The retrieval of aerosol optical properties based on a random forest machine learning approach: Exploration of geostationary satellite images

Abstract

Aerosol optical properties are among the most fundamental parameters in atmospheric environmental studies. Satellite aerosols retrievals that are based on deep learning or machine learning approach have been widely discussed in remote sensing studies, but the flexible random forest (RF) model has not received much attention in the retrieval of geostationary satellite, like Himawari-8. Thus, the Himawari-8 aerosol retrieval achieved by RF model requires further investigation and optimization. Based on the radiative transfer equation, this study proposed a RF model driven by a differential operator, which quantifies a simple linear relationship between aerosol optical depth (AOD) and top-of-atmosphere (TOA) reflectance enhancement. The spectral information of aerosols is achieved by independent TOA reflectance comparison between images rather than one result from multiple band synthesis. The method allows simple feature inputs and shows weak dependence on auxiliary data. It also achieves simultaneous retrievals over different surfaces and maintains mathematical correlation between spectral AODs and Angstrom Exponents (AE). The model performance was evaluated using a series of comprehensive temporal and spatial validation analyses. A sample-based tenfold cross-validation (10-CV) shows that the new method can simultaneously improve the estimation of aerosol properties, with considerably high correlation coefficients (R2) of 0.85 for AODs at the 0.50 μm wavelengths, a mean absolute error (MAE) of 0.08, a root mean square error (RMSE) of 0.13 and >70% of the samples fell within the AOD expected error (EE). The high accuracy of the spectral AOD retrievals also exhibits good performance on AE calculations, with at least 2/3 of the samples falling within the EE. The site based 10-CV also evaluates the spatial predictions on AODs at the 0.50 μm wavelength, with R2 of 0.67, MAE of 0.12 and RMSE of 0.18. It also has outperformed the Himawari operational aerosol products and appeared to be comparable to other popular machine learning models with better AE retrievals in some typical regions. Two typical regional pollution cases also highlight the advantages of the new aerosol monitoring approach. The 5 km resolution aerosol retrievals exhibit good spatial coverage and performance when describing the regional pollution levels and types. The proposed method improves the performance of RF in retrieving aerosol properties from geostationary satellites and also offers a new prospective for aerosol remote sensing using machine learning approaches.