本研究結合虛擬水尺水位辨識系統與基於長短期記憶（LSTM）模型的河川水位預測，旨在提高影像辨識模型於雨天條件下的準確性，並進一步提升淹水預警的可靠性。
研究方法包括資料收集、預處理、模型建構與評估。利用2022年至2023年間基隆河成美大橋的CCTV河流影像進行LSTM模型訓練，並以上升段水位與下降段水位分別建立模型。模型訓練過程中，根據不同lookback參數測試最佳預測窗口，並結合閾值判斷調整預測結果，通過與實際水位的RMSE計算評估模型性能。
預期成果包括實現高效、低成本的河川水位監測系統，展示人工智慧在水資源管理中的應用潛力。未來可將研究成果應用於樹莓派等嵌入式平台，進一步推動環境科學與災害管理的發展。

This study integrates a virtual staff gauge–based water level recognition system with a Long Short-Term Memory (LSTM)–based river water level prediction model to improve the accuracy of image-based water level recognition under rainy conditions and further enhance the reliability of flood warning systems.
The research methodology includes data collection, preprocessing, model construction, and performance evaluation. CCTV river images collected at the Chengmei Bridge on the Keelung River from 2022 to 2023 are used for LSTM model training, with separate models constructed for rising and falling water level segments. During the training process, different lookback parameters are evaluated to determine the optimal prediction window, and threshold-based adjustments are applied to refine the prediction results. Model performance is assessed by calculating the Root Mean Square Error (RMSE) between the predicted and observed water levels.
The expected outcomes include the development of an efficient and low-cost river water level monitoring system, demonstrating the application potential of artificial intelligence in water resource management. Future work may deploy the proposed system on embedded platforms such as Raspberry Pi to further promote advancements in environmental science and disaster management.

目錄
第一章 緒論	1
第一節 研究背景	1
第二節 研究動機	1
第三節 研究目的	2
第四節 論文架構	2
第二章 文獻探討	3
第一節 影像辨識模型	3
第二節 水位在時間序列中的應用	4
第三節 誤差修正機制之研究設計	5
第三章 研究方法	6
第一節 資料收集	6
第二節 資料預處理	8
第三節 模型建構	11
壹、 模型訓練	12
貳、 模型驗證	12
第四節 誤差修正機制	14
壹、 誤差判斷與替換	14
貳、 與人工標記比對	15
參、 正確率驗證	16
第四章 實作驗證	19
第一節 資料收集	19
第二節 資料預處理	20
第三節 模型建構	23
第四節 誤差判斷與替換驗證	26
第五章 結論與未來展望	30
參考文獻	32


 
圖目錄
圖3. 1結合門檻判斷之影像水位辨識與LSTM修正流程架構	6
圖3. 2基隆河成美橋河段之CCTV水位影像監測畫面	7
圖3. 3中央氣象局每日降雨量資料查詢畫面	7
圖3. 4成美橋影像水位之人工標記與參考真值示意圖	8
圖3. 5影像像素座標轉換為時間序列水位資料之流程示意圖	8
圖3. 6人工標記影像像素座標轉換為實際水位高度之流程示意圖	9
圖3. 7中央氣象署降雨量分級標準與本研究採用之門檻示意圖	9
圖3. 8水位時間序列之上升段與下降段資料切割示意圖	10
圖3. 9水位極值點判定與上升段、下降段之人工分段示意圖	10
圖3. 10以人工標記為基準之影像辨識水位分段對應示意圖	11
圖3. 11水位上升段與下降段之分段LSTM建模架構示意圖	12
圖3. 12不同lookback設定下之Rolling-origin驗證策略示意圖	13
圖3. 13上升段與下降段LSTM模型之訓練、驗證與步數選擇流程圖	13
圖3. 14lookback步數設定原則與合理範圍示意圖	14
圖3. 15基於誤差閾值之影像水位替換判斷機制示意圖	14
圖3. 16基於RMSE指標之閾值選擇與評估流程示意圖	15
圖3. 17替換前後水位資料與人工標記之正確率比較示意圖	16
圖3. 18水位影像像素高度與正確率誤差門檻關係示意圖	17
圖3. 19原始影像辨識水位、替換後水位與人工標記水位之誤差分布盒鬚圖比較	18
圖4. 1成美橋CCTV河川影像資料蒐集與原始影像範例	19
圖4. 2臺北氣象站每日雨量資料下載與整理流程示意圖	20
圖4. 3CCTV影像經影像網格分析轉換為時間序列水位資料之流程示意圖	21
圖4. 4透過OpenCV人工標記之影像像素座標轉換為實際水位高度流程示意圖	21
圖4. 5夜間低光源條件下影像辨識水位異常與資料處理示意圖	22
圖4. 6實作驗證用高雨量事件日期之篩選結果	23
圖4. 7四個測試日期依人工標記水位切分之上升段與下降段示意圖	23
圖4. 8不同測試日期與水位段落下之 lookback 與 RMSE 關係(20220328、20220525)	25
圖4. 9不同測試日期與水位段落下之lookback與RMSE關係(20220801、20221016)	25
圖4. 10各測試日期最佳Lookback與眾數統一後RMSE影響評估	26
圖4. 11測試日期影像辨識水位修正前後之比較結果(20220328)	27
圖4. 12測試日期影像辨識水位修正前後之比較結果(20220525)	28
圖4. 13測試日期影像辨識水位修正前後之比較結果(20220801)	28
圖4. 14測試日期影像辨識水位修正前後之比較結果(20221016)	28
圖4. 15四個降雨事件中原始、修正後與人工標記水位之誤差分布盒鬚圖比較	29
圖5. 1以水位變化量（ΔH）作為輸入時之水位預測偏移現象示例	31


 
表目錄
表4. 1四個測試日期各水位段落之資料筆數統計	24
表4. 2各測試日期於上升段與下降段之最佳閾值與RMSE	27
 

[1]	K. He, X. Zhang, S. Ren, and J. Sun, “Deep Residual Learning for Image Recognition,” openaccess.thecvf.com, 2016. https://openaccess.thecvf.com/content_cvpr_2016/html/He_Deep_Residual_Learning_CVPR_2016_paper.html
[2]	S. Jing, H. Kun, Y. Xin, and Hu Juanli, “Optimization of Deep-Learning Network Using Resnet50 Based Model for Corona Virus Disease (COVID-19) Histopathological Image Classification,” 2022 IEEE International Conference on Electrical Engineering, Big Data and Algorithms (EEBDA), pp. 992–997, Feb. 2022, doi: https://doi.org/10.1109/eebda53927.2022.9744883.
[3]	A. Kaur and S. Azad, “Navigating the visual complexity: A deep dive into cifar-10 enhancement using resnet-50,” International Journal of Electronic Devices and Networking, vol. 5, no. 1, pp. 08–14, Jan. 2024, doi: 10.22271/27084477.2024.v5.i1a.47.
[4]	廖御廷。基於影像的網格分析應用於CCTV觀測河川水位〔碩士論文，淡江大學〕。華藝線上圖書館。（2023）https://doi-org.ezproxy.lib.tku.edu.tw/10.6846/tku202300603
[5]	Wu, Z., Huang, Y., Huang, K., Yan, K., & Chen, H. (2023). A review of non-contact water level measurement based on computer vision and radar technology. Water, 15(18), 3233.
[6]	H. Hidayat, A. J. F. Hoitink, M. G. Sassi, and P. Torfs, “Prediction of Discharge in a Tidal River Using Artificial Neural Networks,” vol. 19, no. 8, Aug. 2014, doi: https://doi.org/10.1061/(asce)he.1943-5584.0000970.
[7]	M. Y. A. Khan, F. Hasan, S. Panwar, and G. J. Chakrapani, “Neural network model for discharge and water-level prediction for Ramganga River catchment of Ganga Basin, India,” Hydrological Sciences Journal, vol. 61, no. 11, pp. 2084–2095, Jun. 2016, doi: https://doi.org/10.1080/02626667.2015.1083650.
[8]	F. Kratzert, D. Klotz, C. Brenner, K. Schulz, and M. Herrnegger, “Rainfall–runoff modelling using Long Short-Term Memory (LSTM) networks,” Hydrology and Earth System Sciences, vol. 22, no. 11, pp. 6005–6022, Nov. 2018, doi: https://doi.org/10.5194/hess-22-6005-2018.
[9]	Filter, K., & Data, S. (2019). Kalman filter for noise reducer on sensor readings.
[10]	ASCE Task Committee on Application of Artificial Neural Networks in Hydrology. (2000). Artificial neural networks in hydrology. I: Preliminary concepts. Journal of Hydrologic Engineering, 5(2), 115-123.
[11]	X.-H. Le, H. V. Ho, G. Lee, and S. Jung, “Application of Long Short-Term Memory (LSTM) Neural Network for Flood Forecasting,” Water, vol. 11, no. 7, p. 1387, Jul. 2019, doi: https://doi.org/10.3390/w11071387.
[12]	F. A. Buschman, A. J. F. Hoitink, M. van der Vegt, and P. Hoekstra, “Subtidal water level variation controlled by river flow and tides,” Water Resources Research, vol. 45, no. 10, Oct. 2009, doi: https://doi.org/10.1029/2009wr008167.
[13]	M. Lee, Y. You, S. Kim, K. Kim, and H. Kim, “Decomposition of Water Level Time Series of a Tidal River into Tide, Wave and Rainfall-Runoff Components,” Water, vol. 10, no. 11, p. 1568, Nov. 2018, doi: https://doi.org/10.3390/w10111568.
[14]	Y. Liu, Q. Zhao, C. Hu, and N. Luo, “Prediction of Storm Surge Water Level Based on Machine Learning Methods,” Atmosphere, vol. 14, no. 10, pp. 1568–1568, Oct. 2023, doi: https://doi.org/10.3390/atmos14101568.
[15]	A. Ma, A. Karande, N. Dahlquist, F. Ferrero, and N. R. Nguyen, “Sensor Fusion Enhances Anomaly Detection in a Flood Forecasting System,” Journal of Sensor and Actuator Networks, vol. 14, no. 2, p. 34, Mar. 2025, doi: https://doi.org/10.3390/jsan14020034.
[16]	Mosavi, A., Ozturk, P., & Chau, K. W. (2018). Flood prediction using machine learning models: Literature review. Water, 10(11), 1536.