Performance Analysis of Semantic Segmentation Models Based on Image Enhancement Techniques for KOMPSAT Satellite Imagery

Advances in the aerospace industry have driven growing research into the use of artificial intelligence for analyzing objects of interest in satellite imagery. Unlike typical 8-bit RGB camera images, however, satellite imagery often contains 16-bit pixel values, which can result in outliers that darken the image. This issue leads to difficulty in object identification and negatively impacts analysis performance. To address this issue, various image enhancement techniques have been proposed, but the effectiveness of each technique depends on the specifics of each satellite and the task to be performed. To address this issue, various image enhancement techniques have been proposed. However, since each satellite has unique characteristics and the effectiveness of each technique varies depending on the specific task, it is necessary to carefully evaluate which technique is most suitable. This research analyzed which of the five image enhancement techniques is most suitable for semantic segmentation tasks using the dataset from the KOMPSAT-3A satellite, which is widely used in South Korea. Experimental results using five semantic segmentation models indicated that percentile stretching performed well in three models, suggesting it as the most universally applicable method. In addition, for buildings and roads, which are important objects in urban analysis, recursive separated and weighted histogram equalization (RSWHE) and percentile stretching were found to be effective.