다중 클래스 균형화 기법을 적용한 드론 영상 기반 TransUNet 모델의 내륙습지 분류 성능 개선

Wetlands are becoming increasingly important for climate change mitigation and sustainability due to their ability to absorb carbon and preserve ecosystems. As a result, research is being conducted to accurately classify wetland types using high-resolution imagery and deep learning models. However, in order to improve the classification performance of deep learning models, it is essential to acquire sufficient and balanced training data for each wetland type. However, the actual wetland environment has complex characteristics such as diversity of vegetation types and rapid environmental changes due to seasonal and climatic changes, which limit the acquisition of balanced data. Therefore, in this study, we compared and analyzed the performance of inland wetland classification by applying data balancing techniques using a drone-based high-resolution inland wetland dataset. TransUNet was utilized as a classification model. The F1-scores for the unclassified, herbaceous dominated, woody dominated, water, and bare land classes were 0.849, 0.945, 0.878, 0.985, and 0.769 for the model using only the original data, respectively, and 0.862, 0.951, 0.889, 0.985, and 0.796 for the model applying data balancing, respectively, showing performance improvement for all classes. In particular, we found performance improvements of 0.013, 0.011, and 0.027 for the unclassified, woody-dominated, and bare land classes with small proportions of data. However, even after balancing, the performance of the bare land class was still relatively low, which is likely due to the fact that it does not fully reflect the diversity and complexity of the bare land class. The results of this study are expected to be effectively applied not only to inland wetland classification but also to various unbalanced multi-class environments.