표제지
국문요약
Abstract
목차
1. 서론 14
2. 연구 자료 및 방법 17
2.1. 연구 지역 및 자료 17
2.2. 자료 전처리 20
2.3. 심층 해양 원격 탐지 기술 24
2.3.1. 3차원 합성곱 신경망 24
2.3.2. 기계학습, DNN, 2D-CNN 28
2.3.3. 실험 설정 29
3. 결과 및 토의 34
3.1. 입력 변수 사례 분석 34
3.2. 혼합층에 대한 시/공간 자료의 영향 39
3.3. 3D-CNN 모델의 혼합층 수심 예측 정확도와 자료동화된 수치모델(GLORYS12V1) 결과와의 비교 43
3.4. 계류 부이 자료를 이용한 혼합층 수심 예측 모델 테스트 48
4. 결론 58
5. 참고문헌 60
Table 1. Summary of the data used for the MLD prediction model. 23
Table 2. Summary of 11 model cases run to examine the effects of input parameters on the mixed layer. Capitalized words... 33
Figure. 1. Locations of the Argo floats measured (red dots) at least once since 2003, KEO (yellow star) buoy and PAPA (cyan star) buoy. Blue color represents areas where the water depths are less... 19
Figure. 2. Histograms for summertime mixed-layer depth (MLD) in the (a) northern hemisphere and (b) southern hemisphere. Solid... 22
Figure. 3. Schematic diagram of the 3D-CNN model. The numbers in parentheses mean channel, timestep, height, and width. Red boxes represent the kernel used in each layer which contains... 26
Figure. 4. Structure of the convolutional block in each layer of the 3D-CNN. An individual block consists of three sets of bach... 27
Figure. 5. Histogram of MLD dataset from 2004-2019. Solid black and dotted red lines indicate averaged MLD (60.04 m), and... 31
Figure. 6. RMSE and R² of the 3D-CNN model according to the timestep. Blue and orange bars plot the RMSE values for the... 31
Figure. 7. The training and validation loss along epochs for the 3D-CNN models in the (a) northern and (b) southern... 32
Figure. 8. RMSE and R² values for the 3D-CNN model applied to 11 Cases in the (a) northern and (b) southern hemispheres. Histograms indicate the RMSE for each model, and red lines and dots... 36
Figure. 9. Monthly (a) RMSE, (b) NRMSE, and (c) R² calculated from 11 Cases in the northern hemisphere. Thick red line (Case C3)... 37
Figure. 10. Monthly (a) RMSE, (b) NRMSE, and (c) R² calculated from 11 Cases in the southern hemisphere. Thick blue line (Case D)... 38
Figure. 11. Model performance comparisons of 3D-CNN with 2D-CNN and DORS methods (DNN, RFs, and XGboost) in the northern... 40
Figure. 12. Model performance comparisons of 3D-CNN with 2D-CNN and DORS methods (DNN, RFs, and XGboost) in the southern... 41
Figure. 13. RMSE versus MLD (m) for each method in the (a) northern and (b) southern hemispheres. Thick red line indicates the accuracy of the 3D-CNN. 42
Figure. 14. Scatter plots for the 3D-CNN validation at 0, 6, 12, and 18 hours. Panels in (a) and (b) indicate the accuracy of the northern and southern hemispheres, respectively. Note that log... 45
Figure. 15. Scatter plots for the 3D-CNN at 12h (left), the daily mean 3D-CNN (middle), and GLORYS12V1 (right). Panels in (a)... 46
Figure. 16. Monthly (a, c) RMSE and (b, d) R² of the 3D-CNN (red line) and GLORYS12V1 (yellow line) in the (a, b) northern and (c,... 47
Figure. 17. Comparisons of MLDs from Argo floats with those from the (a) KEO and (b) PAPA buoys. Background color shows the... 50
Figure. 18. Comparisons of MLD time series from the 3D-CNN model with those from (a) KEO and (b) PAPA buoys. Gray lines... 51
Figure. 19. Monthly-mean MLD map from the 3D-CNN model (January & February). 52
Figure. 20. Monthly-mean MLD map from the 3D-CNN model (March & April). 53
Figure. 21. Monthly-mean MLD map from the 3D-CNN model (May & June). 54
Figure. 22. Monthly-mean MLD map from the 3D-CNN model (July & August). 55
Figure. 23. Monthly-mean MLD map from the 3D-CNN model (September & October). 56
Figure. 24. Monthly-mean MLD map from the 3D-CNN model (November & December). 57