HRNet과 Transformer를 활용한 고해상도 위성영상의 구름탐지

The demand for satellite imagery with high spatial resolution has increased since various remotely sensed satellite sensors such as KOMPSAT (Korean Multi-Purpose Satellite) and CAS (Compact Advanced Satellite) have launched. To quickly utilize high-resolution satellite imagery, ARD (Analysis Ready Data) with preprocessing steps such as radiometric and geometric corrections should be required. Various algorithms on cloud detection techniques for satellite imagery have been developed to process satellite imagery in ARD format. In this manuscript, a deep learning model for cloud detection in satellite imagery with high spatial resolution was developed. The Transformer layer was integrated within the channel fusion process of HRNet (High Resolution Network), which is one of the representative CNN (Convolutional Neural Network), to enhance the performance of the deep learning model. Additionally, the training performance by applying preprocessing steps was improved using AIHub's KOMPSAT training dataset for cloud detection. Experimental results represented that preprocessing of the training data improved the learning performance. Furthermore, through performance evaluation with existing deep learning models, it was confirmed that the proposed deep learning model could effectively extract cloud regions.