Satellite imagery is an effective supplementary material for detecting and verifying nuclear activities and is helpful in areas where access and information are limited, such as nuclear facilities. This study aims to build training data using high-resolution KOMPSAT-3/3A satellite images to detect and identify key objects related to nuclear activities and facilities using a semantic segmentation algorithm. First, objects of interest, such as buildings, roads, and small objects, were selected, and the primary dataset was built by extracting them from the AI dataset provided by AIHub. In addition, to reflect the features of the area of interest (e.g., Yongbyon, Pyongsan), satellite images of the area were acquired, augmented, and annotated to construct an additional dataset (approximately 150,000). Finally, we conducted three stages of quality inspection to improve the accuracy of the training data. The training dataset of this study can be applied to semantic segmentation algorithms (e.g., U-Net) to detect objects of interest related to nuclear activities and facilities. Furthermore, it can be used for pixelbased object-of-interest change detection based on semantic segmentation results for multi-temporal images.