AUTOMATED TRAINING DATA GENERATION FOR BUILDING CHANGE DETECTION USING MULTI-TEMPORAL DIGITAL MAPS AND ORTHOPHOTOS

In Korea, about 100,000 building changes occur every year, and it is essential to detect them efficiently and accurately. However, the benchmark dataset used to train existing deep learning models is based on a global environment and does not fully reflect the complex architectural structures of Korea. This study proposes a new methodology for automatically generating building change detection training data using multi-temporal digital maps and orthophotos. This methodology, which consists of three steps: data collection and preprocessing, object matching and classification, and change rate and attribute analysis, was experimentally verified in Mapo, Seoul, and proved to be more efficient and accurate than the existing manual method, classifying building changes with an accuracy of 73.43%. We proposed a criterion that can effectively distinguish between simple footprint changes and meaningful changes through the analysis of key attributes such as the IoUbased change rate and number of layers. This study confirmed the possibility of creating automated training data specific to the Korean environment. Future studies will examine the generalizability of this methodology by considering complex changes and various types of building changes and will apply it to various datasets and regions to enhance reliability.