상세 보기
Improving Mountainous Road Segmentation Using Tree Occlusion-Based Data Augmentation
- Kim, Dae-Beob;
- Yun, Seok-Teak
WEB OF SCIENCE
0SCOPUS
0초록
Mountainous roads are critical spatial data for supporting wildfire response and forest management. However, narrow roads, tree canopy occlusion, and complex backgrounds in mountainous areas reduce the accuracy and connectivity of road segmentation from optical satellite imagery. In this study, we propose Tree Occlusion Generation (TOG) to mitigate road disconnection caused by tree canopy occlusion and apply Image-Adaptive Post-Processing (IAPP) as an auxiliary step to reduce small isolated false positives in the final predictions. TOG is designed to learn road continuity in occluded sections by generating synthetic occlusions over road regions during training while preserving the ground-truth labels. IAPP removes small isolated connected components from the prediction results. The proposed framework was evaluated using Light Roadformer as the baseline network. The dataset was constructed from Compact Advanced Satellite 500-1 (CAS500-1) imagery acquired over seven geographically distinct mountainous regions in South Korea, encompassing forested, rural, settlement-adjacent, and residential-adjacent road environments with narrow roads and tree-canopy occlusion. Four regions were used for model training, while the remaining three geographically separate regions were used exclusively for testing. Using this dataset, we compared TOG with existing masking-based augmentation techniques and analyzed performance variations across the test regions. When the results from the three test regions were combined, TOG improved the Intersection over Union (IoU) from 63.08% to 64.37% and the centerline Dice score (clDice) from 84.83% to 85.89%. With the auxiliary IAPP step, the final results achieved an IoU of 64.60% and a clDice of 86.13%. These results demonstrate that tree canopy occlusion augmentation accounting for road location can improve road connectivity and segmentation performance in mountainous road imagery.
키워드
- 제목
- Improving Mountainous Road Segmentation Using Tree Occlusion-Based Data Augmentation
- 저자
- Kim, Dae-Beob; Yun, Seok-Teak
- 발행일
- 2026-08
- 유형
- Article
- 저널명
- 대한원격탐사학회지
- 권
- 42
- 호
- 4
- 페이지
- 605 ~ 620
- 언어
- ENG
- 출판사
- 대한원격탐사학회
- 발행국가
- 대한민국
- 분량
- 16 페이지
- ISSN
- E 2287-9307
P 1225-6161