Remote Sensing of TOC Concentration in Reservoirs Using Resampled Remote-Sensing-Derived CDOM Estimates and Fluorescence Indices = 담수호에서 리샘플링된 원격탐사 기반 CDOM 추정과 형광지수를 활용한 TOC 농도 원격 모니터링

다국어 초록 (Multilingual Abstract)

Remote Sensing of TOC Concentration in Reservoirs Using Resampled Remote-Sensing-Derived CDOM Estimates and Fluorescence Indices Kim, Jinuk Department of Civil, Environmental and Plant Engineering Graduate School of Konkuk University This study proposes a machine learning-based framework that integrates remote sensing data and synthetic data techniques to predict the spatiotemporal distribution of total organic carbon (TOC) in freshwater systems. TOC is a key water quality parameter linked to aquatic carbon cycling, yet its direct remote estimation remains challenging. To address this, the study leverages indirect optical indicators—colored dissolved organic matter (CDOM), fluorescence indices (HIX, BIX, FI), and water temperature—along with hyperspectral and satellite imagery. Various machine learning models and oversampling methods were applied to overcome data imbalance and optical complexity in inland waters. In Chapter 3, a CDOM retrieval model was developed using airborne hyperspectral imagery, with the absorption coefficient at 355 nm selected as the most representative variable. Using a random forest algorithm and optimized wavelength combinations, the model achieved high accuracy (R² = 0.85, NSE = 0.77), allowing for detailed spatial mapping of CDOM within the reservoir. Chapter 4 focused on improving prediction accuracy in high-concentration ranges using synthetic data generated through the SMOTE algorithm. Machine learning models—Random Forest and LightGBM—were trained with the CDOM field data and hyperspectral reflectance ratios. The SMOTE-applied random forest model showed improved test performance (R² = 0.838, RMSE = 0.777 m⁻¹) compared to the original. In Chapter 5, Sentinel-2 imagery was utilized to develop models that integrate classification- and regression-based oversampling techniques with SVR, RFR, XGB, and DNN algorithms. Among these, DNN combined with SMOTE-SVM and SmoteR achieved the highest accuracy (R² = 0.88 and 0.89, respectively), especially improving predictions in high CDOM concentration areas. The spatial mapping results highlighted the effectiveness of oversampling in addressing challenges related to imbalanced data. Chapter 6 developed a TOC prediction model using fluorescence indices derived from 3D fluorescence excitation-emission matrix (FEEM), CDOM, and MODIS- based water temperature. Input variables were selected using IRFS-SVR-RFE, and secondary models were built to estimate HIX, BIX, and FI. The DNN model outperformed others with a prediction accuracy of R² = 0.903 and RMSE = 1.321 mg/L. Spatial TOC maps from 2017 to 2024 revealed a springtime increase in TOC, closely associated with rising HIX levels, indicating enhanced terrestrial input. In conclusion, this research demonstrates the feasibility of TOC monitoring using an integrated approach that combines remote sensing, synthetic data generation, and deep learning. The proposed framework offers a scalable and practical method for national-scale water quality monitoring and contributes valuable insights for carbon cycle research and environmental policy development.

국문 초록 (Abstract)

본 연구는 원격탐사와 기계학습 기반 알고리즘을 융합하여 담수 수계에서 총유기탄소(Total Organic Carbon, TOC)의 시·공간적 예측 프레임워크를 제안하였다. TOC는 수질 및 탄소 순환 평가에 핵심적인 변수이나, 직접적인 원격탐사 예측은 어렵다는 한계가 있다. 이에 본 연구는 TOC와 밀접한 연관을 가지는 CDOM(Colored Dissolved Organic Matter) 및 형광 기반 지수(HIX, BIX, FI), 수온 등의 간접 지표를 기반으로 TOC를 예측하였다. 또한 데이터 불균형과 수계의 광학적 복잡성을 해결하기 위해 초분광 영상 및 위성 영상을 기반으로 다양한 머신러닝 모델과 합성 데이터 기법을 적용하였다.
3장에서는 CDOM의 광학적 특성과 적정 기준 파장 선정을 위해 항공 초분광 영상 기반 CDOM 예측 모델을 구축하였다. Random Forest 모델을 이용해 입력 파장대를 최적화하고, CDOM의 대표 파장으로 a(355)를 선정하였다. 최종 모델은 R² = 0.85, NSE = 0.77의 높은 성능을 보였으며, CDOM의 공간 분포를 정밀하게 재현하였다. 4장에서는 고농도 예측 정확도 향상을 위해 SMOTE 기반 합성 데이터를 생성하고, Random Forest 및 LightGBM 모델과 비교 분석하였다. 저농도와 고농도 구간을 5 m⁻¹ 기준으로 나누어 모델을 학습한 결과, SMOTE 적용 시 test R² = 0.838, RMSE = 0.777 m⁻¹로 성능이 향상되었으며, 특히 고농도 구간에서의 추정 오류를 효과적으로 줄였다. 5장에서는 Sentinel-2 위성 영상에 기반한 분류기반 및 회귀기반 오버샘플링 기법을 비교하고, SVR, RFR, XGB, DNN 모델을 적용하여 CDOM의 시공간 예측을 수행하였다. DNN 모델은 SMOTE-SVM 및 SmoteR 기법과 결합 시 각각 R² = 0.88, 0.89의 예측력을 나타냈으며, 고농도 경계에서의 공간 분포 추정 성능도 우수하였다. 이를 통해 CDOM 추정 시 데이터 불균형 문제 해결을 위한 오버샘플링 기법의 효과를 입증하였다. 6장에서는 3D 형광 스펙트럼(FEEM) 기반 지수(HIX, BIX, FI), CDOM, MODIS 수온 등의 간접 인자를 입력 변수로 하여 TOC를 예측하였다. 위성 반사도와 형광 데이터를 연계해 머신러닝 기반 예측 모델을 구축한 결과, DNN 모델이 R² = 0.903, RMSE = 1.321 mg/L로 가장 우수한 예측 성능을 보였다. 특히 2017년부터 2024년까지의 TOC 공간 예측 결과는 봄철을 중심으로 지속적인 TOC 증가 경향을 보였으며, 이는 HIX 증가와 함께 나타나 육상 기원의 유기물 유입과 관련됨을 시사한다.
결론적으로, 본 연구는 다양한 광학적 지표와 합성 데이터 기법, 딥러닝 기반 모델을 융합하여 TOC의 원격 예측 가능성을 실증하였다. 본 프레임워크는 향후 국가 수질 모니터링 및 탄소 순환 연구에 적용될 수 있는 실용적이며 확장 가능한 접근법을 제시한다.

목차 (Table of Contents)

• Chapter 1. INTRODUCTION 1
• 1.1. Background and objective 1
• 1.2. Objectives and scope 4
• 1.3. Contents of each chapter 5
• Chapter 2. LITERATURE REVIEW 7
• 2.1. Total organic carbon and its role in aquatic systems 7
• 2.2. Spectroscopic characteristics and environmental impacts of CDOM 11
• 2.3. Fluorescent DOM and Optical Indices Based on Excitation-Emission Matrix (EEM) 14
• 2.4. Remote sensing algorithm for organic carbon 16
• Chapter 3. APPLICATION OF AIRBORNE HYPERSPECTRAL IMAGERY TO RETRIEVE SPATIOTEMPORAL CDOM DISTRIBUTION USING MACHINE LEARNING IN A RESERVOIR 21
• 3.1. Introduction 22
• 3.2. Materials and Methods 24
• 3.2.1. Site description 24
• 3.2.2. Data acquisition 25
• 3.2.2.1. Hyperspectral imagery 25
• 3.2.2.2. CDOM measurements 27
• 3.2.2.3. Selection of optimal input hyperspectral bands 27
• 3.3. Results 30
• 3.3.1. Variations in CDOM absorption coefficient and hyperspectral reflectance spectra 30
• 3.3.2. Construction of CDOM retrieving model using different combinations of hyperspectral reflectance bands in each CDOM absorption coefficient 35
• 3.3.2.1 Selection of hyperspectral bands using backward elimination 35
• 3.3.2.2 Performance of the random forest model in estimating the CDOM absorption coefficient 36
• 3.3.2.3 Determining optimal hyperspectral bands and effective CDOM absorption coefficient 38
• 3.3.3. Generating the spatial and temporal distribution of CDOM absorption coefficients using the best random forest model 40
• 3.4. Discussion 42
• 3.4.1. The effective wavelength for CDOM absorption coefficient 42
• 3.4.2. Optimal input reflectance bands to estimate CDOM in a random forest model 44
• 3.4.3. Comparing the performance of machine learning approaches to retrieve the CDOM absorption coefficient 46
• 3.4.4. Spatial and temporal distribution of CDOM absorption coefficients 47
• 3.5. Conclusion 48
• Chapter 4. ENHANCING MACHINE LEARNING PERFORMANCE IN ESTIMATING CDOM ABSORPTION COEFFICIENT VIA DATA RESAMPLING 49
• 4.1. Introduction 50
• 4.2. Materials and Methods 52
• 4.2.1. Site description and Data Acquisition 52
• 4.2.1.1 Study area 52
• 4.2.1.2 In Situ Reflectance Measurements and Airborne Hyperspectral Image 53
• 4.2.1.3 CDOM Absorption Coefficient 54
• 4.2.2. Feature Selection and Data Resampling Method 55
• 4.2.2.1 Feature Selection 55
• 4.2.2.2 Data Resampling Method 56
• 4.2.3. Construction of Machine Learning Models and Evaluation of Model Performance 57
• 4.2.3.1 Model Process 57
• 4.2.3.2 Random Forest Algorithm 59
• 4.2.3.3 Light Gradient Boost Machine (Light GBM) 59
• 4.2.3.4 Model Accuracy 60
• 4.3. Results 61
• 4.3.1. Descriptive Analysis of Chromophoric Dissolved Organic Matter in Reservoirs 61
• 4.3.2. Results of Feature Selection 63
• 4.3.3. Comparison of Machine Learning Model Performance 64
• 4.3.4. Analysis of CDOM High-Concentration Distribution Area 70
• 4.4. Discussion 72
• 4.4.1. Selection of Input Variables 72
• 4.4.2. Evaluation of Machine Learning Models and Application of Data Resampling 73
• 4.4.3. Spatial Distribution Results 75
• 4.5. Conclusion 77
• Chapter 5. CDOM ESTIMATION USING SENTINEL-2 IMAGERY IN SMALL-SCALE RESERVOIR: CLASSIFICATION AND REGRESSION-BASED DATA RESAMPLING APPROACHES 79
• 5.1. Introduction 80
• 5.2. Materials and Methods 82
• 5.2.1. Site description 82
• 5.2.2. Data acquisition 84
• 5.2.3. Feature Selection 85
• 5.2.4. Oversampling Techniques 86
• 5.2.4.1 Synthetic Minority Oversampling Technique (SMOTE) 87
• 5.2.4.2 Adaptive Synthetic Sampling (ADASYN) 87
• 5.2.4.3 SMOTE-Edited Nearest Neighbor (SMOTE-ENN) 88
• 5.2.4.4 SMOTE-Tomek Link 89
• 5.2.4.5 SMOTE-SVM 89
• 5.2.4.6 ADAENN 90
• 5.2.4.7 SmoteR 91
• 5.2.4.8 Random Oversampling (RO) 92
• 5.2.4.9 Gaussian Noise (GN) 92
• 5.2.4.10 The Weighted Relevance-Based Combination Strategy (WERCS) 93
• 5.2.5. Construction and evaluation of machine learning models 94
• 5.2.5.1 Support Vector Regression (SVR) 97
• 5.2.5.2 Random Forest Regression (RFR) 97
• 5.2.5.3 Extreme Gradient Boost Regression (XGB) 97
• 5.2.5.4 Deep Neural Network (DNN) 98
• 5.3. Results 99
• 5.3.1. Descriptive analysis of variations in CDOM absorption coefficient in Namyang reservoir 99
• 5.3.2. Select input data and Generation of synthetic data 101
• 5.3.3. Results of the performance of machine learning models 103
• 5.3.3.1 Results of overall performance for the machine learning model 103
• 5.3.3.2 Results of each model using generated synthetic data based on CM 103
• 5.3.3.3 Results of each model using generated synthetic data based on RM 108
• 5.3.4. CDOM spatial distribution results by data resampling technique for each model 111
• 5.4. Discussion 114
• 5.4.1. Consideration of input reflectance selection results 114
• 5.4.2. Effects of classification-based and regression-based oversampling on model predictive performance 116
• 5.4.3. Generation of spatiotemporal distribution of CDOM 118
• 5.4.4. Research limitations and future directions 119
• 5.5. Conclusion 121
• Chapter 6. A REMOTE SENSING-BASED MONITORING OF TOTAL ORGANIC CARBON USING FLUORESCENCE EXCITATION-EMISSION MATRIX (FEEM)-DERIVED OPTICAL INDICES IN RESERVOIR SYSTEMS 123
• 6.1. Introduction 123
• 6.2. Materials and Methods 126
• 6.2.1. Site description 126
• 6.2.2. Field sampling and water quality determination 127
• 6.2.3. Optical Analysis 129
• 6.2.4. Remote sensing imagery 131
• 6.2.4.1 Sentinel-2 imagery 131
• 6.2.4.2 MODIS LST 132
• 6.2.5. Machine learning 132
• 6.2.6. Statistical Analysis 134
• 6.3. Results and discussion 136
• 6.3.1. Identification of DOM Components Using the PARAFAC Model 136
• 6.3.2. Variation in water quality parameters and TOC algorithm design 139
• 6.3.2.1 Changes in TOC and DOM optical index in Namyang estuarine reservoir 139
• 6.3.2.2 Statistical analysis of the correlation between optical indices and TOC 140
• 6.3.2.3 Design of TOC algorithm using observation-based machine learning 143
• 6.3.3. DOM optical index and water temperature monitoring using remote sensing 146
• 6.3.3.1 DOM optical index using Sentinel-2 imagery 146
• 6.3.3.2 Water temperature using MODIS LST imagery 148
• 6.3.4. Spatio-temporal variation of total organic carbon 149
• 6.4. Conclusion 154
• Chapter 7. CONCLUSIONS 155
• References 158
• Appendix 187
• Appendix A. Atmospheric correction using MODTRAN6 187
• Appendix B. CDOM spatial distribution by machine learning and data resampling techniques 190
• Appendix C. Data histogram each machine learning and data resampling techniques 202
• Appendix D. Regression coefficients for calculating aCDOM 205
• Abstract (in Korean) 208