딥러닝을 이용한 SAR 영상의 수체 탐지 기법 분석

다국어 초록 (Multilingual Abstract)

The scale of damage caused by disasters have been increased every year. For damage prevention and recovery, satellite images are used as data for disaster monitoring and analysis. Among satellite images, SAR images are useful for waterbody detection because they are not affected by the weather. The existing waterbody detection algorithm utilizes the threshold value of the backscattering coefficient, but there is a limitation in that it is difficult to find the optimal threshold value. In this study, U-Net, U-Net++, and U-Net 3+ techniques among the representative semantic image segmentation techniques based on deep learning were applied. A training data was constructed using satellite images of overseas regions, and then a deep learning model was trained through the training data. The results were derived by applying the learned deep learning model to the domestic region. As a result of the experiment, we confirmed that the deep learning model can effectively detect the waterbody region in the SAR satellite image.

목차 (Table of Contents)

• Ⅰ. 서 론 1
• 1.1 연구의 필요성 및 목적 1
• 1.2 국내⋅외 연구 동향 2
• 1.3 연구 방법 및 범위 3
• Ⅱ. 딥러닝 기법 6
• 2.1 딥러닝 개요 6
• 2.2 딥러닝 네트워크 11
• 2.2.1 CNN 11
• 2.2.2 SegNet 16
• 2.2.3 U-Net 17
• 2.2.3 U-Net++ 21
• 2.2.5 U-Net 3+ 24
• 2.3 딥러닝 학습을 위한 손실함수 25
• 2.4 딥러닝 학습을 위한 최적화 기법 28
• 2.5 딥러닝 평가방법 31
• Ⅲ. 수체 탐지 학습 및 분석 34
• 3.1 실험 영상 및 위성 제원 34
• 3.2 수체 탐지를 위한 훈련자료의 생성 37
• 3.3 위성영상 수체 탐지 및 예측 42
• 3.4 정확도 평가 및 결과 분석 45
• 3.4.1 검증 자료의 평가 및 결과 분석 45
• 3.4.2 테스트 자료의 평가 및 결과 분석 52
• Ⅳ. 결론 58
• 참고문헌 59