1. Evaluation of MACC total aerosol optical depth and its correction model based on the random forest regression.
- Author
-
Zhen, Yang and Shi, Guoping
- Subjects
- *
RANDOM forest algorithms , *AEROSOLS , *MACHINE learning , *ATMOSPHERIC composition , *SOLAR radiation , *AIR pollutants - Abstract
The aerosol optical depth (AOD) is an indicator of particles suspended in the atmosphere, which measures the attenuation of transmitted spectral radiant energy through aerosols. It has been proven to be an essential factor in the estimation of air pollutants concentration and surface solar radiation. However, due to the sparse distribution of ground-based measurements and the high missing rate of satellite retrieval products, the reanalysis data with high spatial–temporal continuity and large spatial range is an excellent supplementary data source. However, compared with satellite monitoring and ground-based measurements, the reliability of the reanalysis data has been questioned. Therefore, in this paper, we evaluate the accuracy of the Monitoring Atmospheric Composition and Climate (MACC) AOD values at the wavelength of 550 nm (AOD550) using the ground-based observations data from 19 AERONET stations in the Chinese mainland for the period 2003–2007. In addition, a data fusion correction model based on the random forest regression is constructed and analyzed. The correction results demonstrate that the correction model, which captures the non-linear relationship, has a satisfactory correction effect overall, with R from 0.771 to 0.978, MAE from 0.225 to 0.047, and RMSE from 0.400 to 0.120. Furthermore, with the aid of model-agnostic machine learning explanation algorithms, we weigh the relative importance of the independent variables and visualize the complicated relationship between the model errors and the important input features. In general, the availability of MACC AOD was improved, and the correction model also provided a reference for the correction of other reanalysis data. [ABSTRACT FROM AUTHOR]
- Published
- 2023
- Full Text
- View/download PDF