| Title |
A Study on Improving the Efficiency of Automatic Land Cover Map Production by Improving Class Imbalance |
| Authors |
서봉상(BongSang Seo) ; 서병석(ByungSuk Seo) |
| DOI |
https://doi.org/10.5573/ieie.2025.62.11.151 |
| Keywords |
Land-cover map; Data automation; Deep learning; Class imbalance; Geospatial information |
| Abstract |
This study proposes an automated deep learning-based data construction and processing pipeline to enhance the efficiency and consistency of Level 3 land-cover map production. The conventional manual workflow for map generation requires significant time and human resources and often results in inconsistent quality across repeated tasks. To overcome these limitations, this research designs an integrated automation framework that includes image preprocessing, class label mapping, data augmentation, automatic tile segmentation, and log-based process tracking. As a result of analyzing the data distribution of 41 level 3 classes, it was found that urbanized dry areas and agricultural areas accounted for more than 70% of the total, and the data proportion of rare classes such as wetlands, salt fields, and grasslands was less than 10%. To mitigate this imbalance, weighted sampling and augmentation strategies were applied. Accordingly, we applied an augmented and weighted sampling strategy to alleviate imbalance, and confirmed that the automated construction procedure reduced the time by approximately 70% compared to manual work and the class accuracy deviation was reduced from ±8.2% to ±3.1%. These results demonstrate that the proposed method significantly improves the efficiency and reproducibility of land-cover map generation and provides a scalable foundation for periodic updates and large-scale geospatial information management systems. |