고해상도 위성영상 자료로부터 토지피복분류를 위한 다중 SVM 비교 연구

다국어 초록 (Multilingual Abstract)

Recently the high-resolution satellite images are helpful for managing the land cover, status data for the natural resources or environmental management. The effective satellite analysis process for these satellite images which requires more development to increase the effectiveness has become increasingly important. The satellite image has the advantage that you can quickly acquire the information of difficult access area, but the actual utilization rate is low. The classification result changed by user choice of classification method, the type of data, classification category. Therefore, a variety of remote sensing analysis has been an important topic for the researchers.
Support vector machines (SVMs) were originally designed for binary classification. And SVM binary classifications have 2 types of SVM, the difference is used optimization parameters C and . How to effectively extend it for multi-class classification is still an ongoing research issue. Several methods have been proposed where typically we construct a multi-class classifier by combining several binary classifiers. Some authors also proposed a method that considers all classes at once, however SVM were originally designed for binary classification. In this study, we give decomposition implementations for two such“all-together”methods.
In the study, multi-class SVM methods are explained, and then SVM methods were applied to KOMPSAT-2 high-resolution satellite images. In the first step of the preprocessing was performed by through calculation and analysis of the statistical values of the training data in the study area. From the results, we found that each classification of the training data follows a normal distribution. Then Linear, Polynomial, RBF and Sigmoid kernel function of SVM methods by changing the parameters was applied in the study area. After that, the land cover result accuracy was analyzed, using reference map. Then we compared their performance with eight multi class SVM methods: “one-versus-all,” “one-versus-one,” and directed acyclic graph SVM with C or parameters, also "Weston & Watkins SVM" and "Crammer & Singer SVM" and those methods based on binary classification and consider the classes at once.
Our experiments indicated that the "Crammer & Singer SVM" and “one-versus-all” with C parameter are more suitable for practical use than the other methods in land cover classification. Also, we showed that Linear, Polynomial and RBF kernel functions are more suitable than Sigmoid kernel function.

목차 (Table of Contents)

• 제 1 장 서론 1
• 1.1 연구배경 및 목적 1
• 1.2 연구연혁 4
• 1.3 연구내용 및 범위 7
• 제 2 장 Support Vector Machine(SVM) 10
• 2.1 이론적 배경 10
• 2.1.1 하드 마진 분류 11
• 2.1.2 소프트 마진 분류 15
• 2.1.3 Kernel 함수 16
• 2.2 이진 분류기 19
• 2.2.1 C-SVM 분류 19
• 2.2.2 nu-SVM 분류 20
• 2.3 이진 분류기 기반 다중 분류기 23
• 2.3.1 일 대 일(One-Versus-One) 다중 분류 25
• 2.3.2 일 대 다(One-Versus-All) 다중 분류 26
• 2.3.3 DAG(Directed Acyclic Graph) 다중 분류 28
• 2.4 한 번에 모든 데이터를 고려하는 다중 분류기 29
• 2.4.1 Weston & Watson 다중 분류 29
• 2.4.2 Crammer & Singer 다중 분류 37
• 제 3 장 적용 및 고찰 42
• 3.1 연구대상지역 선정 42
• 3.2 위성영상 자료 전처리 43
• 3.2.1 방사 보정 44
• 3.2.2 기하 보정 45
• 3.3 토지피복분류 46
• 3.3.1 Training Data 선정 47
• 3.3.2 영상분류 및 매개변수 선정 49
• 3.4 영상분류 결과 및 정확도 평가 52
• 3.4.1 영상분류 정확도 평가 기준 52
• 3.4.2 매개변수 및 커널함수에 따른 SVM 기법별 정확도분석 58
• 3.5 SVM 기법별 분석 결과의 비교 고찰 102
• 제 4 장 결론 118
• 참고문헌 120