Aerosols play a critical role in the Earth’s climate system and hydrological cycle. They not only alter the Earth’s energy balance by directly absorbing or scattering solar radiation but also profoundly influence climate and hydrological processes by indirectly affecting the physical and optical properties of clouds. For instance, aerosols such as carbon particles can enhance scattering effects or act as cloud condensation nuclei, promoting the formation of cloud droplets. Furthermore, aerosols have significant impacts on regional visibility, air quality, and human health. Despite growing attention to the environmental and climatic effects of aerosols, their complex chemical composition, short lifetimes, and highly uneven spatial distribution make it challenging to accurately characterize their global distribution and dynamic changes, remaining one of the critical difficulties in current research. The particulate observing scanning polarimeter (POSP), onboard the GF-5 satellite launched on September 7th, 2021, adopts a cross-track scanning approach and collects polarized radiometric data across nine spectral bands ranging from ultraviolet to shortwave infrared. Equipped with onboard polarization calibration and solar diffuse reflectance calibration functionalities, POSP offers multi-spectral channels with high polarization accuracy, making it particularly suitable for achieving precise aerosol retrievals. While many studies have explored aerosol optical depth (AOD) retrieval using machine learning methods, there is currently no aerosol retrieval algorithm for POSP data that operates without prior knowledge of surface types and atmospheric states. Therefore, this study focuses on developing a neural network-based aerosol retrieval method for POSP data, with case studies in representative regions of China, such as the Beijing-Tianjin-Hebei region and Taiwan region, underscoring its significance for advancing aerosol remote sensing.

The proposed algorithm enables the independent retrieval of AOD over land without requiring prior knowledge of surface types or atmospheric conditions. The model utilizes apparent reflectance and apparent polarized reflectance from seven spectral bands as training inputs. Training data is generated using the unified linearized vector radiative transfer model (UNL-VRTM) and further supplemented by constructing a truth-based training dataset through spatiotemporal matching between POSP observations and aerosol robotic network (AERONET) ground-based AOD measurements, as well as between POSP observations and moderate-resolution imaging spectroradiometer (MODIS) AOD products. After neural network parameter optimization, the model is capable of real-time AOD retrieval without the need for additional radiative transfer computations.

Result and Discussions To evaluate aerosol retrieval accuracy in typical regions of China, four observation sites—Beijing, Baotou, Taiwan, and Hong Kong—are selected for validation [Figs. 5(a)?(d)]. At the Beijing site, Bias, correlation coefficient (Corr), and root-mean-square error (RMSE) are -0.01, 0.94, and 0.06, respectively, demonstrating the algorithm’s excellent performance in urban environments, accurately capturing aerosol optical properties. At Baotou, the correlation is lower (Corr is 0.56) due to high surface reflectance variability and complex aerosol characteristics in arid regions, while RMSE remains low (0.07). Taiwan and Hong Kong showed moderate to strong applicability in island and coastal regions, with correlations of 0.64 and 0.72, and RMSEs of 0.05. Validation using eastern China data (Table 5) further assessed POSP AOD’s applicability in complex environments, highlighting its advantages in diverse conditions. On June 10th, 2024, the Beijing-Tianjin-Hebei region’s AOD retrieval (Fig. 6) shows consistent spatial trends between POSP AOD [Fig. 6(a)] and MODIS AOD [Fig. 6(b)], with higher values in the south and lower values in the north due to limited pollution sources in higher-altitude, vegetated regions. Scatterplot analysis [Fig. 6(c)] shows high agreement (Corr is 0.93, Bias is 0.03, RMSE is 0.11). Similarly, for Hefei on January 12th, 2024 (Fig. 7), POSP exhibits a high correlation (Corr is 0.95, Bias is 0.04, RMSE is 0.07) and superior detail in high AOD regions due to higher spatial resolution and polarization data. For Taiwan on February 14th, 2024 (Fig. 8), POSP and MODIS AOD shows strong consistency (Corr is 0.90, Bias is 0.01, RMSE is 0.06), with lower AOD in the east due to mountainous terrain, vegetation, and monsoon-driven aerosol dispersion.

We propose a neural network-based AOD retrieval method using the multi-spectral and polarization observation data from the POSP sensor. By taking multi-band apparent reflectance and polarized reflectance as inputs, the method achieves high-precision AOD retrieval without relying on prior surface or atmospheric information. Training data are generated using the UNL-VRTM, supplemented with true-value data from joint modeling of satellite and ground-based observations. Optimized neural network parameters enhance the algorithm’s robustness and applicability. Validations in the Beijing-Tianjin-Hebei industrial region, Hefei agricultural region, and Taiwan island region demonstrate the method’s performance. Against AERONET data, POSP achieves Corr of 0.94 and RMSE of 0.06 in Beijing, capturing complex urban aerosol characteristics. In Baotou, despite surface variability, it achieves an RMSE of 0.07, with moderate Corrs in Taiwan (0.64) and Hong Kong (0.72), verifying its adaptability to diverse environments. Overall, Corr with AERONET is 0.90, with RMSE of 0.06. Compared to MODIS, POSP demonstrates better spatial resolution and detail in high-AOD regions, particularly in industrial, agricultural, and coastal regions. This method effectively mitigates surface reflectance variability impacts, showing broad applicability and accuracy in complex environments. Future work will expand the algorithm globally, leveraging multi-source data to enhance retrieval accuracy, supporting large-scale aerosol monitoring and operational applications.