval algorithms need more computing resources but can provide more detailed aerosol and surface properties with higher accuracy. The full inversion algorithms emphasize statistical optimization using the data redundancy available from multi-spectral, multi-directional and polarization measurements such as the Generalized Retrieval of Aerosol and Surface Properties (GRASP) algorithm [21], the SRON RemoTAP algorithm [22,23], and the Jet Propulsion Laboratory (JPL) algorithm [24].

In this work, the SRON RemoTAP retrieval algorithm is applied to DPC measurements to retrieve aerosol properties including the AOD, the fine mode AOD and the SSA. Subsequently, the AERONET aerosol products are used for the validation of the satellite retrievals. To evaluate the application potential of DPC aerosol products, two case studies that include fire and/or dust over northern China are conducted. This paper is organized as follows. Section 2 describes the datasets used in this study. Section 3 introduces the SRON aerosol algorithm. Section 4 contains the aerosol validation results. The case study in northern China is conducted and discussed in Section 5. Finally, the last section summarizes and concludes this study.

2. Data Description

2.1. The Introduction of DPC

The DPC instrument is designed to study the properties of aerosols, clouds and the amount of water vapor by measuring the multi-angular spectral intensity and polarized characteristics of backscattered solar light. It has a sun-synchronous orbit at an altitude of 705 km. The overpass local time is 13:30 p.m. ascending node. The performance specifications of the DPC instrument are shown in

Table 1. The performance specifications of the DPC and POLDER.

Items	DPC	POLDER
Swath(km)	1850	1600
FOV	±50° (Along track/Across track)	±51° (Along track) ±43° (Across track)
Spatial resolution	3.3 km (nadir)	6 km × 7 km (nadir)
pixel number on the CCD	512 × 512	274 × 242
bands (nm, P stands for polarization)	443, 490 (P), 565, 670 (P), 763, 765, 865 (P), 910	443, 490 (P), 565, 670 (P), 763, 765, 865 (P), 910
Bandwidth(nm)	20, 20, 20, 20, 10, 40, 40, 20	20, 20, 20, 20, 10, 40, 40, 20
Polarization direction	0°, 60°, 120°	0°, 60°, 120°
Radiance calibration accuracy	≤5%	2% for shorter wavelength (≤565 nm), 3% for longer wavelength (≥565 nm)
Laboratory calibration accuracy	DoLP: 0.0043, 0.0046, and 0.0037	1%

. The DPC shares a similar design with POLDER. The main differences include: (1) DPC’s spatial resolution is 3.3 km × 3.3 km at nadir, which is higher than POLDER’s pixel size (6 km × 7 km); (2) DPC enables measurements in 8 spectral channels from 443 nm through to 910 nm with the near-infrared channel of 1020 nm excluded. Polarization measurements are performed at 490 nm, 670 nm and 865 nm. (3) DPC provides up to 12 imaging angles (most pixels exceeding 9 angles), slightly less than POLDER’s 16 imaging perspectives. Analysis has concluded that 12 imaging angles can meet the requirement for aerosol retrieval [22,25].

The laboratory calibration system of DPC has been built based on the system used for POLDER-3. An adjustable polarization light source with a large dynamic range is used to provide incident light [26]. The average error in measuring the degree of polarization of a non-polarized light source by all pixels in the three polarized bands is 0.0043, 0.0046, and 0.0037, respectively [27]. Polarization sensors such as POLDER, which do not have on-board calibration devices, generally use the method of observing targets with specific radiation characteristics for on-orbit calibration [28]. The method is also suitable for DPC which means the instrument measurements over specific targets such as sun glint are regularly collected for timely correction of sensor decay.

Stray light is one of the important factors that affect the accuracy of polarization measurement. Huang et al., (2019) proposed a deconvolution method and an extended matrix method to effectively suppress more than 94% of the stray light of the DPC [29]. This stray light correction method has been used to re-process the DPC Level 1b data between December 2019 and April 2020, which are used for validation against the AERONET datasets in this study.

2.2. AERONET Aerosol Products

The retrieved AOD from DPC is validated with AERONET (AErosol RObotic NETwork) level 1.5 data (version 3.0) [30,31]. The data are cloud screened. The uncertainty on AERONET AOD is 0.01 for mid-visible wavelengths and 0.03 for UV wavelengths [32]; it is dominated by a calibration (systematic) error. The AODs of fine modes are compared with AERONET level 1.5 spectral deconvolution algorithm (SDA) data [33]. It should be noted that the inversion and SDA products are quite uncertain themselves at low AOD, so the comparison to these products should not be considered a validation. The retrieved SSA is compared with AERONET level 1.5 almucantar retrieval products of SSA with AOD (675 nm) larger than 0.3. The AERONET SSA itself is not a result from a direct measurement but from an inversion procedure with different kinds of assumptions [34]. The error in the AERONET SSA is at least 0.03 [34]. The global AERONET stations over land with measurements available between December 2019 and April 2020 are used for validation and comparison.

2.3. MODIS and AIRS Products

The MODIS (Moderate Resolution Imaging Spectroradiometer) Fire and Thermal Anomalies product from the Terra (MOD14) and Aqua (MYD14) satellites and the Level 2 dust score from AIRS (Atmospheric Infrared Sounder) are introduced to interpret the DPC SSA variation over fire spots and dust areas. Dust score is determined from multiple tests that compare radiances in select AIRS spectral channels. Radiances measured in dust-sensitive channels are compared to radiances measured in channels that are not sensitive to dust. The differences between several pairs of channels are represented as a dust score.

3. Aerosol Retrieval Algorithm Description

In this study, we employ the SRON RemoTAP retrieval algorithm in a multimode setup [24,35]. The inversion retrieval approach is aimed to invert a forward model equation:

$$y=\mathrm{F}\left(x\right)+e_y$$

(1)

where y is the measurement vector containing the multispectral and multi-angle polarimetric measurements of DPC. $e_y$ represents the measurement error. x is the state vector containing the parameters to be retrieved, which include aerosol properties and land or ocean properties.

3.1. Radiative Transfer

We use the SRON radiative transfer model LINTRAN [36,37,38,39]. All the radiative transfer calculations are performed for a model atmosphere that includes Rayleigh scattering, scattering and absorption by aerosols, and gas absorption. Rayleigh scattering cross sections are used from Bucholtz, 1995 [40]. The forward model simulates Stokes parameters I, Q, and U at the height of the observation for given optical properties (scattering and absorption optical thickness and scattering phase matrix for each vertical layer of the model atmosphere; 15 layers of the atmosphere are assumed).

The fine particles are modeled as spheres, and the optical properties are simulated with Mie scattering theory. The coarse particles are modeled as a size–shape mixture of randomly oriented spheroids [41,42]. We use the Mie-T matrix-improved geometrical optics database by Dubovik et al., 2006 along with their proposed spheroid aspect ratio distribution for computing optical properties for a mixture of spheroids and spheres [43]. The ratio of the spheres $f_{\mathrm{sphere}}^{\mathrm{c}}$ is in the state vector and simultaneously retrieved.

3.2. State Vector

We use the five-mode generic aerosol description according to Fu and Hasekamp, (2018). The mode properties are summarized in

Table 2. Definition of the effective radius (r_eff) and the effective variance (v_eff) in the retrieval.

	Mode 1	Mode 2	Mode 3	Mode 4	Mode 5
reff (μm)	0.094	0.163	0.282	0.882	1.759
veff	0.130	0.130	0.130	0.284	1.718

. It has been used in a global sense by validations with 22 AERONET stations [23] and also applied in the US during ACEPOL (Aerosol Characterization from Polarimeter and Lidar) campaign over land [35] and over ocean [44]. We consider modes 1–3 together as the fine mode and modes 4–5 together as the coarse mode. The refractive indices for the fine and coarse modes are not retrieved directly; however, they can be described as

$$\mathrm{m}\left(\mathsf{\lambda }\right)={{\displaystyle \sum }}_{k=1}^{n_{\alpha }}{\alpha }_k{m}^k\left(\mathsf{\lambda }\right)$$

(2)

where ${\alpha }_k$ (0 ≤ ${\alpha }_k$ ≤ 1) are defined as the weighting coefficients to combine the prescribed refractive index spectra $m^k\left(\mathsf{\lambda }\right)$ from d’Almeida et al., 1991 for different aerosol components [45], e.g., black carbon (BC), inorganic matter (INORG), DUST, and water (H₂O). ${\alpha }_k$ of each aerosol component are included in the state vector to be determined in the retrieval. n_α is set to be 2 for both fine and coarse modes in this study. The fine mode and the coarse mode are assumed to be composed by INORG + BC and DUST + INORG, respectively.

The aerosol parameters included in the retrieval state vector x are the aerosol column numbers for the five modes (

Table 3. State vector for 5-mode retrieval.

	Parameters in the State Vector	5-Mode Retrieval
Aerosol properties	Aerosol loading	N j, (j = 1, 2, …, 5)
	Fraction of spheres	$f_{\mathrm{sphere}}^{\mathrm{c}}$
	Mode component coefficients	${\alpha }_k^{\mathrm{f}}$$, {\alpha }_k^{\mathrm{c}}$, (k = 1, 2)
	Aerosol layer height	z
Surface properties	Scaling parameter for BPDF model	$x_{\mathrm{bpdf}}^{\mathrm{scale}}$
	Coefficient of Li sparse kernel	$x_{\mathrm{brdf}}^{\mathrm{geo}1}$
	Coefficient of Ross thick kernel	$x_{\mathrm{brdf}}^{\mathrm{geo}2}$
	BRDF scaling parameters for wavelength bands	$x_{\mathrm{brdf}}^{\mathrm{iw}}$, (iw = 1, 2, nwave)
	Number of aerosol parameters	nmode + 6
	Number of surface parameters	nwave + 3
	Length of the state vector	nmode + nwave + 9

), two coefficients (inorganic, black carbon) for the fine-mode refractive index, two coefficients (inorganic, dust) for the coarse-mode refractive index, the fraction of spherical particles (assumed the same for two coarse modes), and the central height of a Gaussian aerosol height distribution (assumed the same for all modes). In Table 3, n_mode denotes the number of modes in our retrieval, which is five. There are five wavelength channels (n_wave) of DPC used in our retrieval, including 443 nm, 490 nm (P), 565 nm, 670 nm (P) and 865 nm (P).

3.3. Surface Reflectance Model

For surface models of the bidirectional reflectance distribution function (BRDF), we use the Ross-Li model [46,47] with the same settings as reported in Litvinov et al., 2011 [48]. For modeling surface bidirectional polarization distribution function (BPDF), a Fresnel model is used as introduced by Maignan et al., 2009 [49]. The surface parameters, to be retrieved in the state vector (see Table 3), are scaling parameters for the BPDF model ($x_{\mathrm{bpdf}}^{\mathrm{scale}}$), the coefficient of the Li sparse kernel ($x_{\mathrm{brdf}}^{\mathrm{geo}1}$), the coefficient of the Ross thick kernel ($x_{\mathrm{brdf}}^{\mathrm{geo}2}$), and the BRDF scaling parameters at each wavelength band $\left(x_{\mathrm{brdf}}^{iw}, iw=1, 2, n_{\mathrm{wave}}\right).$

3.4. Inversion

To retrieve the state vector from the DPC measurements, a damped Gauss–Newton iteration method with Phillips–Tikhonov regularization is employed [50]. The inversion algorithm finds the solution $\widehat{x}$, which solves the minimization–optimization problem,

$$\widehat{x}=\min (\parallel S_y^{-\frac{1}{2}}\left(F\left(x\right)-y\right){\parallel }^2+\gamma \parallel W^{-\frac{1}{2}}\left(x-x_a\right){\parallel }^2)$$

(3)

$x_a$ is the a priori state vector, W is a weighting matrix, $\gamma$ is a regularization parameter, and $S_y$ is the measurement error covariance matrix. Since the forward model $F\left(x\right)$ is nonlinear with respect to x, it is needed to linearize the nonlinear forward model so that the retrieval problem can be solved iteratively. For each iteration step, we approximate the forward model $F\left(x\right)$ with

$$\mathrm{F}\left(x\right)\approx \mathrm{F}\left(x_n\right)+\mathrm{K}\left(x-x_n\right)$$

(4)

Here, K is the Jacobian matrix (with ${\mathrm{K}}_{ij}=\raisebox{1ex}{$\partial F_i$}\!\left/ \!\raisebox{-1ex}{$\partial x_j\left(x_n\right)$}\right.$), which contains the derivatives of the forward model with respect to each variable in the state vector x.

The solution of Equation (3) refers to Hasekamp et al., (2011) [50] and is iterated by

$${\widetilde{x }}_{n+1}=\mathsf{\Lambda } \widetilde{G }\widetilde{y }+\widetilde{A }{\widetilde{x }}_n+(\mathrm{I}-\widetilde{A }){\widetilde{x }}_a$$

(5)

with the contribution matrix $\widetilde{G }={({\widetilde{K }}^T\widetilde{K }+\gamma \mathrm{I})}^{-1}{\widetilde{K }}^T$ and the averaging kernel matrix $\widetilde{A }=\widetilde{G }\widetilde{K }. \mathsf{\Lambda }$ is a filter factor, which limits the step size for each iteration of the state vector. In this way, we use a Gauss–Newton scheme with reduced step size to avoid diverging retrievals. The filter factor Λ shows values between 0 and 1. The regularization parameter $\gamma$ and filter factor are chosen optimally (for each iteration) from different values for $\gamma$ (5 values from 0.1 to 5) and for $\mathsf{\Lambda }$ (10 values from 0.1 to 1) by evaluating the goodness of fit using the forward model.

We filter the retrievals based on the goodness of fit χ² between forward model and measurement:

$${\chi }^2=\frac{1}{n_{\mathrm{meas}}}{\displaystyle \sum _{i=1}^{n_{\mathrm{meas}}}\frac{{(F_i-y_i)}^2}{S_y(i,i)}}$$

(6)

Retrievals with χ² < ${\mathsf{\chi }}_{\max }^2$ are considered as valid retrievals. We apply this filter to exclude the retrievals in which the forward model is unable to fit the measurements, i.e., due to cloud-contaminated pixels [51,52], corrupted measurements [50], and cases in which the first guess state vector deviates too much from the truth. ${\mathsf{\chi }}_{\max }^2$ = 5 is set to be the threshold for DPC to define the successfully converged retrievals, based on the previous study on PARASOL [23]. To ensure a sufficient amount of information, the filters on imaging angles include: (1) The number of viewing angles should be more than five; (2) The maximum view zenith angle should be larger than 40° and the minimum view zenith angle should be less than −40°.

4. Validation against AERONET Measurements

To validate DPC retrievals, the AOT, fine mode AOD and SSA data of AERONET are used. To make DPC retrievals and AERONET data comparable, DPC retrievals located within 10 km from AERONET stations are selected and averaged. For the AOD, we use AERONET measurements performed within 0.5 h of the DPC overpass. For the SSA, maximal temporal variation between AERONET and DPC retrievals does not exceed 2 h. AERONET data are averaged for validation and comparison.

4.1. AOD

According to the average FMF measured by AERONET, the AERONET sites are divided into urban type and dust type. The contribution of anthropogenic fine particles is dominant in urban sites with the average FMF larger than 0.5, while the average FMF is less than 0.5 for dust sites where coarse particles contribute more. Figure 1 shows the retrieved values from DPC versus AERONET measurements for the total AOD at 500 nm. To quantitatively evaluate the performances of different retrieval cases, the RMSE (Root-Mean-Square Error) and absolute mean bias are shown. The bias can be positive or negative, meaning the overestimation or the underestimation. Bréon et al., 2011 defined the Fraction of Good retrieval (Gfrac) as AOD difference against the sun photometer data within ±(0.05 + 15% × AOD_truth) [53]. The Global Climate Observing System (GCOS) Essential Climate variables Data Access Matrix formulated accuracy requirement on AOD as ±(0.03 + 10% × AOD_truth). We used GCOS_frac of ±(0.04 + 10% × AOD_truth) to take AERONET uncertainty into account. Both Gfrac and GCOS_frac are calculated and shown in Figure 1. Overall, the DPC AOD retrievals at 500 nm compare reasonably well with AERONET. Nevertheless, the retrievals show a trend from overestimation at small AOD (possibly related to cloud contamination) to underestimation at large AOD. A similar, but weaker trend was observed by Fu and Hasekamp (2018) [23] for POLDER-3 retrievals using the RemoTAP algorithm under the same setup. It can be seen that DPC retrievals perform better over urban regions than over dust regions with higher correlation and Gfrac shown in Figure 1a. The normalized RMSE values calculated as RMSE/$\overline{\tau }$ are 0.392 and 0.451 for urban and dust matchups, respectively, where $\overline{\tau }$ is the average of the retrieved AOD. However, it should also be noted that there are significantly less dust cases than urban cases. The more challenging (high) surface reflection and the mix of fine and coarse particles probably render the retrievals over dust sites more challenging.

The application to POLDER data of the same algorithm in the same setup resulted in quite similar RMSE values as we obtain for DPC (~0.19 for PARASOL and ~0.166 for DPC, respectively), although it should be noted that the global performance evaluation cannot be compared one-to-one to the regional evaluation here. The Gfrac of DPC AOD within the error envelope of ±(0.05 + 15%) for all matchups is 47.3%, close to that of PARASOL AOD (45.1% at 675 nm) with the GRASP algorithm [54].

Figure 2 shows the probability distribution of the AOD retrieval bias. It can be seen that there is a significant overestimation when AOD < 0.3, but it is not obvious when AOD > 0.3. The AOD retrieval biases in different AOD intervals are calculated and shown in Figure 3. It demonstrates that there is significant overestimation when AOD < 0.2, and significant underestimation when AOD > 1.5. A similar performance was found in AOD retrieval with GRASP algorithm using PARASOL data. Tan et al., 2019 found that the PARASOL/GRASP AOD is overestimated in low aerosol loadings, and underestimated in heavy aerosol loadings, as validated against AERONET measurements over China [54]. On the other hand, optimization of the algorithms may be possible to improve the retrieval capability for moderate aerosol loading.

4.2. Fine Mode AOD

The multi-mode retrieval approach has more freedom in fitting different shapes of size distribution when the number of chosen modes is sufficiently large (Fu and Hasekamp, 2018) [23]. In this sub-section, the AOD of fine mode consisting of modes 1–3 retrieved with DPC is evaluated. Figure 4 shows the comparison with AERONET SDA fine mode AOD at 500 nm for urban, dust-dominated stations and all stations, which are represented by the three sub-plots. It indicates good agreement with AERONET ground-based fine mode AOD, with overall RMSE of 0.1412 and bias of 0.0217 shown in Figure 4c. Similar to the total AOD validation results, the retrieval performs better over urban stations than dust stations with average bias values of 0.001 and 0.132, and RMSE values of 0.129 and 0.194, respectively. In addition, comparing Figure 4a,b, higher correlation is obtained over urban stations (0.931 and 0.786, respectively). For the dust stations, the fine AOD retrievals are significantly overestimated, which is not obvious for the urban stations. A possible reason for this can be attributed to the more challenging (high) surface reflection and the mix of fine and coarse particles that render the retrievals over dust sites more challenging.

Compared with the total AOD validation as shown in Figure 1c, the Gfrac of fine mode AOD retrievals is significantly improved (63.8% and 47.3% for fine and total AOD, respectively), which further indicates that DPC retrievals have better capability for scenes dominated by fine particles. Wu et al., 2015 concluded that the inclusion of shortwave infrared bands (especially 1590 nm) can yield smaller retrieval errors particularly for the coarse mode [26].

Figure 5 shows the probability distribution of fine mode AOD retrieval error. Compared with the situation of total AOD, the peak probability of fine AOD bias is closer to bias = 0, which indicates that the overestimation of fine mode AOD is not obvious. A similar result can also be found from Figure 6, which shows the retrieval errors in different fine mode AOD intervals. The mean biases represented with blue circles show that the overestimation at low AODs which was found for the AOD retrieval is not significant for the fine AOD retrieval.

4.3. Single Scattering Albedo (SSA)

As described in Section 2.2, the AERONET SSA itself bears significant uncertainty of at least 0.03 [34]. This should be taken into account for the comparisons shown in this section. It is known that retrieval of SSA becomes more difficult at low AOD, both for satellite-based and ground-based remote sensing measurements. Thus in Figure 7 the SSA retrievals at 675 nm with AOD (675 nm) larger than 0.3 are shown.

Some differences are found between DPC and AERONET SSA retrievals, but these discrepancies decrease with increasing AOD. There is a good agreement when discrepancies between modeled and DPC measurements fall within χ² < 3, where most of the SSA differences do not exceed 0.05 for AOD (675 nm) larger than 0.3. Regardless of the AODs, most of our SSA retrievals (~71%) show a difference with AERONET SSA retrievals lower than 0.05. DPC tends to retrieve higher SSA values than AERONET. The coefficient of correlation (R) is 0.45. The results are encouraging, considering that sensitivities of space-based and ground-based instruments are different and that the AERONET SSA inversion itself also bears substantial uncertainty. Moreover, it is worth mentioning that most of the SSA retrievals included in our comparison have rather high SSA values: 85.5% and 51.6% of the AERONET SSA inversions are higher than 0.90 and 0.95, respectively. Overall, the comparison between DPC SSA retrievals and AERONET, as performed here, shows a similar performance to POLDER-3 SSA retrievals [23,55,56,57,58], although it should be mentioned that the POLDER-3 evaluations have been performed for multi-year global data sets.

5. Case Studies

In order to further explore the application potential of DPC aerosol products, we perform two typical cases including fire spots and dust over northern China. Different DPC aerosol retrieval products are used for different scenes. The MODIS fire spot products and the AIRS dust products are used to interpret the variance of DPC aerosol retrievals.

5.1. Fire and Dust

Figure 8 shows the regional aerosol retrievals of DPC at 550 nm on 13 March 2020. As the SSA retrievals tend to have large bias at low AOD, the SSA retrievals with AOD less than 0.3 are excluded. Figure 9 shows the MODIS fire and thermal anomalies and AIRS dust score distribution over Inner Mongolia and Shanxi provinces in China on 13 March 2020. Higher dust scores indicate more certainty that dust is present. Dust is probable when the score is above 380. The retrieved AODs over northern China from 10 to 15 March 2020 are shown in Figure 10.

It can be seen from Figure 8a that high values of total AOD appear in western Inner Mongolia (near Baotou shown in Figure 9, ~40.85°N, 109.63°E) and southwest Shanxi (~34.72°N, 110.38°E). Though from the difference between the fine AOD and the coarse AOD over these two regions, the contributions of coarse particles and fine particles are difficult to distinguish. However, the SSA retrievals shown in Figure 8d indicate obviously different aerosol sources (dust particles and fire smoke particles) in these two regions. In southwest Shanxi, the SSA values are significantly higher (~0.94) representative of dust. While the SSA retrievals are lower in western Inner Mongolia with some values close to 0.8, indicating typical absorbing biomass burning aerosols. It should be noted that during the spring season, there is a high occurrence of dust transport originating from Xinjiang and Western Inner Mongolia in China. The dust particles transported from west to east lead to an increase in coarse particles in northern China. The existence of dust over southwest Shanxi can be also found from the peak AIRS dust score of about 500 in Figure 9. Figure 9 also shows fire spots around Baotou identified by MODIS. The difference in aerosol source over these two regions can be further verified by MODIS fire spots and the AIRS dust score distribution.

5.2. Fire Only

Figure 11 shows the regional aerosol retrievals with DPC over northeast China on 6 February 2020. The simultaneous MODIS fire products are shown in Figure 12. The retrieved AOD from 4 to 9 February 2020 are shown in Figure 13. The total AOD and fine mode AOD retrievals indicate moderate haze weather in western Liaoning province and southeastern Inner Mongolia. The discrete high AOD values in northern Hebei province (approximate latitude of 40–42°N) are probably due to poor cloud screening. Comparing the fine and coarse mode AOD, the western part of Liaoning province is obviously dominated by fine particles, while the contribution of coarse and fine particles is almost equivalent in the southeast of Inner Mongolia. Figure 11d shows obviously lower SSA values over the central and southern parts of Liaoning Province. It can be interpreted by MODIS fire spots at the same area shown with orange dots in Figure 11. The strong absorption aerosol produced by biomass combustion is considered to be the main cause for the low SSA value.

5.3. Discussion

From the above two cases, we find that while the total AOD represents aerosol loading, the difference between fine and coarse mode AOD can provide certain information regarding aerosol source. Generally, in populated areas such as urban environments, the contribution of fine particle AOD affected by human activities is greater, but in some seasons, it may be affected by dust transport and become dominated by coarse particles. The regions dominated by coarse particles generally occur in deserts and surrounding areas affected by dust transport. The SSA is particularly sensitive to biomass burning aerosols, and when there is a significant low value, it usually corresponds to the ground fire point. In addition, to be a good identifier for smoke aerosols, accurate SSA retrieval is of great significance to quantify the radiative forcing and the changes in climate.

6. Conclusions

Among passive satellite sensors, the Multi-Angle Polarimeter (MAP) can provide the richest information on aerosol properties. The DPC onboard the Chinese GF-5 satellite, launched in May 2018, bears similar specifications as POLDER-3 on the PARASOL microsatellite. In this study, we apply the SRON RemoTAP retrieval algorithm to DPC measurements to obtain the aerosol optical properties, including the total AOD, the fine mode and coarse mode AOD and the SSA.

The DPC Level 1b data after correction of stray light between December 2019 and April 2020 are used for validation against the AERONET measurements. According to the average FMF measured by AERONET, the selected stations are divided into urban type (FMF ≥ 0.5) and dust type (FMF < 0.5). For the total AOD at 500 nm, DPC retrievals show better performance over urban stations than over dust stations with average biases of 0.055 and 0.106, and RMSEs of 0.151 and 0.228, respectively. There is a significant overestimation when AOD < 0.2 and when AOD becomes large (AOD > 1.5), there is a significant underestimation. A similar retrieval bias can be also found from GRASP AOD retrievals with PARASOL measurements [54]. A possible reason for this may be the prior values, which lead to significant deviations of the AOD inversion in extreme cases. However, the performance is improved for ordinary aerosol loadings with AOD between 0.3 and 1.5 at 500 nm.

The comparison with AERONET SDA fine mode AOD indicates good agreement, with an overall RMSE of 0.1412 and bias of 0.0217. Similar to the total AOD validation results, the retrieval performs better over urban stations than dust stations with the average biases of 0.001 and 0.132, and RMSEs of 0.129 and 0.194, respectively. Compared with validation of the total AOD, relatively lower bias (0.021 and 0.065) and higher Gfrac (63.8% and 47.3%) are found for the fine mode AOD.

For the SSA validation, it should be noted that the AERONET SSA itself bears significant uncertainty of at least 0.03. This should be taken into account for the comparison. For AOD (675 nm) larger than 0.3, most of our SSA retrievals (~71%) show differences with AERONET SSA retrievals lower than 0.05. DPC tends to retrieve higher SSA values than AERONET. The similar performance observed between DPC and POLDER-3 SSA retrievals can be considered as encouraging.

To further explore the application potential of DPC aerosol products, two typical cases including fire spots or dust over northern China are analyzed and interpreted with simultaneous MODIS fire and AIRS dust products. From the case studies, we find that the difference between fine and coarse mode AOD can provide more information on aerosol source compared with using the total AOD alone. SSA retrievals are particularly sensitive to absorbing aerosols, thus they can be useful to track biomass burning aerosols. In addition, accurate SSA retrieval is of great significance to evaluate the effect of aerosols on radiative forcing and the global change.