MULTI-LABEL VEGETATION SEGMENTATION DATASET FOR URBAN AND FOREST REGIONS IN SOUTH KOREA

Urban vegetation plays a crucial role in urban ecosystems, but existing datasets are often focused on specific regions or large forests, failing to adequately reflect the characteristics and diversity of urban vegetation. In particular, high-resolution satellite imagery-based datasets are essential for vegetation analysis in regions like South Korea, where seasonal variations are pronounced. In this study, a multi-label vegetation segmentation dataset (MultiVeg) was constructed to reflect the geographical characteristics of South Korea, primarily consisting of urban and forest regions. The dataset was created through handcrafted labeling and visual inspection, and oversampling techniques were applied to mitigate class imbalance issues. To validate the quality and utility of the MultiVeg dataset, performance evaluation was conducted using the DeepLabV3+ model. In identifying the target classes of tree and grass, MultiVeg was shown to outperform or exhibit higher reliability compared to benchmark datasets (Wuhan Urban Semantic Understanding (WUSU), Zurich summer). Future research aims to enhance the performance of MultiVeg by collecting additional regional data to increase the diversity of the grass class, which showed relatively lower accuracy compared to the tree class. Furthermore, the multi-temporal data in MultiVeg will be utilized to perform precise geometric corrections and expand the dataset into a time-series vegetation change detection dataset. The dataset constructed in this study can be accessed via the following link: https://github.com/pang914/MultiVeg.