학술논문
PlanetScope와 Sentinel-2 위성영상을 활용한 Swin Transformer 기반 산불피해 탐지 성능 개선: 다종 데이터 융합의 효과 분석
Enhancing Wildfire Damage Detection Performance with Swin Transformer Using PlanetScope and Sentinel-2 Satellite Imagery: Analysis of Multi-Source Data Fusion
Enhancing Wildfire Damage Detection Performance with Swin Transformer Using PlanetScope and Sentinel-2 Satellite Imagery: Analysis of Multi-Source Data Fusion
Document Type
Article
Author
이도이 / Doi Lee; 손상훈 / Sanghun Son; 배재구 / Jaegu Bae; 박소련 / Soryeon Park; 이성혁 / Seonghyuk Lee; 서정민 / Jeongmin Seo; 김예지 / Yeji Kim; 최민하 / Minha Choi; 이양원 / Yangwon Lee; 김진수 / Jinsoo Kim
Source
대한원격탐사학회지 / Korean Journal of Remote Sensing. Dec 31, 2024 40(6):991
Subject
Language
Korean
English
English
ISSN
1225-6161
Abstract
산불 피해지 탐지는 산림관리와 복구전략 수립에 있어 핵심적인 역할을 한다. 본 연구는 딥러닝 Swin (shifted windows) Transformer 모델을 활용하여 PlanetScope와 Sentinel-2 고해상도 위성 영상을 융합함으로써 산불 피해지 탐지 성능을 효과적으로 향상시킬 수 있음을 보여주었다. 한국에서 발생한 대형산불 5개 지역을 대상으로 5-fold 교차검증을 수행하여, 각 지역별로 독립적인 탐지 성능을 도출함으로써 모델의 신뢰성과 일반화를 평가하였다. 기존의 기하학적 증강기법과 AugMix 기법의 성능비교 결과, AugMix는 평균 F1-score 76.48로 기존 증강기법보다 일관되고 향상된 성능을 보였다. 또한, Sentinel-2 데이터를 추가한 모델은 지역별 평균 F1-score 81.06을 달성하며 탐지 정확도와 피해지역 성능이 개선되었음을 확인하였다. 이는 두 데이터셋의 상호 보완적인 공간적·분광적 정보를 융합하여 더 높은 정확도와 정밀도로 산불 피해지를 탐지할 수 있음을 보여주며, PlanetScope와 Sentinel-2 데이터 간 높은 상관계수는 다종위성 융합 접근법의 효과적인 활용가능성을 뒷받침한다. 본 연구는 지역별 독립적 평가를 통해 다종위성 영상을 활용한 접근법이 산불 피해지 탐지의 정밀성과 신뢰성을 높일 수 있는 가능성을 제시한다.
Wildfire damage detection has a pivotal role in forest management and recovery strategy planning. This study demonstrates that the use of a deep learning Swin (shifted windows) Transformer model, by integrating high-resolution PlanetScope and Sentinel-2 satellite imagery, can effectively enhance wildfire damage detection performance. A 5-fold cross-validation was conducted on five major wildfire areas in South Korea, with independent detection performance evaluated for each region to assess the model’s reliability and generalization. The comparison between traditional geometric augmentation methods and the AugMix method showed that AugMix resulted in a more consistent and improved performance, achieving an average F1-score of 76.48. Furthermore, the model incorporating Sentinel-2 data achieved an average F1-score of 81.06 across regions, confirming an improvement in detection accuracy and performance in identifying burned areas. This demonstrates that combining the complementary spatial and spectral information from both datasets allows for higher accuracy and precision in detecting wildfire damage. The high correlation coefficients between PlanetScope and Sentinel-2 data support the effective potential of multi-satellite fusion approaches. This study suggests that the use of multi-satellite imagery, through independent regional evaluation, can enhance the precision and reliability of wildfire damage detection.
Wildfire damage detection has a pivotal role in forest management and recovery strategy planning. This study demonstrates that the use of a deep learning Swin (shifted windows) Transformer model, by integrating high-resolution PlanetScope and Sentinel-2 satellite imagery, can effectively enhance wildfire damage detection performance. A 5-fold cross-validation was conducted on five major wildfire areas in South Korea, with independent detection performance evaluated for each region to assess the model’s reliability and generalization. The comparison between traditional geometric augmentation methods and the AugMix method showed that AugMix resulted in a more consistent and improved performance, achieving an average F1-score of 76.48. Furthermore, the model incorporating Sentinel-2 data achieved an average F1-score of 81.06 across regions, confirming an improvement in detection accuracy and performance in identifying burned areas. This demonstrates that combining the complementary spatial and spectral information from both datasets allows for higher accuracy and precision in detecting wildfire damage. The high correlation coefficients between PlanetScope and Sentinel-2 data support the effective potential of multi-satellite fusion approaches. This study suggests that the use of multi-satellite imagery, through independent regional evaluation, can enhance the precision and reliability of wildfire damage detection.