Multiscale Validation and Trend Evolution of Global Aerosol Reanalysis Datasets: A Comprehensive Comparative Study of CAMS and MERRA-2

Highlights

What are the main findings?

• CAMS AOD shows a higher global correlation, while MERRA-2 AOD is more robust; both perform best in low–mid latitudes, and MERRA-2 AE is overall superior to CAMS AE.
• Spatiotemporal validation reveals strong seasonal and hourly discrepancies between the two reanalysis datasets; 2003–2023 global AOD declines monotonically, whereas MERRA-2 AE shows an increasing trend after EEMD.

What are the implications of the main findings?

• CAMS reanalysis is recommended for short-term, real-time, and urban/biomass burning applications, while MERRA-2 is more suitable for coarse-mode aerosol and long-term climate studies.
• The identified biases and performance gaps provide clear directions for improving aerosol emission inventories, parameterization schemes, and data assimilation in future reanalysis systems.

Abstract

Aerosol optical depth (AOD) and Ångström exponent (AE) are critical parameters for characterizing atmospheric aerosols, playing a pivotal role in atmospheric environmental monitoring and climate change studies. This study addressed the imperative need for a systematic evaluation of mainstream reanalysis products by conducting a comprehensive multi-scale assessment of the CAMS and MERRA-2 datasets (2003–2023), encompassing data quality verification, spatiotemporal pattern analysis, and trend evolution investigation. The following key findings emerge: (1) Both AOD data exhibited the best performance observed in low–mid latitudes. CAMS AOD (AODC) showed a slightly better correlation, while MERRA-2 AOD (AODM) demonstrated superior robustness. Both AE data performed similarly, and MERRA-2 AE (AEM) was superior. Both AE data performed better in low latitudes and near Europe. (2) CAMS and MERRA-2 showed good performance in annual and seasonal variations, with significant fluctuations and biases in the annual cycle. Both models achieved the highest AE performance in summer. MERRA-2 AOD demonstrated better hourly performance during daytime. The hourly stability of AE was slightly worse than AOD, with notably degraded performance during midday hours. (3) The distribution and trends of AOD over land showed spatial consistency. The distribution of AEM was generally lower than AEC’s. After ensemble empirical mode decomposition (EEMD), all datasets showed monotonically decreasing trends except for AEM. This study provides valuable insights into the strengths and limitations for CAMS and MERRA-2 and suggests possible areas for improvement in future data assimilation and parameterization.

1. Introduction

Atmospheric aerosols refer to solid and liquid particles suspended in the air, with aerodynamic diameters ranging from 10⁻³ μm to 10 μm [1]. Aerosols primarily originate from natural and anthropogenic sources that are typically associated with dust, smoke, soot, and sea salt [2]. Aerosol particles in the atmosphere can affect the Earth’s radiation balance, altering the Earth’s albedo by scattering and absorbing solar radiation and affecting cloud formation and properties [3]. It can also impact the climate through direct and indirect radiative effects [4], such as precipitation [5] and temperature [6]. Excessive aerosols can degrade air quality and have adverse effects on human health [7]. Therefore, it plays a critical role in global climate change, air quality, and human health [8]. As an important physical parameter of aerosols, the aerosol optical depth (AOD) represents the integral of the atmospheric vertical extinction coefficient, used to study the optical properties of aerosols and estimate radiative efficiency [9], air quality [10], and global radiative forcing [11]. The Ångström exponent (AE) represents the relative intensity of extinction or absorption at different wavelengths and has been widely used to infer aerosol size and type [12]. AOD and AE are considered key factors in understanding the state of aerosols in the environment [13].

Compared to ground-based observations, satellite remote sensing provides global and real-time aerosol observations. Due to the complexity of aerosol physicochemical properties, differences in satellite sensors, assumptions in retrieval algorithms, limitations in observational conditions, and regional environmental heterogeneity, accurate monitoring of aerosol spatiotemporal distribution and modeling of the Earth’s climate system always face the challenge of multi-source uncertainties [1,14,15,16]. To address this issue, many current studies aim to optimize the shortcomings of individual products by combining various data sources. As a technique that integrates satellite observations with climate models, data assimilation has the advantage of dynamically integrating real-time observational data with model physical processes. It can fill gaps in both space and time, systematically reducing the uncertainty of aerosol spatio-temporal distribution [17].

In this context, the Copernicus Atmosphere Monitoring Service (CAMS) from European Centre for Medium-Range Weather Forecasts (ECMWF) and the Modern-Era Retrospective Analysis for Research and Applications-2 (MERRA-2) from National Aeronautics and Space Administration (NASA) aerosol products provide globally continuous aerosol data with spatial and temporal coverage, particularly in regions lacking observations [18]. CAMS and MERRA-2 are increasingly being widely used in climate system research, with both possessing long-term development potential. As a result, a number of evaluative studies have emerged for CAMS and MERRA-2, respectively. For example, Júnior et al. [19] evaluated CAMS AOD (AODC) in Brazil, and the results showed good consistency with ground-based AOD observations. There has been relatively more validation of MERRA-2, with research focused on specific countries or regions (Saudi Arabia [20], China [21], Turkey [22], Australia [12], the Indochina Peninsula [23], West Africa [24], etc.). Su et al. [25] also evaluated and compared the accuracy of MERRA-2 and various AOD products globally, but independent global validation studies for MERRA-2 are still limited.

The differences in the selection of assimilation strategies, model frameworks and observation sources lead to significant differences in their product characteristics, directly influencing the applicability of the data in different application scenarios. At the same time, both involve the assimilation of Moderate Resolution Imaging Spectroradiometer (MODIS) data. As a result, a small number of regional studies have emerged to compare and evaluate them. But the systematic evaluations of CAMS and MERRA-2 remain fragmented. Due to their relatively coarse spatial resolution, which limits their application at finer spatial scales, related discussions mainly focus on large-scale study areas, including southern Morocco [26], China [27,28], Asia [18], and global scales [29,30]. The current limitations include excessive focus on AOD, neglecting AE validation, especially over the oceans. Additionally, time analysis is limited to coarse resolution, overlooking day and annual variations. There is also insufficient analysis of the long-term trend consistency between products.

Therefore, this study jointly evaluates AOD and AE to clarify the link between aerosol microphysics and the climate. The study hopes to fill these gaps by providing a comprehensive, multi-scale assessment of the CAMS and MERRA-2 AOD/AE datasets (2003–2023) through validation and trend detection. This work provides strong support for climate modeling and policy-making while identifying priority areas for future aerosol reanalysis data optimization.

2. Materials and Methods

2.1. Ground-Based AERONET Measurements

The Aerosol Robotic Network (AERONET) is a ground-based aerosol automatic observation network established by NASA, PHOtométrie pour le Traitement Opérationnel de Normalisation Satellitaire (PHOTONS), and other international organizations to monitor relevant aerosol parameters (https://aeronet.gsfc.nasa.gov/, accessed on 23 May 2024) [31]. Given the high precision of its observational data, many studies use it as a fundamental reference for validating and assessing the accuracy of AOD products [32]. Version 3 of AERONET provides three levels of AOD data, including unfiltered Level 1.0, cloud-filtered Level 1.5, and quality-assured Level 2.0 products [33]. To evaluate the reliability and long-term trends of AOD products, this study only considers sites with valid Level 2.0 aerosol data records for 5 years or more. A total of 504 sites were selected for this paper, including 488 land sites and 16 oceanic sites. For spatial validation, the world is divided into 10 sub-regions based on conventional classification criteria to facilitate the validation analysis. The spatial distribution of the sites and the regional division are shown in Figure 1.

AERONET collects data at multiple wavelengths but does not directly provide observational data at 550 nm. Therefore, the study used AE (α) to obtain the AOD (τ) at 550 nm by interpolation from 440 nm to 870 nm, as shown in Formulas (1) and (2).

$$\alpha =-{\displaystyle \frac{\ln ({\tau }_{440}/{\tau }_{870})}{\ln (440/870)}}$$

(1)

$${\tau }_{550}={{\tau }_{870}(550/870)}^{-\alpha }$$

(2)

2.2. Aerosol Reanalysis

CAMS is a global reanalysis dataset built on the ECMWF Integrated Forecasting System (IFS) (https://ads.atmosphere.copernicus.eu/, accessed on 15 February 2024) [34]. CAMS uses the four-dimensional variational (4D-Var) assimilation algorithm to integrate data from products such as the Envisat Advanced Along-Track Scanning Radiometer (AATSR), MODIS Terra, and MODIS Aqua [35]. It was started in 2003 and has a temporal resolution of 3 h and a spatial resolution of 0.75° × 0.75° [27]. It provides the total aerosol information at wavelengths of 469, 550, 670, 865, and 1240 nm. This study used the total AOD data at 550 nm from CAMS, but CAMS does not directly provide AE. AE reflects the wavelength dependence of aerosol particle size on light absorption. The wavelength difference between 469 nm and 865 nm is larger, making it more sensitive to fine-mode aerosols. Furthermore, numerical simulations have shown that the difference between AE obtained from the 440–870 nm and 470–870 nm wavelength ranges is negligible [36]. Therefore, AE is derived from the AOD at 469 nm and 865 nm.

MERRA-2 is a reanalysis product obtained by NASA using the Goddard Earth Observing System, version 5 (GEOS-5) (https://gmao.gsfc.nasa.gov/, accessed on 17 May 2024). Sources for MERRA-2 assimilation include information from the Advanced Very High Resolution Radiometer (AVHRR), Multiangle Imaging SpectroRadiometer (MISR), AERONET, and MODIS Terra and MODIS Aqua [37]. This product has a temporal resolution of 1 h with time stamps at the midpoint of each hour (e.g., 00:30, 01:30, … UTC), and a spatial resolution of 0.5° × 0.625°. This study used the 550 nm AOD and AE of 470–870 nm directly provided by MERRA-2.

2.3. Evaluation Methods

Both reanalysis products have relatively coarse spatial resolutions and are quite similar. To preserve the reliability of their original values for practical applications, we did not unify their spatial resolutions. AERONET performs repeated measurements at fixed points, while reanalysis products provide gridded data. For spatial matching, the reanalysis grid cell containing each AERONET site’s coordinates was identified, and the corresponding grid value was assigned to the site. For temporal matching, considering the different temporal resolutions of the reanalysis products, the AERONET data within ±30 min of each reanalysis timestamp were averaged and matched with the corresponding reanalysis values [38].

The study mainly evaluates the correlation between the reanalysis product and AERONET data through various statistical analysis methods. The accuracy of the reanalysis product is quantitatively validated using five indicators: Pearson correlation coefficient (R), Relative Mean Bias (RMB), Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and Bias [39]. The expected error (EE) definition quantifies the acceptable uncertainty envelope, and the Within EE percentage represents the fraction of reanalysis values falling within this envelope relative to AERONET observations [40]. Linear regression models ($y=kx+b$) are also used to assess the quality of the product, as shown in Formulas (3)–(7).

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

(3)

$$\mathrm{RMB}={\displaystyle \frac{\overline{y}}{\overline{x}}}$$

(4)

$$\mathrm{MAE}={\displaystyle \frac{1}{n}}{\displaystyle \sum _{i=1}^n\left|y_i-x_i\right|}$$

(5)

$$\mathrm{RMSE}=\sqrt{{\displaystyle \frac{1}{n}}{\displaystyle \sum _{i=1}^n{\left(y_i-x_i\right)}^2}}$$

(6)

$$\mathrm{EE}=\pm \left(0.05+0.15\times x_i\right)$$

(7)

Here, x and y represent the true values from AERONET and the reanalysis values, respectively; n denotes the number of validation samples; i is the index of the selected validation points; and x and y represent the mean values of the observed and reanalysis values, respectively.

2.4. TS Slope and MK Test

The Theil–Sen median (TS) method is a non-parametric statistical technique used to analyze trends. Due to its high computational efficiency, it is often applied in long-term series analysis. In this study, a TS slope is used to describe the trends of AOD or AE over the years.

$$S=\mathrm{median}\left({\displaystyle \frac{X_j-X_i}{j-i}}\right),\hspace{1em}2003\le i<j\le 2023$$

(8)

where X_i and X_j represent the values in the i-th and j-th years, respectively; when S > 0, the variable shows an increasing trend; otherwise, it shows a decreasing trend.

The Mann–Kendall (MK) test does not assume that the measured values follow a normal distribution and is robust to missing values and outliers. The applicability of the MK test for trend significance testing has been widely confirmed. Therefore, this study uses the MK test to assess the significance level of AOD trends.

$$S={\displaystyle \sum _{i=1}^{n-1}{\displaystyle \sum _{j=i+1}^n\mathrm{sgn}\left(X_j-X_i\right)}}$$

(9)

$$\mathrm{sgn}(X_j-X_i)=\left\{\begin{array}{cc}1\hfill & X_j-X_i>0\hfill \\ 0\hfill & X_j-X_i=0\hfill \\ -1\hfill & X_j-X_i<0\hfill \end{array}\right.$$

(10)

In Formulas (9) and (10), X_i and X_j represent the values in the i-th and j-th years, respectively, and n represents the length of the time series. Then, the variance of S is calculated using the following formula.

$$\mathrm{Var}(S)={\displaystyle \frac{n(n-1)(2n+5)-{\displaystyle \sum _{i=1}^m{t}_i}(t_i-1)(2t_i+5)}{18}}$$

(11)

where m is the number of tied ranks, and each group has t_i tied observations. The test statistic Z is used for the trend test, and its calculation method is as follows.

$$Z=\left\{\begin{array}{cc}\left(S-1\right)/\sqrt{Var\left(S\right)}\hfill & S>0\hfill \\ 0\hfill & S=0\hfill \\ \left(S+1\right)/\sqrt{Var\left(S\right)}\hfill & S<0\hfill \end{array}\right.$$

(12)

For a significance level α, when |Z| > Z_1−α/2, it indicates a significant increasing or decreasing trend. At the 99% and 95% confidence levels, |Z| is greater than or equal to 2.58 and 1.96, respectively.

On the basis of the above MK trend significance test, the MK mutation analysis was further adopted to identify the temporal mutation characteristics of AOD or AE during the study period. To implement the mutation detection, the upward fluctuation (UF) sequence was first calculated by processing the original time series in forward order. Specifically, the order sequence S_k (cumulative count of X_i > X_j for 1 ≤ j ≤ i) was constructed, and the UF statistic was standardized as follows:

$${\mathrm{UF}}_k=\frac{S_k-E(S_k)}{\sqrt{\mathrm{Var}(S_k)}}\hspace{1em}(k=1,2,\dots ,n)$$

(13)

where $E(S_k)={\displaystyle \frac{k(k-1)}{4}}$ and $\mathrm{V}\mathrm{a}\mathrm{r}(S_k)={\displaystyle \frac{k(k-1)(2k+5)}{72}}$ are the mean and variance of S_k, respectively. Subsequently, the backward fluctuation (UB) sequence was obtained by reversing the original time series, repeating the above calculation, and negating the results to align with the forward trend. Similarly, the calculation formula for the UB statistic in the reversed sequence is as follows:

$${\mathrm{UB}}_k=-{\displaystyle \frac{Sk{2}_k-E(Sk{2}_k)}{\sqrt{\mathrm{Var}(Sk{2}_k)}}}\hspace{1em}(k=1,2,\dots ,n)$$

(14)

where Sk2_k is the order sequence calculated from the reversed time series, and E(Sk2_k), Var(Sk2_k) have the same calculation formulas as E(S_k) and Var(S_k).

Using the same significance level of α = 0.05 (critical value ±1.96 as mentioned above), the mutation time of the target variables was determined by the intersection of the UF and UB sequences. Specifically, if the intersection point lies within the 95% confidence interval [−1.96, 1.96], the corresponding time was identified as a significant mutation point [41].

2.5. Ensemble Empirical Mode Decomposition (EEMD)

The EEMD method is an extension of Empirical Mode Decomposition (EMD), which effectively addresses the mode mixing problem in EMD while retaining the ability of binary filters. It uses a “sifting” process to decompose the time series $x\left(t\right)$ into multiple intrinsic mode functions (${IMF}_i=1,...,k$) with progressively decreasing frequencies, as well as a long-term trend $R_n\left(t\right)$. It has been widely used in climate change research.

Specifically, the sifting process identifies the local maxima and minima of the current signal and connects them with cubic spline interpolation to form the upper and lower envelopes. The mean of these two envelopes is then subtracted from the signal to produce a candidate component that isolates the highest-frequency oscillation present at that stage. This subtraction step is repeated iteratively on the candidate until it satisfies the IMF conditions (i.e., the number of extrema and zero-crossings differs by at most one and the local mean is sufficiently close to zero). The resulting IMF is subtracted from the original series to obtain the first residue, and the entire sifting procedure is applied to this residue to extract the next IMF with a lower frequency. The process continues until the final residue $R_n\left(t\right)$ becomes monotonic or contains at most one extremum, which is taken as the long-term trend. Each IMF thus captures a distinct timescale of variability in the original signal, while $R_n\left(t\right)$ represents the slowly varying nonlinear trend. The original signal can be perfectly reconstructed as follows:

$$x\left(t\right)={\displaystyle {\sum }_{i=1}^k{IMF}_i(t)-R_n(t)}$$

(15)

To cope with the mode-mixing issue that is inherent in standard EMD, EEMD introduces a noise-assisted ensemble approach: white noise with a finite amplitude scaled to the standard deviation of the original data is added to the time series, EMD is performed on each noisy realization, and the resulting IMFs and residues are averaged over a large number of independent trials. This averaging effectively cancels the added noise while stabilizing the decomposition and eliminating mode mixing.

At a specific time t, the EEMD trend is the increment of R_n since 2003.

$$Trend(t)=R_n(t)-R_n(2003)$$

(16)

These trends’ rate of change can be calculated as follows:

$${Rate}_{trend}(t)=Trend(t)-Trend(t-1)$$

(17)

If the EEMD trend is monotonic, it is considered to have no trend change. For a trend with one extreme value, the location of the extreme value is referred to as a turning point. This study implements the EEMD simulation based on Python (version 3.9.2), with an ensemble size of 250 and a noise amplitude of 0.2 times the standard deviation of the original data.

In addition, the instantaneous rate of the EEMD trend refers to the rate of change in the nonlinear trend $R_n\left(t\right)$ at each specific time point, calculated as the first-order derivative ${dR}_n\left(t\right)/dt$. Unlike the linear trend rate, which remains constant over time, the instantaneous rate captures the time-varying characteristics of the nonlinear evolution. This approach allows for the identification of acceleration or deceleration phases in the trend, as well as the timing of turning points [42]. For practical implementation using discrete annual data, the instantaneous rate at year t is approximated as follows:

$$Rate(t)={\displaystyle \frac{\left[Rn(t+1)-Rn(t-1)\right]}{2}}$$

(18)

3. Results and Discussion

3.1. Comparative Validation of Global and Regional Performance

3.1.1. Land and Ocean Validation

From the perspective of land (Figure 2a,c and Figure 3a,c), the sample size of MERRA-2 was approximately three times that of CAMS for both AOD and AE. For AOD, AODC showed a slightly higher correlation coefficient (R = 0.841) than MERRA-2 AOD (AODM) (R = 0.835), with better performance in R², RMSE, and RMB. However, AODM demonstrated superior performance in MAE (0.061 vs. 0.070) and Within EE (79.63% vs. 71.38%). For AE, both models achieved identical MAE (0.295). MERRA-2 AE (AEM) demonstrated better performance in R (0.658 vs. 0.630), R², and RMSE, while CAMS AE (AEC) showed advantages in RMB, Bias and Within EE. Notably, CAMS exhibited a tendency toward overestimation (AODC: Bias = 0.004 and Above EE = 17.75%, AEC: Bias = 0.067, Above EE = 33.93%), while MERRA-2 products showed underestimation (AODM: Bias = −0.007 and Below EE = 9.77%, AEM: Bias = −0.070 and Below EE = 32.30%). Both models performed excellently on land, reflecting the increasingly mature simulation ability of reanalysis products for AOD. The contrasting Bias patterns can be attributed to differences in emission source estimates and model assumptions between the two reanalysis systems [43].

From the perspective of ocean (Figure 2b,d and Figure 3b,d), both models achieved comparable performance for AOD with correlation coefficients of 0.907 (AODC) and 0.908 (AODM). The R² values were similarly close (0.822 vs. 0.825), as were RMSE (both 0.078) and MAE (0.047 vs. 0.042). AODM showed a higher proportion of data Within EE (87.69% vs. 84.04%). For AE, AEM demonstrated notably better performance than AEC with higher R (0.746 vs. 0.723). AEC showed substantial overestimation with RMB of 1.448, while AEM’s RMB was closer to one. Both models exhibited systematic Bias patterns that were consistent with land validation: CAMS products tended toward overestimation (AODC: Bias = 0.009 and Above EE = 11.58%, AEC: Bias = 0.270 and Above EE = 64.81%), while MERRA-2 products showed underestimation (AODM: Below EE = 8.12%, AEM: Below EE = 21.58%). Over oceans, both products achieved higher correlations than over land, owing to simpler aerosol types and lower surface reflectivity [44]. CAMS’s IFS-AER model overestimated marine AOD and AE due to its simulation of sea salt emissions [43], while MERRA-2 improved consistency through AVHRR assimilation and refined aerosol typing [45]. The assimilation of AERONET data over land significantly reduced model differences, while more reliance on satellite data and model parameters amplified the differences over the ocean. Due to the complexity of land aerosols and the diversity and wide coverage of observations, land consistency was relatively poorer, but the percentage of Within EE was better than over the ocean [45].

3.1.2. Regional Performance Comparison

Regional validation results (Figures S1–S4) revealed spatial heterogeneity in model performance over both land and ocean. Over land, AODM outperformed AODC across most metrics in AF, EU, AU, and NA. In SA, AODC showed better performance in R and R², likely due to its effective capture of biomass burning aerosols [34,46]. For AE, AEC outperformed AEM across most metrics in EU, while AEM showed advantages in AF and SA. Mixed results were observed in AS, AU, and NA, with each model excelling in different aspects. AODM demonstrated better robustness in different aerosol environments, particularly in high latitudes of the Northern Hemisphere and Oceania. This advantage stems from MERRA-2’s assimilation of AERONET AOD observations, which provide ground-truth constraints on aerosol loading, and the assimilation of MISR retrievals over bright surfaces such as deserts and snow/ice where MODIS retrievals are less reliable, thereby improving surface reflectance correction in high-latitude regions. Additionally, the GOCART model employed by MERRA-2 simulates multiple aerosol species including dust, sea salt, organic carbon, black carbon, and sulfate with relatively mature parameterizations for these aerosol types [47,48]. However, MERRA-2’s limited ability to resolve extreme events such as severe dust storms and intense biomass burning episodes led to underestimation of high AOD values, as the data assimilation system tends to smooth out extreme values and the model’s emission inventories may not adequately capture episodic emission events [12]. In biomass burning areas, MERRA-2 showed deviations in handling aerosol categories and the ultraviolet aerosol index because the GOCART model has limited representation of organic aerosol evolution from biomass burning and may misclassify smoke aerosols as other types such as dust or sulfate, leading to poorer AE results [37]. AEM showed better consistency in the western United States and southern Africa due to its advantages in handling dust aerosols [29].

Over ocean, AODM outperformed AODC in MA, NATL, and MP with Within EE exceeding 87%, while AODC performed better in IO. For AE, AEM outperformed AEC in MA and MP across all metrics, while in IO and NATL, the results were mixed. Overall, MERRA-2 products demonstrated more robust performance across regions with smaller systematic errors.

3.1.3. Validation at Station Level

Figure 4 displayed the spatial distribution characteristics of the correlation coefficients between AOD and AE for CAMS and MERRA-2 with AERONET at each station. Both types of reanalysis data performed excellently in AOD retrieval. The best-performing regions were mainly located in the low–mid latitudes, concentrated in eastern North America, the interior of South America, Southeast Asia, southern and western Africa, and the border between Africa and Europe. The regions with relatively poor overall performance for both datasets were few, and mainly concentrated on the western coast and southern tip of South America, which can be attributed to aerosol misclassification, the inversion layer, and humidity effects [49,50,51]. In comparison, the advantage of AODM was more apparent, mainly in the high-latitude regions of the Northern Hemisphere and Oceania. As shown in Figure 4b, there were many regions where the correlations of AEC and AEM with AERONET performed consistently well. The better-performing regions were mainly concentrated in Europe and low–mid latitudes. However, there were significant spatial differences compared to the AOD retrieval performance. AEM showed a clear advantage in western North America and southern Africa, while AEC’s strengths were mainly in the low latitude interior of South America.

3.2. Comparative Validation at Specific Time Cycles

3.2.1. Annual Variation

Figure 5 and Table S1 displayed the annual validation results (R and Bias) of AOD and AE for CAMS and MERRA-2 from 2003 to 2023. Overall, both models demonstrated moderate to good correlations for AOD, with the annual R values ranging from 0.811 to 0.865 for CAMS and 0.789 to 0.870 for MERRA-2, but both exhibited significant interannual fluctuations and systematic Biases. The mean annual R value for AODC (0.840) was slightly higher than AODM (0.835). However, MERRA-2 exhibited greater interannual variability in AOD performance, with R standard deviation of 0.025 compared to CAMS’s 0.014. Notably, the relative performance shifted over time: MERRA-2 showed higher R values during 2003–2011 (mean R = 0.857 vs. 0.842), while CAMS outperformed during 2012–2023 (mean R = 0.838 vs. 0.819), which was linked to VIIRS data assimilation in CAMS [52]. In terms of Bias, AODC showed a slight tendency toward overestimation (mean Bias = 0.002) with a transition from negative Bias in early years to positive bias in later years, whereas MERRA-2 consistently exhibited underestimation (mean Bias = −0.009).

For AE, both models showed lower correlations compared to AOD, with CAMS ranging from 0.607 to 0.721 and MERRA-2 from 0.614 to 0.712. MERRA-2 achieved a higher mean annual R value (0.674 vs. 0.648) and demonstrated better consistency over time, with R standard deviation of 0.021 compared to CAMS’s 0.033. A notable decline in AEC correlation was observed from the early to late period. Regional analysis (Table S2) revealed that the most pronounced decline occurred over AF, where R decreased steadily from 0.879 in 2003 to 0.521 in 2023. Moderate declines were also found over SA and AS. This regional degradation in the AE correlation, particularly over Africa and biomass burning-influenced areas, can largely be attributed to the growing complexity of mixed aerosols (e.g., dust–smoke interactions), which challenges the IFS-AER scheme’s ability to represent particle size distributions [43]. Compounding this are limitations in emission inventories, such as GFAS fire emissions, that struggle to capture the changing intensity and optical properties of biomass burning aerosols under climate and land-use shifts. Additionally, sensitivities in the assimilation of multi-wavelength data for AE derivation further contribute to the observed decline. These results are consistent with earlier findings reporting elevated uncertainties in CAMS aerosol typing over biomass burning regions [34]. AEM maintained relatively stable performance and showed advantages after 2010 [53]. AEC consistently showed positive Bias (mean = 0.068) with a small fluctuation (std = 0.026), indicating systematic overestimation, whereas AEM showed negative Bias (mean = −0.072) with larger fluctuation (std = 0.060), which was attributed to its systematic underestimation of coarse-mode aerosols [37].

3.2.2. Seasonal Variation

The quarterly validation results of AOD (Figure 6) showed that both CAMS and MERRA-2 achieved good correlations across all seasons, with R values ranging from 0.817 to 0.861. AODC achieved the highest R in spring (0.860) but showed the lowest Within EE in summer (68.78%), with higher Bias (0.019) indicating systematic overestimation during summer months. AODM demonstrated more consistent performance across seasons, with Within EE exceeding 79% in all seasons and the highest R being in winter (0.861). MERRA-2 excelled in winter despite slight underestimation (Bias = −0.017), which may be attributed to the lack of nitrate aerosols in the model and underestimation of organic carbon emissions during winter [37]. Regardless of the season, MERRA-2 consistently had higher Within EE than CAMS.

For AE (Figure S5), both models showed lower correlations compared to AOD, with R values ranging from 0.602 to 0.736 for CAMS and 0.605 to 0.703 for MERRA-2. Both datasets achieved the highest R values in summer (CAMS: 0.736, MERRA-2: 0.703) and the lowest in winter (CAMS: 0.602, MERRA-2: 0.605). The Within EE values were highest in summer and autumn and lowest in winter for both models. The poorer performance in winter was attributed to cloud contamination and other factors that caused data loss or invalid measurements, resulting in uncertainty [47], while the larger MAE and RMSE in winter indicated that errors were dominated by random errors. In terms of Bias, CAMS AE showed positive Bias across all seasons except summer, with the most pronounced overestimation being in spring (Bias = 0.155), while MERRA-2 AE consistently showed negative Bias, most notably in spring (Bias = −0.111). The more pronounced Bias in spring compared to winter indicated stronger systematic deviation due to aerosol mixing complexity and meteorological dynamics during the transition season [54].

3.2.3. Hourly Variation

Figure 7 showed the hourly validation results of AOD and AE for CAMS and MERRA-2 during daytime hours. For AOD, both models showed relatively stable R values throughout the day, with AODC ranging from 0.821 to 0.857 and AODM from 0.817 to 0.853. AODC achieved higher R values during morning hours (8:00–10:00 and 12:00–13:00), while AODM showed advantages during afternoon hours (14:00–17:00). MERRA-2 consistently maintained higher Within EE values (79–81%) compared to CAMS (68–74%) throughout the entire period. The contrasting Bias patterns persisted across all hours: AODC showed slight overestimation in the late morning and afternoon, while AODM consistently exhibited underestimation, consistent with the systematic Bias pattern discussed in seasonal analysis.

For AE, both models showed larger fluctuations compared to AOD. AEC R values ranged from 0.547 to 0.716 with greater variability, while AEM showed more stable performance ranging from 0.625 to 0.744. AEM demonstrated consistently higher R values during most hours. However, AEC achieved higher Within EE values than AEM for the majority of hours. The AE performance was notably worse during midday hours (10:00–13:00), with R values dropping below 0.66 for both models and Within EE reaching the lowest points. Both models showed notable Bias patterns in AE. AEC exhibited overestimation throughout the day, while AEM displayed an increasingly pronounced underestimation toward late afternoon. These differences can be attributed to the distinct assimilation strategies and AE derivation methods between the two systems. CAMS employs a 4D-Var assimilation scheme incorporating near-real-time satellite observations (MODIS and VIIRS), which enhances the responsiveness to instantaneous changes in aerosol properties but also renders the estimates more susceptible to noise and model errors, leading to larger fluctuations in correlation performance [34]. In contrast, MERRA-2 uses a three-dimensional variational (3D-Var) assimilation system with more gradual updates and persistent model physics, which smooths out high-frequency variability and produces more stable AE estimates [45]. Furthermore, CAMS does not provide AE directly but derives it from AOD at 469 nm and 865 nm; uncertainties in multi-wavelength AOD assimilation may be amplified when converting to AE, especially on hourly timescales. MERRA-2 outputs AE directly through aerosol optical parameters, benefiting from the internal consistency between aerosol species and thus contributing to more stable hourly performance [48]. This is consistent with the common tendency in reanalyses to provide a moderate estimation of aerosol size distribution: they tend to underestimate fine-mode aerosols in high-AE situations and overestimate coarse-mode aerosols in low-AE situations [28]. Such systematic biases reflect limitations in the models’ representation of aerosol size distributions, possibly arising from uncertainties in emission inventories, microphysical parameterizations, or the assimilation of multi-wavelength AOD data. These deficiencies are particularly evident during midday hours (10:00–13:00), when the AE performance degrades for both models, which is likely due to increased errors in AE retrieval at lower AOD levels [55].

3.3. Spatial Patterns and Long-Term Trend Analysis

3.3.1. Spatial Distribution of Multi-Year Mean

This study conducted a long-term spatial distribution analysis focusing on continental areas that were significantly influenced by both natural and anthropogenic factors. Figure 8 revealed high overall similarity between CAMS and MERRA-2, particularly in the general land patterns of AOD. In most regions, the AOD values remained below 0.3, with high-latitude areas showing values below 0.1. High-AOD hotspots were concentrated in low- to mid-latitudes (West Africa, West Asia, South Asia, East China, and the Taklamakan Desert), primarily due to intense emission sources and the cumulative effects of regional circulation patterns that promote aerosol buildup [56]. The differences between the two reanalyses were primarily in magnitude and specific hotspot locations. In low-AOD regions such as northern Asia, Greenland, and Oceania, AODC tended to be lower than AODM, which was consistent with CAMS’s potential underestimation in data-scarce or low-aerosol conditions due to inversion uncertainties and emission source biases [57,58]. In contrast, AODM was lower than AODC in northern South America, while in high-AOD regions like South Asia and eastern China, AODC was notably higher than AODM—aligning with the overall overestimation tendency of CAMS products observed in validation results. In West Africa (AOD > 0.5), a symmetrical distribution around the equator was evident. However, the two models displayed contrasting spatial patterns. AODC showed higher values south of the equator, reflecting its stronger sensitivity to biomass burning aerosols. In contrast, AODM emphasized higher values north of the equator, aligning more closely with dust dominance in that region. For AE, more pronounced differences emerged. AEM was generally lower than AEC across land areas, which was consistent with MERRA-2’s systematic underestimation of AE. High-AE hotspots appeared in central South America, southern Africa, and southeastern China, though AEM exhibited relatively higher values in southern Africa. Low-AE regions, such as the Sahara Desert and Central Asia, displayed broader spatial extent and lower values in AEM, reflecting MERRA-2’s stronger emphasis on coarse-mode aerosols (e.g., dust and sea salt) compared to CAMS.

3.3.2. Trend Patterns and Significance

The AOD trend patterns (Figure 9a,b) showed general consistency between the two reanalyses: an upward trend at high latitudes and a downward trend at low- to mid-latitudes. The fastest increases (>0.002) occurred in Russia (95–140°E, 55–70°N) and South Asia, which may be related to enhanced transport or emissions from combustion, weather systems, monsoons, urbanization, and industrial activities [59,60,61]. In contrast, the sharp decline in eastern China (<−0.005) is attributed to national policies, energy structure adjustments, and effective pollution control measures [62]. Regional discrepancies were notable in South America (where AODC decreased while AODM increased) and West Asia (where AODC increased while AODM decreased). These opposite trends in South America and West Asia stem from the differing expertise of the two reanalysis systems in handling aerosol types. CAMS (using the IFS-AER model) better captures biomass burning aerosols through its GFAS fire assimilation system and more detailed treatment of organic matter, leading to stronger sensitivity to emission reductions or variability in biomass burning-dominated areas. In contrast, MERRA-2 (based on the GOCART model) has relatively mature parameterizations for dust and sea salt but tends to smooth extreme biomass burning events and has limitations in representing brown carbon or organic aerosol evolution from fires, which can result in underestimation of high AOD/AE episodes and different trend responses [37,48]. In West Africa, the high AOD on both sides of the equator further highlights this: CAMS reflects biomass burning dominance more prominently south of the equator, while MERRA-2 emphasizes coarse-mode dust to the north.

MK significance tests (Figure 9c,d and Table S3) indicated that regions with significant AOD increases accounted for 13.408% (AODC) and 6.975% (AODM), with AODC showing a larger area of significant increase in West Asia. Significant decreases were similarly distributed but broader for AODC (e.g., China, eastern South America, Oceania, northern Europe). AE trends (Figure 9 and Table S3) displayed larger contrasts: AEC was dominated by a highly significant decrease (37.424%), while AEM generally showed no significant positive trend. Differences were evident in central North America, eastern South America, northern Africa, and eastern Oceania, where AEC decreased but AEM increased. This aligns with the models’ Bias patterns—AEC’s overestimation of fine-mode aerosols (high AE) leads to stronger decreasing signals when emissions are controlled, whereas AEM’s tendency to underestimate fine-mode while overestimating coarse-mode produces flatter or opposing trends in mixed-aerosol regions. MK mutation analysis (Figure 9e,f) showed potential intersection points within the 95% confidence interval for some regions. However, to ensure robustness, we further applied the sliding t-test (Figure S7) and Pettitt test (Figure S8). A statistically significant breakpoint was identified only when all three methods consistently indicated a mutation. Under this criterion, no robust breakpoints were detected for either AODC or AODM (Pettitt p > 0.05). Similar conclusions were held for AE, despite apparent intersections in the MK analysis alone.

3.3.3. EEMD-Based Trend Decomposition

EEMD analysis (Figure 10 and Figures S9–S12, and Tables S4–S7) further decomposed the trends. All products except AEM showed a monotonic decreasing trend overall. AODC and AODM had similar average levels, with linear decrease rates of approximately 0.0003 and 0.0002 per year, respectively. The instantaneous rates of EEMD changed in 2014, possibly linked to a slight increase in high-latitude interior regions during the later stage. Notably, MERRA-2 ceased assimilating AERONET AOD observations after October 2014 due to data latency issues [48]. This limitation in evaluating MERRA-2 performance against AERONET post-2014 has been noted by Gueymard et al. [29]. This change may have contributed to the observed slowdown in the decreasing trend of AODM after 2014, as the absence of direct ground-based constraints could reduce the model’s ability to capture certain regional variations compared to the earlier period when AERONET data provided strong assimilation constraints. Combining with Tables S4–S7, negative trends were predominantly distributed in low–mid latitudes, while high latitudes mostly showed positive trends for all time periods. The linear trends of the original data and the EEMD trends at different latitudes revealed that AODC before the turning point and AODM after the turning point were more aligned with the full-time sequence. The largest rate changes before and after the turning points for both occurred in 0–30°N. Specifically, before and after the turning point, AODC’s positive trend shifted from 0.00279 to 0.00217: a decrease of 0.00062. The negative trend shifted from −0.00131 to −0.00195: an increase of 0.00064 with 48.855%. AODM’s positive trend decreased from 0.00136 to 0.00094: a decrease of 0.00042.

AEC and AEM had significantly different average levels, with rates of decrease and increase of 0.002 and 0.0009 per year, respectively (Figures S11 and S12). AEC and AEM exhibited significant differences, which were consistent with the results before EEMD. The EEMD instantaneous decrease rate of AEC slowed down in 2016, while the EEMD instantaneous increase rate of AEM slowed in 2011. The turning points for their instantaneous rates were different. During the second stage, AEC’s coarse/fine particle estimates may have declined in Mongolia, while AEM may have increased coarse/fine particle estimates in areas such as the west coast of Africa and the interior of South America. AEC mainly showed a decreasing trend, except for the 60–90°S region for all time periods (Tables S6 and S7). AEM exhibited positive trends in low–mid latitudes and negative trends in high latitudes. After the turning point, both AEC and AEM were more aligned with the full-time sequence. The largest rate value change for AEC before and after the turning point occurred in 30–60°S, where the positive trend decreased from 0.00211 to 0.00070, a decrease of 0.00141 with 66.825%. The largest percentage change occurred in 60–90°N, where the positive trend decreased from 0.00054 to 0.00007, a change of 87.037%, indicating that the most significant changes occurred in mid–high latitudes. The largest rate value change for AEM occurred in 0–30°S, where the positive trend increased from 0.00195 to 0.00231, an increase of 0.00036. The largest percentage change occurred in 30–60°S, where the positive trend increased from 0.00161 to 0.00272, a change of 68.944%. Due to natural and anthropogenic activities, the largest positive rate changes in AEC occurred in mid–high latitudes, and AEM’s occurred in the low–mid latitudes of the Southern Hemisphere.

4. Conclusions

This study quantitatively assessed the suitability of AOD and AE for CAMS and MERRA-2 reanalysis products globally from 2003 to 2023 at different spatio-temporal scales. The similarities, differences, and reliability of their spatial patterns and long-term trends on land were discussed through a TS Slope, the EEMD method, and various time-series tests.

1. The correlation of AODC was slightly better than AODM, while AODM exhibited better robustness. The best-performing regions for both were in the low–mid latitudes. The AOD correlation for ocean was significantly higher than land, but the differences between ocean regions were larger than those on land. The performance of AE was similar, with AEM performing better. Regions with good AE performance were generally distributed in low latitudes and near Europe.

2. CAMS and MERRA-2 showed moderate to good performance in annual cycles, but with significant fluctuations and systematic biases. AOD demonstrated good correlations across different seasons. AE performed better in summer for both models, while AEM performed better than AEC in other seasons. From an hourly perspective, AODM performed better with higher Within EE values throughout the day.

3. The distribution and trends of AOD products on land showed high spatial consistency, with a decreasing trend. The high AOD values and decreasing trends were mainly found in low–mid latitudes. AEM was generally lower than AEC in spatial distribution. AEC and AEM exhibited downward and upward trends, respectively, and their trend distributions showed large differences. Through EEMD analysis, it was found that the others exhibited monotonic decreasing trends except for AEM. AODC and AODM were relatively consistent. AEC and AEM showed significant differences, consistent with before EEMD, and the trends after the turning points better represented their full temporal sequence.