토지 보상 지장물 식별 자동화를 위한 SegFormer 기반 드론 영상 의미론적 분할 기법

초록

Land compensation involves the acquisition of private land and property rights for public projects, along with compensation for the resulting losses, and therefore requires precise assessment of existing land status. Although drones enable the periodic collection of high-resolution imagery, conventional identification of land compensation structures based on manual inspection raises concerns regarding efficiency and objectivity. To automate the identification of land compensation structures from drone imagery, this study employed SegFormer, a Transformer-based semantic segmentation model. To address the severe class imbalance inherent in the land compensation structure dataset, we systematically evaluated the individual and combined effects of three data augmentation strategies: Pixel-based Weighted Sampling, Refined Copy-Paste Augmentation, and Geometric Augmentation. Experimental results indicate that the performance improvement obtained from Refined Copy-Paste Augmentation was limited, likely due to insufficient consideration of spatial consistency between foreground objects and background context. When Refined Copy-Paste Augmentation was excluded, the combination of Pixel-based Weighted Sampling and Geometric Augmentation achieved the best segmentation performance, yielding a weighted F1-score of 0.9049. This combination also improved the F1-scores of minority classes (container and tomb) from 0.4622 to 0.5650 and from 0.7028 to 0.7288, respectively, compared with the baseline model. These findings confirm the effectiveness of the proposed strategy in mitigating class imbalance in drone-based land compensation structure segmentation. © 2026, Korean Society of Surveying. All rights reserved.

키워드