Abstract: In order to fully exploit the environmental monitoring capability of domestic Gaofen-4 satellite (GF-4), overcome the complexity of retrieving aerosol optical depth (AOD) based on look-up table method, make up for the lack of aerosol monitoring station in Guiyang City, China, and improve spatial and temporal resolution of MODIS AOD products, the AOD inversion model for Guiyang was constructed based on LightGBM algorithm using the elevation data of the study area and the surface reflectance, solar zenith angle, satellite zenith angle, relative azimuth angle and normalized different vegetation index extracted from the data of GF-4 panchromatic multispectral sensor (PMS) as the feature variables, and taking MODIS aerosol products as the label. The research results show that the model can achieve high precision AOD intelligent inversion based on GF-4 PMS single-phase remote sensing data, greatly simplify the AOD inversion steps, and has high modeling accuracy (with mean absolute error EMA, root mean square error ERMS and coefficient of determination R2 of 0.042, 0.057, and 0.751, respectively) and prediction accuracy (with EMA and ERMS of 0.077 and 0.086 respectively in urban area, and EMA and ERMS of 0.094 and 0.101 respectively in non-urban area). In addition, it is shown that the predicted AOD with the proposed inversion model and MODIS AOD have a similar variation trend, and the Pearson correlation coefficient of them is 0.697.

Key words: Gaofen-4, moderate-resolution imaging spectroradiometer, aerosol optical depth, light gradient boosting machine, Guiyang City