Abstract: Aerosol optical depth (AOD) is a key parameter reflecting the scattering and absorption characteristics of aerosol particles in the atmosphere. In tower-based ecological remote sensing observation, obtaining accurate AOD is crucial for quantifying the scattering and absorption effects of aerosols on the quantitative processing of spectral observation data. We developed an AOD retrieval method based on tower-based hyperspectral observation data and machine learning. Firstly, based on the atmospheric radiative transfer model, we analyzed the sensitivity of solar irradiance to AOD variations and its spectral characteristics. Then, we selected the 785 and 665 nm bands irradiance (E₇₈₅ and E₆₆₅), and constructed a dual-channel ratio index (E₇₈₅/E₆₆₅) to sensitively track the variations of AOD. Finally, we systematically evaluated the retrieval accuracy of three machine learning models (random forest, support vector machine and artificial neural network). The results showed that the AOD estimated from the three machine learning models based on tower-based hyperspectral observation data all demonstrated high accuracy, with coefficients of determination of 0.950, 0.936, and 0.947; root mean square errors of 0.025, 0.028, and 0.027; and mean absolute errors of 0.017, 0.020, and 0.019, respectively. The random forest model achieved the best performance among the three models. Machine learning methods have the potential to accurately estimate AOD from solar irradiance spectral data without the need for additional auxiliary data, which can provide reliable and synchronous AOD estimation data for the atmospheric correction methods of the tower-based platform.

Key words: aerosol optical depth, machine learning, tower-based hyperspectral observation, atmospheric correction