Comparison between Object-based Method and Deep Learning Method for Extracting Road Features Using Submeter-grade High-resolution Satellite Imagery

Abstract

In recent years, the Ministry of Land, Infrastructure, and Transport of the Republic of Korea has been developing two land observation satellites, KAS500-1 and KAS500-2, to rapidly observe the territory of South Korea and maximize the utilization of satellite images. Essential data, such as high-quality satellite images and ground control points, must be quickly provided to users, to ensure the successful utilization of the land observation satellite images. Furthermore, the users should be able to easily use application technologies, such as land use (LU) classification and spatial feature extraction, which are essential for utilizing satellite images. Object-based methods are mainly used for extracting spatial features from submeter-grade high-resolution satellite images with a ground sample distance (GSD) of 0.5 m. In recent years, advances in artificial intelligence (AI) have led to an increase in the use of deep learning. Herein, an object-based spatial feature extraction software tool based on open-source software and a deep learning-based spatial feature extraction module are developed to successfully utilize the data from land observation satellites. Images from the KOMPSAT-3A satellite, which have specifications similar to those of the KAS500 satellite images, were used to extract road features, and a quality analysis of the two methods was performed. For the quality analysis, the object-based spatial feature extraction software developed in this study, a commercial software product, and the deep learning-based spatial extraction module were each used to extract road features, and the extracted road features were analyzed. The results showed that extraction of road features using the commercial software showed 84.1% accuracy and 54.0% recall, whereas that using the software developed in this study using an open-source software product showed 83.9% accuracy and 50.2% recall. Thus, the software developed in this study and the commercial software can extract spatial features at similar levels of accuracy and recall. Road features were extracted using the module based on deep learning with 88.6% accuracy and 29.7% recall.