국토지리학회지 ISSN:1225-3766(Print) THE GEOGRAPHICAL JOURNAL OF KOREA

Atmospheric correction is a crucial preprocessing step that involves the removal or minimization of atmospheric effects from top of atmosphere (TOA) reflectance observed by sensors. Recently, there has been a growing trend in the use of machine learning and deep learning-based approaches capable of handling complex non-linear relationships for atmospheric correction. Accordingly, in this study, atmospheric correction of CAS500-1 (Compact Advanced Satellite 500-1) images was performed based on a machine learning algorithm simulating the 6SV radiation transfer model using Sentinel-2 satellite images as learning material. When comparing the results of the atmospheric correction model constructed through the Random Forest algorithm with those of the 6S model, excluding specific situations in the bands, the machine learning approach exhibited an accuracy of RMSE less than 0.003 and a correlation coefficient (CC) exceeding 0.99 when compared to the 6SV calculations. Additionally, in unusual situations, it produced more stable results than physics-based model. Therefore, in order to perform precise atmospheric correction for ultra-high-resolution satellite images such as CAS500-1, a machine learning-based approach that is excellent in terms of accuracy and speed is considered appropriate.